Search This Blog

Thursday, May 12, 2022

History of podcasting

From Wikipedia, the free encyclopedia
 

Podcasts, previously known as "audioblogs", has its roots dating back to the 1980s. With the advent of broadband Internet access and portable digital audio playback devices such as the iPod, podcasting began to catch hold in late 2004. Today there are more than 115,000 English-language podcasts available on the Internet, and dozens of websites available for distribution at little or no cost to the producer or listener.

Precursors

The logo of the project The Illusion of Independent Radio (1989)

The Illusion of Independent Radio is a Russian samizdat "radio program" created in 1989 in Rostov-on-Don and distributed on magnetic tape and cassettes. It was the first Soviet Russian prototype of the media phenomenon that was widely developed in the 2000s as podcasting.

Before the advent of the internet, in the 1980s, RCS (Radio Computing Services), provided music and talk-related software to radio stations in a digital format. Before online music digital distribution, the MIDI format as well as the Mbone, Multicast Network was used to distribute audio and video files. The MBone was a multicast network over the Internet used primarily by educational and research institutes, but there were audio talk programs.

Many other jukeboxes and websites in the mid-1990s provided a system for sorting and selecting music or audio files, talk, segue announcements of different digital formats. There were a few websites that provided audio subscription services. In 1993, the early days of Internet radio, Carl Malamud launched Internet Talk Radio which was the "first computer-radio talk show, each week interviewing a computer expert". It was distributed "as audio files that computer users fetch one by one". A 1993 episode of The Computer Chronicles described the concept as "asynchronous radio". Malamud said listeners could pause and restart the audio files at will, as well as skip content they did not like.

Some websites allowed downloadable audio shows, such as the comedy show The Dan & Scott Show, available on AOL.com from 1996. Additionally, in 1998, Radio Usach, radio station from the University of Santiago, Chile, explored the option to broadcast online and on demand streaming talk shows. However, the development of downloaded music did not reach a critical mass until the launch of Napster, another system of aggregating music, but without the subscription services provided by podcasting or video blogging aggregation client or system software. Independent of the development of podcasting via RSS, a portable player and music download system had been developed at Compaq Research as early as 1999 or 2000. Called PocketDJ, it would have been launched as a service for the Personal Jukebox or a successor, the first hard-disk based MP3-player.

In 2001, Applian Technologies of San Francisco introduced Replay Radio (later renamed into Replay AV), a TiVo-like recorder for Internet Radio Shows. Besides scheduling and recording audio, one of the features was a Direct Download link, which would scan a radio publisher's site for new files and copy them directly to a PC's hard disk. The first radio show to publish in this format was WebTalkGuys World Radio Show, produced by Rob and Dana Greenlee.

Timeline

In September 2000, the first system that enabled the selection, automatic downloading and storage of serial episodic audio content on PCs and portable devices was launched from early MP3 player manufacturer, i2Go. To supply content for its portable MP3 players, i2Go introduced a digital audio news and entertainment service called MyAudio2Go.com that enabled users to download episodic news, sports, entertainment, weather, and music in audio format for listening on a PC, the eGo portable audio player, or other MP3 players. The i2GoMediaManager and the eGo file transfer application could be programmed to automatically download the latest episodic content available from user selected content types to a PC or portable device as desired. The service lasted over a year, but succumbed when the i2Go company ran out of capital during the dot-com crash and folded.

The RSS connection

In October 2000, the concept of attaching sound and video files in RSS feeds was proposed in a draft by Tristan Louis. The idea was implemented by Dave Winer, a software developer and an author of the RSS format. Winer had received other customer requests for "audioblogging" features and had discussed the enclosure concept (also in October 2000) with Adam Curry, a user of Userland's Manila and Radio blogging and RSS aggregator software.

Winer included the new functionality in RSS 0.92 by defining a new element called "enclosure", which would simply pass the address to a media aggregator. On January 11, 2001, Winer demonstrated the RSS enclosure feature by enclosing a Grateful Dead song in his Scripting News weblog.

For its first two years, the enclosure element had relatively few users and many developers simply avoided using it. Winer's company incorporated both RSS-enclosure and feed-aggregator features in its weblogging product, Radio Userland, the program favored by Curry, audioblogger Harold Gilchrist and others. Since Radio Userland had a built-in aggregator, it provided both the "send" and "receive" components of what was then called "audioblogging". All that was needed for "podcasting" was a way to automatically move audio files from Radio Userland's download folder to an audio player (either software or hardware)—along with enough compelling audio to make such automation worth the trouble.

In June 2003, Stephen Downes demonstrated aggregation and syndication of audio files in his Ed Radio application. Ed Radio scanned RSS feeds for MP3 files, collected them into a single feed, and made the result available as SMIL or Webjay audio feeds.

The first on-demand radio show and the first podcast

In August 2000, the New England Patriots launched the Internet radio show PFW in Progress. It was a live show that was recorded and made available for on-demand download to visitors of Patriots.com, although this wasn't technically a podcast at the time, since the technology had not yet been invented to automatically download new episodes—a key differentiator that sets podcasts apart from simple audio files that can be downloaded manually. In 2005, two years after the introduction of the iTunes platform, the show was also offered there as a bona fide podcast. Today, it is still in existence, under the name Patriots Unfiltered, and is available on all podcast platforms. However, this was not the first podcast. That honour goes to IT Conversations by Doug Kaye; the show ran from 2003 to 2012.

In September 2003, the aforementioned Dave Winer created a special RSS-with-enclosures feed for his Harvard Berkman Center colleague Christopher Lydon's weblog, which previously had a text-only RSS feed. Lydon, a former New York Times reporter, Boston TV news anchor and NPR talkshow host, had developed a portable recording studio, conducted in-depth interviews with bloggers, futurists and political figures, and posted MP3 files as part of his Harvard blog. When Lydon had accumulated about 25 audio interviews, Winer gradually released them as a new RSS feed. Announcing the feed in his weblog, Winer challenged other aggregator developers to support this new form of content and provide enclosure support.

Not long after, Pete Prodoehl released a skin for the Amphetadesk aggregator that displayed enclosure links. Doug Kaye, who had been publishing MP3 recordings of his interviews at IT Conversations since June, created an RSS feed with enclosures, thus creating the first true podcast. Lydon's blog eventually became Radio Open Source; its accompanying podcast, titled Open Source (not to be confused with Adam Curry's Daily Source Code, which was also one of the first podcasts), is now the oldest still-running podcast.

BloggerCon

October 2003, Winer and friends organized the first BloggerCon weblogger conference at Berkman Center. CDs of Lydon's interviews were distributed as an example of the high-quality MP3 content enclosures could deliver; Bob Doyle demonstrated the portable studio he helped Lydon develop; Harold Gilchrist presented a history of audioblogging, including Curry's early role, and Kevin Marks demonstrated a script to download RSS enclosures and pass them to iTunes for transfer to an iPod. Curry and Marks discussed collaborating.

Pushing audio to a device

After the conference, Curry offered his blog readers an RSS-to-iPod script (iPodder) that moved MP3 files from Userland Radio to iTunes, and encouraged other developers to build on the idea.

In November 2003, the company AudioFeast (later renamed PodBridge, then VoloMedia) filed a patent application for “Method for Providing Episodic Media” with the USPTO based on its work in developing the AudioFeast service launched in September 2004. Although AudioFeast did not refer to itself as a podcasting service and was not built on RSS, it provided a way of downloading episodic audio content through desktop software and portable devices, with a system similar to the MyAudio2Go.com service four years before it. (AudioFeast shut down its service in July 2005 due to the unwillingness of its free customers to pay for its $49.95 paid annual subscription service, and a lack of a strong competitive differentiation in the market with the emergence of free RSS podcatchers.)

In May 2004, Eric Rice, then of SlackStreet.com, along with Randy Dryburgh of VocalSpace.com launched Audioblog.com as the first commercial podcasting hosting service. Audioblog.com became Hipcast.com in June 2006 and has hosted hundreds of thousands of podcasts since.

In September 2004, the media-in-newsfeed idea was picked up by multiple developer groups. While many of the early efforts remained command-line based, the very first podcasting client with a graphic user interface was iPodderX (later called Transistr after a trademark dispute with Apple), developed by August Trometer and Ray Slakinski. It was released first for the Mac, then for the PC. Shortly thereafter, another group (iSpider) rebranded their software as iPodder and released it under that name as Free Software (under GPL). The project was terminated after a cease and desist letter from Apple (over iPodder trademark issues). It was reincarnated as Juice and CastPodder. Commonly known as:

The name

Writing for The Guardian in February 2004, journalist Ben Hammersley suggested the term "podcasting" as a name for the nascent technology. Seven months later, Dannie Gregoire used the term "podcasting" to describe the automatic download and synchronization of audio content (as opposed to the broadcasting of digital audio, which is how the word is usually used today); he also registered several "podcast"-related domains (e.g. podcast.net). The first documented use of "podcasting" in the definition known today (i.e., broadcasting rather than downloading) was mentioned in a podcast episode of the Evil Genius Chronicles on September 18, 2004, by Dave Slusher, who also mentioned the emerging technology of torrenting as well as pondering if he should monetise the podcast (and, if so, whether it should be through sponsorship or through voluntary donations, which is a dilemma that many professional podcasters face today). As of March 2021, the recording is still available to be streamed or downloaded.

The use of "podcast" by Gregoire was picked up by podcasting evangelists such as Slusher, Winer and Curry, and entered common usage. Also in September, Adam Curry launched a mailing list; then Slashdot had a 100+ message discussion, bringing even more attention to the podcasting developer projects in progress.

On September 28, 2004, Blogger and technology columnist Doc Searls began keeping track of how many "hits" Google found for the word "podcasts". His first query reportedly returned 24 results. On September 28, 2004, there were 526 hits on Google's search engine for the word "podcasts". Google Trends marks the beginning of searches for "podcast" at the end of September. On October 1, 2004, there were 2,750 hits on Google's search engine for the word "podcasts". This number continued to double every few days.

By October 11, 2004, capturing the early distribution and variety of podcasts was more difficult than counting Google hits. However, by the end of October, The New York Times had reported on podcasts across the United States and in Canada, Australia and Sweden, mentioning podcast topics from technology to veganism to movie reviews.

Wider notice

USA Today told its readers about the "free amateur chatfests" the following February, profiling several podcasters, giving instructions for sending and receiving podcasts, and including a "Top Ten" list from one of the many podcast directories that had sprung up. Those Top Ten programs gave further indication of podcast topics: four were about technology (including Curry's Daily Source Code, which also included music and personal chat), three were about music, one about movies, one about politics, and—at the time number one on the list—The Dawn and Drew Show, described as "married-couple banter", a program format that (as USA Today noted) was popular on American broadcast radio in the 1940s (e.g. Breakfast with Dorothy and Dick). After Dawn and Drew, such "couplecasts" became quite popular among independent podcasts, the most notable being the London couple Sowerby and Luff (consisting of comedy writers Brian West (Luff) and Georgina Sowerby), whose talk show The Big Squeeze quickly achieved a global audience via the podcast Comedy 365. On October 18, 2004, the number of hits on Google's search engine for the word "podcasts" ballooned to more than 100,000 after being just 24 results three weeks prior.

In October 2004, detailed how-to-podcast articles had begun to appear online, and a month later, Liberated Syndication launched the first Podcast Service Provider, offering storage, bandwidth, and RSS creation tools. This was the same month that Podtrac started providing its free download tracking service and audience demographics survey to the podcasting industry. "Podcasting" was first defined in Wikipedia. In November 2004, podcasting networks started to appear on the scene with podcasters affiliating with one another. One of the earliest adopters from the mainstream media of on-demand audio (although not strictly a podcast) was the BBC, with the BBC World Service show, Go Digital, in August 2001. The first domestic BBC show to be podcasted was In Our Time, made available as a podcast in November 2004.

Apple adds podcasts to iTunes

In June 2005, Apple added podcasting to its iTunes 4.9 music software and building a directory of podcasts at its iTunes Music Store. The new iTunes could subscribe to, download and organize podcasts, which made a separate aggregator application unnecessary for many users. Apple also promoted creation of podcasts using its GarageBand and QuickTime Pro software and the MP4 format instead of MP3. Prior to iTunes' seamless integration, acquiring and organising podcasts required dedicated "podcatching" software that was often clunky and intimidating for the average user.

In July 2005, U.S. President George W. Bush became a podcaster of sorts, when the White House website added an RSS 2.0 feed to the previously downloadable files of the president's weekly radio addresses. Also in July, the first People's Choice Podcast Awards were held during the Podcast Expo. Awards were given in 20 categories. On September 28, 2005, exactly a year after first tracking hits for the word "podcasts" on Google's search engine, Google found more than 100 million hits on the word "podcasts". In November 2005, the first Portable Media Expo and Podcasting Conference was held at the Ontario Convention Center in Ontario, California. The annual conference changed its name to the Podcast and New Media Expo, which stopped being held in 2015. On December 3, 2005, "podcast" was named the word of the year in 2005 by the New Oxford American Dictionary and would be in the dictionary in 2006.

Expansion

In February 2006, following London radio station LBC's successful launch of the first premium-podcasting platform, LBC Plus, there was widespread acceptance that podcasting had considerable commercial potential. UK comedian Ricky Gervais, whose first season of The Ricky Gervais Show became a big hit, launched a new series of the popular podcast. The second series of the podcast was distributed through audible.co.uk and was the first major podcast to charge consumers to download the show (at a rate of 95 pence per half-hour episode). The first series of The Ricky Gervais Show podcast had been freely distributed by the Positive Internet Company and marketed through The Guardian newspaper's website, and it was the world's most successful podcast for several years, eventually gaining more than 300 million unique downloads by March 2011. Even in its new subscription format, The Ricky Gervais Show was regularly the most-downloaded podcast on iTunes. The Adam Carolla Show claimed a new Guinness world record, with total downloads approaching 60 million, but Guinness failed to acknowledge that Gervais's podcast had more than 5 times as many downloads as Carolla's show at the time that this new record was supposedly set.

In February 2006, LA podcaster Lance Anderson became nearly the first to take a podcast and create a live venue tour. The Lance Anderson Podcast Experment (sic) included a sold-out extravaganza in The Pilgrim, a central Liverpool (UK) venue (February 23, 2006), followed by a theatrical event at The Rose Theatre, Edge Hill University (February 24, 2006), which included appearances by Mark Hunter from The Tartan Podcast, Jon and Rob from Top of the Pods, Dan Klass from The Bitterest Pill via video link from Los Angeles, and live music from The Hotrod Cadets. In addition, Anderson was also invited to take part in the first-ever Podcast Forum at CARET, the Centre for Applied Research in Educational Technologies at the University of Cambridge (February 21, 2006). Organised and supported by Josh Newman, the university's Apple Campus Rep, Anderson was joined at this event by Dr. Chris Smith from the Naked Scientists podcast; Debbie McGowan, an Open University lecturer and advocate for podcasting in education; and Nigel Paice, a professional music producer and podcasting tutor. In March 2006, Canadian Prime Minister Stephen Harper became the second head of government to issue a podcast, the Prime Minister of Canada's Podcast (George W. Bush technically being the first one back in July 2005). In July 2009, the company VoloMedia is awarded the "Podcast patent" by the USPTO in patent number 7,568,213. Dave Winer, the co-inventor of podcasting (with Adam Curry), points out that his invention predated this patent by two years.

On February 2, 2006, Virginia Tech (Virginia Polytechnic Institute and State University) launched the first regular schedule of podcast programming at the university. Having four regularly scheduled podcasts was a first for a major American university, which was launched as part of Virginia Tech's "Invent the Future" campaign.

In April 2006, comedy podcast Never Not Funny began when Matt Belknap of ASpecialThing Records interviewed comedian Jimmy Pardo on the podcast for his popular alternative comedy forum A Special Thing. The two had previously discussed producing a podcast version of Jimmy's Los Angeles show "Running Your Trap", which he hosted at the Upright Citizens Brigade Theatre, but they hit it off so well on AST Radio that Pardo said "This is the show." Shortly after, Never Not Funny started simulcasting both a podcast stream and a paid video version. The podcast still uses this format, releasing two shows a week—one free and one paid—along with paid video feed.

In October 2006, the This American Life radio program began to offer a podcast version to listeners. Since debuting, This American Life has consistently been one of the most-listened-to podcasts, averaging around 2.5 million downloads per episode.

In March 2007, after being on-air talent and being fired from KYSR (STAR) in Los Angeles, California, Jack and Stench started their own subscription-based podcast. At $5.00 per subscription, subscribers had access to a one-hour podcast, free of any commercials. They had free local events at bars, ice cream parlors and restaurants all around Southern California. With a successful run of 12 years and over 2,700 episodes, the Jack and Stench Show is among the longest-running monetized podcasts.

In March 2007, the Cambridge CARET Centre also helped to give birth to the first as-live podcast channel for women politicians in the UK and globally called Women's Parliamentary Radio. A former BBC correspondent and political editor in the East, Boni Sones OBE, worked with three other broadcast journalists—Jackie Ashley, Deborah McGurran, and Linda Fairbrother—to create an online radio station where women MPs of all parties could be interviewed impartially. The MP3 files could be streamed or downloaded. Their resulting 550 interviews over 15 years can now be found in one of four audio archives nationally at the British Library, the London School of Economics, The History of Parliament Trust and the Churchill Archives University of Cambridge. Sones has also written four books about these podcast interviews and archives, which are in all the major libraries in the UK.

The Adam Carolla Show started as a regular weekday podcast in March 2009; by March 2011, 59.6 million episodes had been downloaded in total, claiming a record; however, as previously mentioned, Gervais's podcast had already received five times Carolla's downloads by the time the record was supposedly set. The BBC noted in 2011 that more people (eight million in the UK or about 16% of the population, with half listening at least once a week—a similar proportion to the USA) had downloaded podcasts than had used Twitter.

Besides the aforementioned Adam Carolla Show, 2009 saw a huge influx of many other popular new comedy podcasts, including the massively successful talk-style podcasts with a comedic bent such as WTF with Marc Maron, The Joe Rogan Experience, and the David Feldman Show. 2009 also saw the launch of the surrealist comedy show Comedy Bang! Bang! (which was known as Comedy Death-Ray Radio at the time), which was later turned into a TV show with the same name.

With a run of eight years (as of October 2013), the various podcasts provided by Wrestling Observer/Figure Four Online, including Figure Four Daily and the Bryan and Vinny Show with host Bryan Alvarez, and Wrestling Observer Radio with hosts Alvarez and Dave Meltzer, have produced over 6,000 monetized podcasts at a subscription rate of $10.99 per month. Their subscription podcast model launched in June 2005. Alvarez and Meltzer were co-hosts in the late 1990s at Eyada.com, the first Internet-exclusive live streaming radio station, broadcasting out of New York City.

In 2014, This American Life launched the first season of their Serial podcast. The podcast was a surprise success, achieving 68 million downloads by the end of Season 1 and becoming the first podcast to win a Peabody Award. The program was referred to as a "phenomenon" by media outlets and popularized true crime podcasts. True crime programs such as My Favorite Murder, Crimetown, and Casefile were produced after the release of Serial and each of these titles became successful in their own right. From 2012 to 2013, surveys showed that the number of podcast listeners had dropped for the first time since 2008. However, after Serial debuted, audience numbers rose by 3%.

Podcasting reached a new stage of growth in 2017 when The New York Times debuted The Daily news podcast. The Daily is designed to match the fast pace of modern news, and the show features original reporting and recordings of the newspaper's top stories. As of May 2019, it has the highest unique monthly US audience of any podcast.

Download records

Due to the fragmented delivery mechanisms and various other factors, it is difficult to externally nail down a precise listenership figure for any one podcast (although podcasters themselves can generally get fairly accurate data if they so please, which is especially useful for securing advertising contracts). As of December 2018, Serial was believed by some sources to be the most downloaded podcast of all time, with 420 million total downloads, surpassing Gervais's 300 million figure from back in 2011. However, Stuff You Should Know has accrued more than a billion downloads, and there are others still that have also hit this figure. According to Podtrac, NPR is the most popular podcast publisher, with over 175 million downloads and streams every month; however, Joe Rogan claimed in 2019 that his podcast alone was receiving 190 million downloads a month—a claim that is very likely true—and therefore makes his show the most downloaded podcast of all time in terms of both average viewership and total downloads. Indeed, Rogan signed a $100 million licensing deal with Spotify due to his unprecedented success with the medium.

Nielsen and Edison Research reported in April 2019 that they had logged 700,000 active podcasts worldwide. Their research also revealed that, per capita, South Korea leads the world in podcast listeners, with 58% of South Koreans listening to podcasts every month. For comparison, in 2019, 32% of Americans had listened to podcasts in the last month. In 2020, 24% of Americans had listened to podcasts weekly. Comedy is the most popular podcast genre in the United States. There are more than 1,700,000 shows and nearly 44 million episodes as of January 19, 2021. Podtrac reports iHeartRadio's shows had more than 243 million downloads. IAB and PWC project that U.S. podcast advertising revenues will surpass $1 billion by 2021.

Video podcasting

A video podcast or vodcast is a podcast that contains video content. Web television series are often distributed as video podcasts. Dead End Days, a serialized dark comedy about zombies released from 31 October 2003 through 2004, is commonly believed to be the first video podcast. Never Not Funny was a pioneer in providing video content in the form of a podcast. H3H3's H3 podcast and the Joe Rogan Experience are two examples of a litany of video podcasts, with many of them now being hosted on YouTube rather than as part of a feed (which was much more common when video podcasting was a brand-new medium). The key difference between a vlog and a video podcast is the length. While a vlog could technically be a video podcast, long-form conversational-style videos are generally considered to be a video podcast.

Popularization

Business model studies

Classes of MBA students have been commissioned to research podcasting and compare possible business models, and venture capital flowing to influential content providers.

Podnography

As is often the case with new technologies, pornography has become a part of the scene, producing what is sometimes called podnography.

Podsafe music

The growing popularity of podcasting introduced a demand for music available for use on the shows without significant cost or licensing difficulty. Out of this demand, a growing number of tracks, by independent as well as signed acts, are now being designated "podsafe".

Use by conventional media

Podcasting has been given a major push by conventional media. (See Podcasting by traditional broadcasters.)

Broadcast media

Podcasting has presented both opportunities and challenges for mainstream radio outlets, which on one hand see it as an alternative medium for their programs while on the other hand struggle to identify its unique affordances and subtle differences. In a famous example of the way online statistics can be misused by those unused to the nuances of the online world, marketing executives from the ABC in Australia were unsure of how to make sense of why Digital Living, at that stage a little-known podcast from one of their local stations, outrated all of their expensively produced shows. It turned out that a single segment on Blu-ray had been downloaded a massive 150,000 times in one day from a single location in China.

Print media

For example, podcasting has been picked up by some print media outlets, which supply their readers with spoken versions of their content. One of the first examples of a print publication to produce an audio podcast to supplement its printed content was the international scientific journal Nature. The Nature Podcast was set up in October 2005 by Cambridge University's award-winning "Naked Scientist", Chris Smith, who produces and presents the weekly show.

Although firm business models have yet to be established, podcasting represents a chance to bring additional revenue to a newspaper through advertising, subscription fees and licensing.[citation needed]

Podcamps

Chris Brogan and Christopher S. Penn launched the PodCamp unconference series aimed at bringing together people interested in blogging, social media, social networking, podcasting, and video on the net, and in so doing, Brogan won the Mass High Tech All Stars award for 2008.

Podcast Movement

Veteran podcaster Gary Leland joined forces with Dan Franks and Jared Easley to form a new international conference for podcasters in early 2014 called Podcast Movement. Unlike other new media events, Podcast Movement was the first conference of its size in over a decade that was focused specifically on podcasting, and has tracks for both new and experienced podcast creators, as well as industry professionals. The sixth annual conference is expected to be attended by over 3,000 podcasters, and is scheduled for August 2019 in Orlando, FL.

Adaptions

Some popular podcasts, such as Lore, Homecoming, My Brother, My Brother, and Me and Serial, have been adapted as films or television series.

Coping with growth

While podcasting's innovators took advantage of the sound-file synchronization feature of Apple Inc.'s iPod and iTunes software—and included "pod" in the name—the technology was always compatible with other players and programs. Apple was not actively involved until mid-2005, when it joined the market on three fronts: as a source of "podcatcher" software, as publisher of a podcast directory, and as provider of tutorials on how to create podcasts with Apple products GarageBand and QuickTime Pro. Apple CEO Steve Jobs demonstrated creating a podcast during his January 10, 2006 keynote address to the Macworld Conference & Expo using new "podcast studio" features in GarageBand 3.

When it added a podcast-subscription feature to its June 28, 2005, release of iTunes 4.9, Apple also launched a directory of podcasts at the iTunes Music Store, starting with 3,000 entries. Apple's software enabled AAC-encoded podcasts to use chapters, bookmarks, external links, and synchronized images displayed on iPod screens or in the iTunes artwork viewer. Two days after release of the program, Apple reported one million podcast subscriptions.

Some podcasters found that exposure to iTunes' huge number of downloaders threatened to make great demands on their bandwidth and related expenses. Possible solutions were proposed, including the addition of a content delivery system, such as Liberated Syndication; Podcast Servers; Akamai; a peer-to-peer solution, BitTorrent; or use of free hosting services, such as those offered by the Internet Archive or Buzzsprout.

Since September 2005, a number of services began featuring video-based podcasting, including Apple (via its iTunes Music Store), the Participatory Culture Foundation, and Loomia. These services handle both audio and video feeds.

United States abortion-rights movement

Albert Wynn and Gloria Feldt on the steps of the U.S. Supreme Court to rally for legal abortion on the anniversary of Roe v. Wade

The United States abortion-rights movement (also known as the pro-choice movement) is a sociopolitical movement in the United States supporting the view that a woman should have the legal right to an elective abortion, meaning the right to terminate her pregnancy, and is part of a broader global abortion-rights movement. The movement consists of a variety of organizations, with no single centralized decision-making body.

A key point in abortion rights in the United States was the U.S. Supreme Court's 1973 decision in Roe v. Wade, which struck down most state laws restricting abortion, thereby decriminalizing and legalizing elective abortion in a number of states.

On the other side of the abortion debate in the United States is the anti-abortion movement (or pro-life movement), which believes that human embryos and fetuses have a right to life, and abortion violates this right and should be outlawed or otherwise restricted. Within this group, many argue that human personhood begins at conception, a position rejected by many abortion rights groups.

Overview

Abortion-rights advocates argue that whether or not a pregnant woman continues with a pregnancy should be her personal choice, as it involves her body, personal health, and future. They also argue that the availability of legal abortions reduces the exposure of women to the risks associated with illegal abortions. More broadly, abortion-rights advocates frame their arguments in terms of individual liberty, reproductive freedom, and reproductive rights. The first of these terms was widely used to describe many of the political movements of the 19th and 20th centuries (such as in the abolition of slavery in Europe and the United States, and in the spread of popular democracy) whereas the latter terms derive from changing perspectives on sexual freedoms and bodily integrity.

Abortion-rights supporters rarely consider themselves "pro-abortion", because they consider termination of a pregnancy as a bodily autonomy issue, and find forced abortion to be as legally and morally indefensible as the outlawing of abortion. Indeed, some who support abortion rights consider themselves opposed to some or all abortions on a moral basis, but believe that abortions would happen in any case and that legal abortion under medically controlled conditions is preferable to illegal back-alley abortion without proper medical supervision. Such people believe the death rate of women due to such procedures in areas where abortions are only available outside of the medical establishment is unacceptable.

Some who argue from a philosophical viewpoint believe that an embryo has no rights as it is only a potential and not an actual person and that it should not have rights that override those of the pregnant woman at least until it is viable.

Many abortion-rights campaigners also note that some anti-abortion activists also oppose practices that correlate with less demand for abortion, namely sex education and the ready availability of contraception. Proponents of this argument point to cases of areas with limited sex education and contraceptive access that have high abortion rates, either legal or illegal. Some women also travel to another jurisdiction or country where they may obtain an abortion. For example, a large number of Irish women would visit the United Kingdom for abortions, as would Belgian women who travelled to France before Belgium legalized abortion. Similarly, women would travel to the Netherlands when it became legal to have abortions there in the 1970s.

Some people who support abortion rights see abortion as a last resort and focus on situations where they feel abortion is a necessary option. Among these situations are those where the woman was raped, her health or life (or that of the fetus) is at risk, contraception was used but failed, the fetus has acute congenital disorder and defects, incest, financial constraints, or she feels unable to raise a child. One common reason women give for terminating unintended pregnancies is that having a baby would prevent them from achieving goals such as pursuing an education. Some abortion-rights moderates, who would otherwise be willing to accept certain restrictions on abortion, feel that political pragmatism compels them to oppose any such restrictions, as they could be used to form a slippery slope against all abortions. On the other hand, even some abortion rights advocates feel uncomfortable with the use of abortion for sex-selection, as is practiced in some countries, such as India.

History

Abortion-rights activists before the Washington Monument in Washington, D.C., at the March for Women's Lives in 2004
 
November 12, 1989 Washington, D.C. abortion rights march

Prior to 1973, abortion rights in the United States were not seen as a constitutional issue. Abortion was seen as a purely state matter, all of which had some type of restrictions. The first legal restrictions on abortion appeared in the 1820s, forbidding abortion after the fourth month of pregnancy. A number of early feminists expressed decided skepticism about outlawing abortion. They disliked abortion but thought anti-abortion laws did not apply "the proper remedies", according to one nineteenth-century women's rights pioneer. By 1900, legislators at the urgings of the American Medical Association (AMA) had enacted laws banning abortion in most U.S. states. The AMA played a vital role in stigmatizing abortions by using their status and power to create a moral stance against abortion. The AMA viewed abortion providers as unwanted healthcare competitors. Due to the high maternal morbidity and mortality rates caused by back alley abortions, physicians, nurses, and social workers pushed for legalization of abortion from a pro-public health perspective. Support for abortion rights went beyond feminists and medical professions. The broad support for legalizing abortion in the 1960s also derived from certain religious leaders. For example, there were 1,400 clergy operating on the East Coast for the Clergy Consultation Service on Abortion—an underground network that connected women seeking abortions to doctors—during the 1960s. As the historian Christine Stansell explained, many religious leaders came to approach the abortion rights argument from a position of individual conscience instead of from dogma by witnessing the "strains unwanted pregnancies put on members of their congregations".

In its landmark 1973 case Roe v. Wade, in which a woman challenged the Texas laws criminalizing abortion, the U.S. Supreme Court reached two important conclusions:

  • That state abortion laws are subject to the due process clause of the Fourteenth Amendment to the United States Constitution; and
  • That the procurement of an abortion was a constitutional right during the first and second trimesters of a pregnancy based on the constitutional right to privacy, but that the state's interest in protecting "potential life" prevailed in the third trimester unless the woman's health was at risk. In subsequent rulings, the Court rejected the trimester framework altogether in favor of a cutoff at the point of fetal viability (cf. Planned Parenthood v. Casey).

Abortion-rights groups are active in all American states and at the federal level, campaigning for legal abortion and against the reimposition of anti-abortion laws, with varying degrees of success. Only a few states allow abortion without limitation or regulation, but most do allow various limited forms of abortion.

In the wake of extreme abortion restrictions in states like Georgia, Alabama, Missouri and Ohio, transgender, intersex and other gender-nonconforming people felt forgotten in this movement.

Abortion is one of the most common medical interventions undergone by women aged 15–44. In the US childbirth carries a risk of death about 14 times higher than abortion.

The Democratic Party's platform endorses the abortion-rights position, stating that abortion should be "safe and legal". Not all Democrats agree with the platform, however, and there is a small anti-abortion faction within the party, expressed in such groups as Democrats for Life of America. Similarly, there is a small abortion-rights faction within the Republican Party. The Libertarian Party platform holds "that government should be kept out of the matter" of abortion.

Roe v. Wade

Roe v. Wade was an essential court case in the U.S abortion rights movement. The case was first tried in June 1970 by a Texas district court. The court ruled that Texas' law for abortions was unconstitutional and was appealed to the United States Supreme Court. This particular law made it a crime to have an abortion unless the mother's life would be in danger if she had the child. This issue was brought to the U.S Supreme Court when a woman in Texas by the name of Norma McCorvey wanted to get an abortion. In the case, she was given the pseudonym "Jane Roe", and Henry Wade was the district attorney of Dallas County who the case was against. A lawsuit was filed on behalf of McCorvey by her lawyers, Linda Coffee and Sarah Weddington, claiming that particular articles for the Texas penal code, 1191-1194 and 1196, were unconstitutional. The lawyers' stance was that these articles took away a women's right to choose which went against the U.S Constitution's Ninth Amendment.

On January 22, 1972, the United States Supreme Court ruled in the favor of Jane Roe and banned the law in Texas that only allowed abortions if the mother's life was at stake. This was a monumental decision that impacted the entire country, legalizing abortions nationwide. The court's decision was decided and based on the three different trimesters in pregnancy. The ruling allowed for women to have the choice to have an abortion in the first trimester, without any Government regulation. Once a woman enters the second trimester, the Government would be allowed to regulate any abortions. The Government's interference at this stage does not mean that abortions are banned, but that they have the power to decide if a woman should be able to get an abortion if her health is at stake. Once a woman enters the third trimester, the state has the authority to approve or deny a woman having an abortion. This decision by the state is based on whether the fetus can survive outside the womb. If it can, than the state will not approve of the abortion unless the mother's life and or health is at stake.

Roe v. Wade has had a major impact on the United States abortion rights movement. Before this Supreme Court case, some women would have to resort to having unsafe and illegal abortions that could cost them their health or even lives. This case allowed women to have more of a choice when it comes to the decision of having an abortion or not. Since its ruling, states have been creating laws that make it hard for women to have the choice to get an abortion.

Organizations and individuals

The abortion rights movement includes a variety of organizations, with no single centralized decision-making body. Many more individuals who are not members of these organizations also support their views and arguments.

Planned Parenthood, NARAL Pro-Choice America, the National Abortion Federation, the National Organization for Women, and the American Civil Liberties Union are the leading abortion-rights advocacy and lobbying groups in the United States. Most major feminist organizations also support abortion-rights positions, as do the American Medical Association, the American Congress of Obstetricians and Gynecologists, and pro-abortion rights physicians such as Eugene Gu and Warren Hern who have fought political opposition from anti-abortion Senator Marsha Blackburn. Faith-based groups that advocate for abortion rights include notably the Religious Coalition for Reproductive Choice and Catholics for Choice.

Planned Parenthood

Planned Parenthood was founded on October 16, 1916, in Brownsville, New York City, New York. This organization was created so that women could have access to healthcare services and information that could help them live strong and healthy lives. Planned Parenthood is important to the United States abortion rights movement because their members are advocates for abortion access, as they believe that it is a healthcare right. Some of the issues surrounding abortion that this organization are advocating against include bans on abortion at 20 weeks, bans on abortions at 6 weeks, and the Hyde Amendment.

On July 1, 1976, the Planned Parenthood v. Danforth case was taken to the United States Supreme Court. The Planned Parenthood of Central Missouri, Dr. David Hall, and Dr. Michael Freiman challenged a Missouri law for abortion, known as House Bill 1211. Bill 1211 regulated that women who were married had to receive consent from their husband for an abortion, or consent from the parents if they were a minor and not married. Attorney Frank Susman represented Planned Parenthood in the case. Hall, Freiman, Susman, and Planned Parenthood argued against House Bill 1211 and its definition of viability. They argued that it gave a vague definition of viability that allowed for any fetus to be considered viable, essentially making abortions illegal. It was also argued that the passage of House Bill 1211 was targeting specifically individuals who wanted an abortion by requiring consent from the husband or parents for abortions, but not for other medical procedures.  Frank Susman also challenged the ban on saline amniocentesis, a procedure to induce abortions, and the regulation of doctors who performed abortions. This regulation required that the doctor care for the fetus as if it were delivered normally to save its life. John Danforth, who was the general attorney for Missouri in the case, argued that the Missouri law, House Bill 1211, was constitutional. His views were that the consent allowed for the women to properly understand and make a cautious decision about the abortion.

The Court came to a decision and kept some parts of House Bill 1211, while omitting others. The definition of fetal viability and the recordkeeping requirements were upheld. The Court decided to omit the requirement of written consent for the procedure, the requirement of doctors to care for the fetus after the abortion as if it was a conventional delivery, and the ban of saline amniocentesis.

Other than Planned Parenthood's advocacy efforts for the abortion rights movement, their members also provide information at their clinics and website for the public. Regarding abortions, they provide information for people who may be considering abortions, information on the abortion pill, where to find abortion clinics and what to expect when experiencing abortions.

NARAL Pro-Choice America

The NARAL Pro-Choice America Foundation advocates for access to abortion, birth control, and paid parental leave, and against discrimination toward pregnant people. Their members provide education about the adverse effects of policies that go against a woman's choice, advocate for the group's policies, and advocate voting for government officials who support those policies. Another sector of this organization, NARAL Pro-Choice America PAC, focuses on endorsing political candidates who are pro-choice and willing to defend the right to abortion.

The NARAL Pro-Choice America Foundation advocates against targeted regulation of abortion providers (TRAP) laws, legislative restriction on access to abortions, and abortion refusal laws. They also have been a part of successful campaigns that included support for the EACH Woman Act and Whole Woman's Health v. Hellerstedt. The EACH Woman Act is legislation that was created to eliminate the Hyde amendment. This amendment prohibits women on government health care programs, such as Medicaid, from receiving funds to pay for an abortion. In support of the EACH Woman act, the NARAL Pro-Choice Foundation gathered petition signatures. In the Supreme Court case Whole Woman's Health v. Hellerstedt the NARAL Pro-Choice Foundation advocated for this case. This particular case was taken to court because restrictions in Texas made it difficult for women to have access to abortions. The Court ruled that this was unconstitutional.

The National Abortion Federation

The National Abortion Federation supports providers who perform abortions with delivering patient care. Their members advocate for women to have a choice when it comes to having an abortion, ensuring quality abortion care, and providing a platform where providers and patients can share their personal experiences regarding abortions.  During the COVID-19 pandemic, the National Abortion Federation advocated for abortion clinics to remain open, deeming it an essential service. The National Abortion Federation also provides a Hotline that helps patients have access to abortions and financial assistance. They provide a Patient Partnership program that allows people who have had abortions to share their personal experiences and stories to lawmakers.

Religious organizations in support of the abortion rights movement

Religious Coalition for Reproductive Choice

The Religious Coalition for Reproductive Choice (RCRC) is a national nonprofit organization that was founded in 1973 shortly following the Roe v. Wade ruling. Founders of the Coalition included clergy and lay leaders from many mainstream religions. Many of these leaders had previously helped women find abortion services before the Supreme Court's ruling.

The organization's website provides several resources to people in need. Four specific issues of interest are listed and delved into: the moral case, reproductive health, reproductive rights, and reproductive justice. Beyond this, the organization offers religious resources and services to those who are considering or recovering from an abortion, as well as clinics and practitioners. Some services that are offered include clinic blessings, workshops, "Compassion School", and advocacy. Several states have a state-level coalition with varying amounts of resources:

  • Colorado has a website dedicated to RCRC called Colorado Religious Coalition for Reproductive Choice. Apart from having similar resources as the parent website, the Colorado RCRC website offers "all options counseling" with a promise to help all pregnant people who are seeking advice
  • Kentucky has a website dedicated to RCRC called Kentucky Religious Coalition for Reproductive Choice. Here, the organization shares important updates on the EMW Clinic, the state's only remaining abortion clinic. KRCRC is one of many organizations that continuously advocates to keep the clinic's doors open. Volunteers from the organization also provide escorting to people who are utilizing the clinic's services.
  • Minnesota does not have an independent for RCRC, but there is a Facebook group dedicated to the cause. It is titled "Minnesota Religious Coalition for Reproductive Choice" and focuses on many of the same issues that the parent RCRC organization highlights.
  • New Mexico has a website specifically dedicated to its state RCRC. In addition to the resources offered by the parent website, New Mexico also offers letters of support from clergy members and an abortion fund to assist those who cannot afford the procedure. NMRCRC also lists volunteer opportunities such as being a legal observer or a patient host and/or driver.
  • Ohio's independent RCRC website is called Ohio Religious Coalition for Reproductive Choice. The organization focuses on the same general goals as its parent organization as well as bringing local issues to light. Ohio RCRC highlights issues such as sidewalk harassment (people against abortion lobbying outside of abortion clinics) and provides resources such as abortion care packages and patient advocates.
  • The Pennsylvania RCRC website includes a list of clergy members from all across the state who support the organization. It also provides a link to the parent website and several songs and prayers for reproductive justice.
  • Wisconsin's RCRC does not have a website, but its supporters have organized a Facebook group titled "Wisconsin Religious Coalition for Reproductive Choice". Here, members share articles that highlight issues that are of interest to the organization.
  • California, Connecticut, Illinois, Indiana, Nebraska, and Oklahoma also have coalitions, but these states do not have websites or Facebook groups dedicated to their state coalitions.

Catholics for Choice

Catholics for Choice is a Catholic-run nonprofit that supports reproductive freedom, including abortion rights. The organization asserts that abortion is a matter that should be decided by the individual's own conscience. Members of the organization push for abortion services to be available to everyone so people have the ability to make that choice. The organization also asserts that it is a part of the majority of the faithful within the Catholic church that view abortion as a social justice value, not a religious value.

Deep learning

From Wikipedia, the free encyclopedia

Representing images on multiple layers of abstraction in deep learning
Representing images on multiple layers of abstraction in deep learning
 

Deep learning (also known as deep structured learning) is part of a broader family of machine learning methods based on artificial neural networks with representation learning. Learning can be supervised, semi-supervised or unsupervised.

Deep-learning architectures such as deep neural networks, deep belief networks, deep reinforcement learning, recurrent neural networks and convolutional neural networks have been applied to fields including computer vision, speech recognition, natural language processing, machine translation, bioinformatics, drug design, medical image analysis, climate science, material inspection and board game programs, where they have produced results comparable to and in some cases surpassing human expert performance.

Artificial neural networks (ANNs) were inspired by information processing and distributed communication nodes in biological systems. ANNs have various differences from biological brains. Specifically, artificial neural networks tend to be static and symbolic, while the biological brain of most living organisms is dynamic (plastic) and analogue.

The adjective "deep" in deep learning refers to the use of multiple layers in the network. Early work showed that a linear perceptron cannot be a universal classifier, but that a network with a nonpolynomial activation function with one hidden layer of unbounded width can. Deep learning is a modern variation which is concerned with an unbounded number of layers of bounded size, which permits practical application and optimized implementation, while retaining theoretical universality under mild conditions. In deep learning the layers are also permitted to be heterogeneous and to deviate widely from biologically informed connectionist models, for the sake of efficiency, trainability and understandability, whence the "structured" part.

Definition

Deep learning is a class of machine learning algorithms that uses multiple layers to progressively extract higher-level features from the raw input. For example, in image processing, lower layers may identify edges, while higher layers may identify the concepts relevant to a human such as digits or letters or faces.

Overview

Most modern deep learning models are based on artificial neural networks, specifically convolutional neural networks (CNN)s, although they can also include propositional formulas or latent variables organized layer-wise in deep generative models such as the nodes in deep belief networks and deep Boltzmann machines.

In deep learning, each level learns to transform its input data into a slightly more abstract and composite representation. In an image recognition application, the raw input may be a matrix of pixels; the first representational layer may abstract the pixels and encode edges; the second layer may compose and encode arrangements of edges; the third layer may encode a nose and eyes; and the fourth layer may recognize that the image contains a face. Importantly, a deep learning process can learn which features to optimally place in which level on its own. This does not eliminate the need for hand-tuning; for example, varying numbers of layers and layer sizes can provide different degrees of abstraction.

The word "deep" in "deep learning" refers to the number of layers through which the data is transformed. More precisely, deep learning systems have a substantial credit assignment path (CAP) depth. The CAP is the chain of transformations from input to output. CAPs describe potentially causal connections between input and output. For a feedforward neural network, the depth of the CAPs is that of the network and is the number of hidden layers plus one (as the output layer is also parameterized). For recurrent neural networks, in which a signal may propagate through a layer more than once, the CAP depth is potentially unlimited. No universally agreed-upon threshold of depth divides shallow learning from deep learning, but most researchers agree that deep learning involves CAP depth higher than 2. CAP of depth 2 has been shown to be a universal approximator in the sense that it can emulate any function. Beyond that, more layers do not add to the function approximator ability of the network. Deep models (CAP > 2) are able to extract better features than shallow models and hence, extra layers help in learning the features effectively.

Deep learning architectures can be constructed with a greedy layer-by-layer method. Deep learning helps to disentangle these abstractions and pick out which features improve performance.

For supervised learning tasks, deep learning methods eliminate feature engineering, by translating the data into compact intermediate representations akin to principal components, and derive layered structures that remove redundancy in representation.

Deep learning algorithms can be applied to unsupervised learning tasks. This is an important benefit because unlabeled data are more abundant than the labeled data. Examples of deep structures that can be trained in an unsupervised manner are deep belief networks.

Interpretations

Deep neural networks are generally interpreted in terms of the universal approximation theorem or probabilistic inference.

The classic universal approximation theorem concerns the capacity of feedforward neural networks with a single hidden layer of finite size to approximate continuous functions. In 1989, the first proof was published by George Cybenko for sigmoid activation functions and was generalised to feed-forward multi-layer architectures in 1991 by Kurt Hornik. Recent work also showed that universal approximation also holds for non-bounded activation functions such as the rectified linear unit.

The universal approximation theorem for deep neural networks concerns the capacity of networks with bounded width but the depth is allowed to grow. Lu et al. proved that if the width of a deep neural network with ReLU activation is strictly larger than the input dimension, then the network can approximate any Lebesgue integrable function; If the width is smaller or equal to the input dimension, then a deep neural network is not a universal approximator.

The probabilistic interpretation derives from the field of machine learning. It features inference, as well as the optimization concepts of training and testing, related to fitting and generalization, respectively. More specifically, the probabilistic interpretation considers the activation nonlinearity as a cumulative distribution function. The probabilistic interpretation led to the introduction of dropout as regularizer in neural networks. The probabilistic interpretation was introduced by researchers including Hopfield, Widrow and Narendra and popularized in surveys such as the one by Bishop.

History

Some sources point out that Frank Rosenblatt developed and explored all of the basic ingredients of the deep learning systems of today. He described it in his book "Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms", published by Cornell Aeronautical Laboratory, Inc., Cornell University in 1962.

The first general, working learning algorithm for supervised, deep, feedforward, multilayer perceptrons was published by Alexey Ivakhnenko and Lapa in 1967. A 1971 paper described a deep network with eight layers trained by the group method of data handling. Other deep learning working architectures, specifically those built for computer vision, began with the Neocognitron introduced by Kunihiko Fukushima in 1980.

The term Deep Learning was introduced to the machine learning community by Rina Dechter in 1986, and to artificial neural networks by Igor Aizenberg and colleagues in 2000, in the context of Boolean threshold neurons.

In 1989, Yann LeCun et al. applied the standard backpropagation algorithm, which had been around as the reverse mode of automatic differentiation since 1970, to a deep neural network with the purpose of recognizing handwritten ZIP codes on mail. While the algorithm worked, training required 3 days.

In 1994, André de Carvalho, together with Mike Fairhurst and David Bisset, published experimental results of a multi-layer boolean neural network, also known as a weightless neural network, composed of a 3-layers self-organising feature extraction neural network module (SOFT) followed by a multi-layer classification neural network module (GSN), which were independently trained. Each layer in the feature extraction module extracted features with growing complexity regarding the previous layer.

In 1995, Brendan Frey demonstrated that it was possible to train (over two days) a network containing six fully connected layers and several hundred hidden units using the wake-sleep algorithm, co-developed with Peter Dayan and Hinton. Many factors contribute to the slow speed, including the vanishing gradient problem analyzed in 1991 by Sepp Hochreiter.

Since 1997, Sven Behnke extended the feed-forward hierarchical convolutional approach in the Neural Abstraction Pyramid by lateral and backward connections in order to flexibly incorporate context into decisions and iteratively resolve local ambiguities.

Simpler models that use task-specific handcrafted features such as Gabor filters and support vector machines (SVMs) were a popular choice in the 1990s and 2000s, because of artificial neural network's (ANN) computational cost and a lack of understanding of how the brain wires its biological networks.

Both shallow and deep learning (e.g., recurrent nets) of ANNs have been explored for many years. These methods never outperformed non-uniform internal-handcrafting Gaussian mixture model/Hidden Markov model (GMM-HMM) technology based on generative models of speech trained discriminatively. Key difficulties have been analyzed, including gradient diminishing and weak temporal correlation structure in neural predictive models. Additional difficulties were the lack of training data and limited computing power.

Most speech recognition researchers moved away from neural nets to pursue generative modeling. An exception was at SRI International in the late 1990s. Funded by the US government's NSA and DARPA, SRI studied deep neural networks in speech and speaker recognition. The speaker recognition team led by Larry Heck reported significant success with deep neural networks in speech processing in the 1998 National Institute of Standards and Technology Speaker Recognition evaluation. The SRI deep neural network was then deployed in the Nuance Verifier, representing the first major industrial application of deep learning.

The principle of elevating "raw" features over hand-crafted optimization was first explored successfully in the architecture of deep autoencoder on the "raw" spectrogram or linear filter-bank features in the late 1990s, showing its superiority over the Mel-Cepstral features that contain stages of fixed transformation from spectrograms. The raw features of speech, waveforms, later produced excellent larger-scale results.

Many aspects of speech recognition were taken over by a deep learning method called long short-term memory (LSTM), a recurrent neural network published by Hochreiter and Schmidhuber in 1997. LSTM RNNs avoid the vanishing gradient problem and can learn "Very Deep Learning" tasks that require memories of events that happened thousands of discrete time steps before, which is important for speech. In 2003, LSTM started to become competitive with traditional speech recognizers on certain tasks. Later it was combined with connectionist temporal classification (CTC) in stacks of LSTM RNNs. In 2015, Google's speech recognition reportedly experienced a dramatic performance jump of 49% through CTC-trained LSTM, which they made available through Google Voice Search.

In 2006, publications by Geoff Hinton, Ruslan Salakhutdinov, Osindero and Teh showed how a many-layered feedforward neural network could be effectively pre-trained one layer at a time, treating each layer in turn as an unsupervised restricted Boltzmann machine, then fine-tuning it using supervised backpropagation. The papers referred to learning for deep belief nets.

Deep learning is part of state-of-the-art systems in various disciplines, particularly computer vision and automatic speech recognition (ASR). Results on commonly used evaluation sets such as TIMIT (ASR) and MNIST (image classification), as well as a range of large-vocabulary speech recognition tasks have steadily improved. Convolutional neural networks (CNNs) were superseded for ASR by CTC for LSTM. but are more successful in computer vision.

The impact of deep learning in industry began in the early 2000s, when CNNs already processed an estimated 10% to 20% of all the checks written in the US, according to Yann LeCun. Industrial applications of deep learning to large-scale speech recognition started around 2010.

The 2009 NIPS Workshop on Deep Learning for Speech Recognition was motivated by the limitations of deep generative models of speech, and the possibility that given more capable hardware and large-scale data sets that deep neural nets (DNN) might become practical. It was believed that pre-training DNNs using generative models of deep belief nets (DBN) would overcome the main difficulties of neural nets. However, it was discovered that replacing pre-training with large amounts of training data for straightforward backpropagation when using DNNs with large, context-dependent output layers produced error rates dramatically lower than then-state-of-the-art Gaussian mixture model (GMM)/Hidden Markov Model (HMM) and also than more-advanced generative model-based systems. The nature of the recognition errors produced by the two types of systems was characteristically different, offering technical insights into how to integrate deep learning into the existing highly efficient, run-time speech decoding system deployed by all major speech recognition systems. Analysis around 2009–2010, contrasting the GMM (and other generative speech models) vs. DNN models, stimulated early industrial investment in deep learning for speech recognition, eventually leading to pervasive and dominant use in that industry. That analysis was done with comparable performance (less than 1.5% in error rate) between discriminative DNNs and generative models.

In 2010, researchers extended deep learning from TIMIT to large vocabulary speech recognition, by adopting large output layers of the DNN based on context-dependent HMM states constructed by decision trees.

Advances in hardware have driven renewed interest in deep learning. In 2009, Nvidia was involved in what was called the “big bang” of deep learning, “as deep-learning neural networks were trained with Nvidia graphics processing units (GPUs).” That year, Andrew Ng determined that GPUs could increase the speed of deep-learning systems by about 100 times. In particular, GPUs are well-suited for the matrix/vector computations involved in machine learning. GPUs speed up training algorithms by orders of magnitude, reducing running times from weeks to days. Further, specialized hardware and algorithm optimizations can be used for efficient processing of deep learning models.

Deep learning revolution

How deep learning is a subset of machine learning and how machine learning is a subset of artificial intelligence (AI).

In 2012, a team led by George E. Dahl won the "Merck Molecular Activity Challenge" using multi-task deep neural networks to predict the biomolecular target of one drug. In 2014, Hochreiter's group used deep learning to detect off-target and toxic effects of environmental chemicals in nutrients, household products and drugs and won the "Tox21 Data Challenge" of NIH, FDA and NCATS.

Significant additional impacts in image or object recognition were felt from 2011 to 2012. Although CNNs trained by backpropagation had been around for decades, and GPU implementations of NNs for years, including CNNs, fast implementations of CNNs on GPUs were needed to progress on computer vision. In 2011, this approach achieved for the first time superhuman performance in a visual pattern recognition contest. Also in 2011, it won the ICDAR Chinese handwriting contest, and in May 2012, it won the ISBI image segmentation contest. Until 2011, CNNs did not play a major role at computer vision conferences, but in June 2012, a paper by Ciresan et al. at the leading conference CVPR showed how max-pooling CNNs on GPU can dramatically improve many vision benchmark records. In October 2012, a similar system by Krizhevsky et al. won the large-scale ImageNet competition by a significant margin over shallow machine learning methods. In November 2012, Ciresan et al.'s system also won the ICPR contest on analysis of large medical images for cancer detection, and in the following year also the MICCAI Grand Challenge on the same topic. In 2013 and 2014, the error rate on the ImageNet task using deep learning was further reduced, following a similar trend in large-scale speech recognition.

Image classification was then extended to the more challenging task of generating descriptions (captions) for images, often as a combination of CNNs and LSTMs.

Some researchers state that the October 2012 ImageNet victory anchored the start of a "deep learning revolution" that has transformed the AI industry.

In March 2019, Yoshua Bengio, Geoffrey Hinton and Yann LeCun were awarded the Turing Award for conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing.

Neural networks

Artificial neural networks

Artificial neural networks (ANNs) or connectionist systems are computing systems inspired by the biological neural networks that constitute animal brains. Such systems learn (progressively improve their ability) to do tasks by considering examples, generally without task-specific programming. For example, in image recognition, they might learn to identify images that contain cats by analyzing example images that have been manually labeled as "cat" or "no cat" and using the analytic results to identify cats in other images. They have found most use in applications difficult to express with a traditional computer algorithm using rule-based programming.

An ANN is based on a collection of connected units called artificial neurons, (analogous to biological neurons in a biological brain). Each connection (synapse) between neurons can transmit a signal to another neuron. The receiving (postsynaptic) neuron can process the signal(s) and then signal downstream neurons connected to it. Neurons may have state, generally represented by real numbers, typically between 0 and 1. Neurons and synapses may also have a weight that varies as learning proceeds, which can increase or decrease the strength of the signal that it sends downstream.

Typically, neurons are organized in layers. Different layers may perform different kinds of transformations on their inputs. Signals travel from the first (input), to the last (output) layer, possibly after traversing the layers multiple times.

The original goal of the neural network approach was to solve problems in the same way that a human brain would. Over time, attention focused on matching specific mental abilities, leading to deviations from biology such as backpropagation, or passing information in the reverse direction and adjusting the network to reflect that information.

Neural networks have been used on a variety of tasks, including computer vision, speech recognition, machine translation, social network filtering, playing board and video games and medical diagnosis.

As of 2017, neural networks typically have a few thousand to a few million units and millions of connections. Despite this number being several order of magnitude less than the number of neurons on a human brain, these networks can perform many tasks at a level beyond that of humans (e.g., recognizing faces, playing "Go").

Deep neural networks

A deep neural network (DNN) is an artificial neural network (ANN) with multiple layers between the input and output layers. There are different types of neural networks but they always consist of the same components: neurons, synapses, weights, biases, and functions. These components functioning similar to the human brains and can be trained like any other ML algorithm.

For example, a DNN that is trained to recognize dog breeds will go over the given image and calculate the probability that the dog in the image is a certain breed. The user can review the results and select which probabilities the network should display (above a certain threshold, etc.) and return the proposed label. Each mathematical manipulation as such is considered a layer, and complex DNN have many layers, hence the name "deep" networks.

DNNs can model complex non-linear relationships. DNN architectures generate compositional models where the object is expressed as a layered composition of primitives. The extra layers enable composition of features from lower layers, potentially modeling complex data with fewer units than a similarly performing shallow network. For instance, it was proved that sparse multivariate polynomials are exponentially easier to approximate with DNNs than with shallow networks.

Deep architectures include many variants of a few basic approaches. Each architecture has found success in specific domains. It is not always possible to compare the performance of multiple architectures, unless they have been evaluated on the same data sets.

DNNs are typically feedforward networks in which data flows from the input layer to the output layer without looping back. At first, the DNN creates a map of virtual neurons and assigns random numerical values, or "weights", to connections between them. The weights and inputs are multiplied and return an output between 0 and 1. If the network did not accurately recognize a particular pattern, an algorithm would adjust the weights. That way the algorithm can make certain parameters more influential, until it determines the correct mathematical manipulation to fully process the data.

Recurrent neural networks (RNNs), in which data can flow in any direction, are used for applications such as language modeling. Long short-term memory is particularly effective for this use.

Convolutional deep neural networks (CNNs) are used in computer vision. CNNs also have been applied to acoustic modeling for automatic speech recognition (ASR).

Challenges

As with ANNs, many issues can arise with naively trained DNNs. Two common issues are overfitting and computation time.

DNNs are prone to overfitting because of the added layers of abstraction, which allow them to model rare dependencies in the training data. Regularization methods such as Ivakhnenko's unit pruning or weight decay (-regularization) or sparsity (-regularization) can be applied during training to combat overfitting. Alternatively dropout regularization randomly omits units from the hidden layers during training. This helps to exclude rare dependencies. Finally, data can be augmented via methods such as cropping and rotating such that smaller training sets can be increased in size to reduce the chances of overfitting.

DNNs must consider many training parameters, such as the size (number of layers and number of units per layer), the learning rate, and initial weights. Sweeping through the parameter space for optimal parameters may not be feasible due to the cost in time and computational resources. Various tricks, such as batching (computing the gradient on several training examples at once rather than individual examples) speed up computation. Large processing capabilities of many-core architectures (such as GPUs or the Intel Xeon Phi) have produced significant speedups in training, because of the suitability of such processing architectures for the matrix and vector computations.

Alternatively, engineers may look for other types of neural networks with more straightforward and convergent training algorithms. CMAC (cerebellar model articulation controller) is one such kind of neural network. It doesn't require learning rates or randomized initial weights for CMAC. The training process can be guaranteed to converge in one step with a new batch of data, and the computational complexity of the training algorithm is linear with respect to the number of neurons involved.

Hardware

Since the 2010s, advances in both machine learning algorithms and computer hardware have led to more efficient methods for training deep neural networks that contain many layers of non-linear hidden units and a very large output layer. By 2019, graphic processing units (GPUs), often with AI-specific enhancements, had displaced CPUs as the dominant method of training large-scale commercial cloud AI. OpenAI estimated the hardware computation used in the largest deep learning projects from AlexNet (2012) to AlphaZero (2017), and found a 300,000-fold increase in the amount of computation required, with a doubling-time trendline of 3.4 months.

Special electronic circuits called deep learning processors were designed to speed up deep learning algorithms. Deep learning processors include neural processing units (NPUs) in Huawei cellphones and cloud computing servers such as tensor processing units (TPU) in the Google Cloud Platform.

Atomically thin semiconductors are considered promising for energy-efficient deep learning hardware where the same basic device structure is used for both logic operations and data storage. In 2020, Marega et al. published experiments with a large-area active channel material for developing logic-in-memory devices and circuits based on floating-gate field-effect transistors (FGFETs).

In 2021, J. Feldmann et al. proposed an integrated photonic hardware accelerator for parallel convolutional processing. The authors identify two key advantages of integrated photonics over its electronic counterparts: (1) massively parallel data transfer through wavelength division multiplexing in conjunction with frequency combs, and (2) extremely high data modulation speeds. Their system can execute trillions of multiply-accumulate operations per second, indicating the potential of integrated photonics in data-heavy AI applications.

Applications

Automatic speech recognition

Large-scale automatic speech recognition is the first and most convincing successful case of deep learning. LSTM RNNs can learn "Very Deep Learning" tasks that involve multi-second intervals containing speech events separated by thousands of discrete time steps, where one time step corresponds to about 10 ms. LSTM with forget gates is competitive with traditional speech recognizers on certain tasks.

The initial success in speech recognition was based on small-scale recognition tasks based on TIMIT. The data set contains 630 speakers from eight major dialects of American English, where each speaker reads 10 sentences. Its small size lets many configurations be tried. More importantly, the TIMIT task concerns phone-sequence recognition, which, unlike word-sequence recognition, allows weak phone bigram language models. This lets the strength of the acoustic modeling aspects of speech recognition be more easily analyzed. The error rates listed below, including these early results and measured as percent phone error rates (PER), have been summarized since 1991.

Method Percent phone
error rate (PER) (%)
Randomly Initialized RNN 26.1
Bayesian Triphone GMM-HMM 25.6
Hidden Trajectory (Generative) Model 24.8
Monophone Randomly Initialized DNN 23.4
Monophone DBN-DNN 22.4
Triphone GMM-HMM with BMMI Training 21.7
Monophone DBN-DNN on fbank 20.7
Convolutional DNN 20.0
Convolutional DNN w. Heterogeneous Pooling 18.7
Ensemble DNN/CNN/RNN 18.3
Bidirectional LSTM 17.8
Hierarchical Convolutional Deep Maxout Network 16.5

The debut of DNNs for speaker recognition in the late 1990s and speech recognition around 2009-2011 and of LSTM around 2003–2007, accelerated progress in eight major areas:

  • Scale-up/out and accelerated DNN training and decoding
  • Sequence discriminative training
  • Feature processing by deep models with solid understanding of the underlying mechanisms
  • Adaptation of DNNs and related deep models
  • Multi-task and transfer learning by DNNs and related deep models
  • CNNs and how to design them to best exploit domain knowledge of speech
  • RNN and its rich LSTM variants
  • Other types of deep models including tensor-based models and integrated deep generative/discriminative models.

All major commercial speech recognition systems (e.g., Microsoft Cortana, Xbox, Skype Translator, Amazon Alexa, Google Now, Apple Siri, Baidu and iFlyTek voice search, and a range of Nuance speech products, etc.) are based on deep learning.

Image recognition

A common evaluation set for image classification is the MNIST database data set. MNIST is composed of handwritten digits and includes 60,000 training examples and 10,000 test examples. As with TIMIT, its small size lets users test multiple configurations. A comprehensive list of results on this set is available.

Deep learning-based image recognition has become "superhuman", producing more accurate results than human contestants. This first occurred in 2011 in recognition of traffic signs, and in 2014, with recognition of human faces.

Deep learning-trained vehicles now interpret 360° camera views. Another example is Facial Dysmorphology Novel Analysis (FDNA) used to analyze cases of human malformation connected to a large database of genetic syndromes.

Visual art processing

Closely related to the progress that has been made in image recognition is the increasing application of deep learning techniques to various visual art tasks. DNNs have proven themselves capable, for example, of

  • identifying the style period of a given painting
  • Neural Style Transfer – capturing the style of a given artwork and applying it in a visually pleasing manner to an arbitrary photograph or video
  • generating striking imagery based on random visual input fields.

Natural language processing

Neural networks have been used for implementing language models since the early 2000s. LSTM helped to improve machine translation and language modeling.

Other key techniques in this field are negative sampling and word embedding. Word embedding, such as word2vec, can be thought of as a representational layer in a deep learning architecture that transforms an atomic word into a positional representation of the word relative to other words in the dataset; the position is represented as a point in a vector space. Using word embedding as an RNN input layer allows the network to parse sentences and phrases using an effective compositional vector grammar. A compositional vector grammar can be thought of as probabilistic context free grammar (PCFG) implemented by an RNN. Recursive auto-encoders built atop word embeddings can assess sentence similarity and detect paraphrasing. Deep neural architectures provide the best results for constituency parsing, sentiment analysis, information retrieval, spoken language understanding, machine translation, contextual entity linking, writing style recognition, Text classification and others.

Recent developments generalize word embedding to sentence embedding.

Google Translate (GT) uses a large end-to-end long short-term memory (LSTM) network. Google Neural Machine Translation (GNMT) uses an example-based machine translation method in which the system "learns from millions of examples." It translates "whole sentences at a time, rather than pieces. Google Translate supports over one hundred languages. The network encodes the "semantics of the sentence rather than simply memorizing phrase-to-phrase translations". GT uses English as an intermediate between most language pairs.

Drug discovery and toxicology

A large percentage of candidate drugs fail to win regulatory approval. These failures are caused by insufficient efficacy (on-target effect), undesired interactions (off-target effects), or unanticipated toxic effects. Research has explored use of deep learning to predict the biomolecular targets, off-targets, and toxic effects of environmental chemicals in nutrients, household products and drugs.

AtomNet is a deep learning system for structure-based rational drug design. AtomNet was used to predict novel candidate biomolecules for disease targets such as the Ebola virus and multiple sclerosis.

In 2017 graph neural networks were used for the first time to predict various properties of molecules in a large toxicology data set. In 2019, generative neural networks were used to produce molecules that were validated experimentally all the way into mice.

Customer relationship management

Deep reinforcement learning has been used to approximate the value of possible direct marketing actions, defined in terms of RFM variables. The estimated value function was shown to have a natural interpretation as customer lifetime value.

Recommendation systems

Recommendation systems have used deep learning to extract meaningful features for a latent factor model for content-based music and journal recommendations. Multi-view deep learning has been applied for learning user preferences from multiple domains. The model uses a hybrid collaborative and content-based approach and enhances recommendations in multiple tasks.

Bioinformatics

An autoencoder ANN was used in bioinformatics, to predict gene ontology annotations and gene-function relationships.

In medical informatics, deep learning was used to predict sleep quality based on data from wearables and predictions of health complications from electronic health record data.

Medical image analysis

Deep learning has been shown to produce competitive results in medical application such as cancer cell classification, lesion detection, organ segmentation and image enhancement. Modern deep learning tools demonstrate the high accuracy of detecting various diseases and the helpfulness of their use by specialists to improve the diagnosis efficiency.

Mobile advertising

Finding the appropriate mobile audience for mobile advertising is always challenging, since many data points must be considered and analyzed before a target segment can be created and used in ad serving by any ad server. Deep learning has been used to interpret large, many-dimensioned advertising datasets. Many data points are collected during the request/serve/click internet advertising cycle. This information can form the basis of machine learning to improve ad selection.

Image restoration

Deep learning has been successfully applied to inverse problems such as denoising, super-resolution, inpainting, and film colorization. These applications include learning methods such as "Shrinkage Fields for Effective Image Restoration" which trains on an image dataset, and Deep Image Prior, which trains on the image that needs restoration.

Financial fraud detection

Deep learning is being successfully applied to financial fraud detection, tax evasion detection, and anti-money laundering.

Military

The United States Department of Defense applied deep learning to train robots in new tasks through observation.

Partial differential equations

Physics informed neural networks have been used to solve partial differential equations in both forward and inverse problems in a data driven manner. One example is the reconstructing fluid flow governed by the Navier-Stokes equations. Using physics informed neural networks does not require the often expensive mesh generation that conventional CFD methods relies on.

Image Reconstruction

Image reconstruction is the reconstruction of the underlying images from the image-related measurements. Several works showed the better and superior performance of the deep learning methods compared to analytical methods for various applications, e.g., spectral imaging and ultrasound imaging.

Relation to human cognitive and brain development

Deep learning is closely related to a class of theories of brain development (specifically, neocortical development) proposed by cognitive neuroscientists in the early 1990s. These developmental theories were instantiated in computational models, making them predecessors of deep learning systems. These developmental models share the property that various proposed learning dynamics in the brain (e.g., a wave of nerve growth factor) support the self-organization somewhat analogous to the neural networks utilized in deep learning models. Like the neocortex, neural networks employ a hierarchy of layered filters in which each layer considers information from a prior layer (or the operating environment), and then passes its output (and possibly the original input), to other layers. This process yields a self-organizing stack of transducers, well-tuned to their operating environment. A 1995 description stated, "...the infant's brain seems to organize itself under the influence of waves of so-called trophic-factors ... different regions of the brain become connected sequentially, with one layer of tissue maturing before another and so on until the whole brain is mature."

A variety of approaches have been used to investigate the plausibility of deep learning models from a neurobiological perspective. On the one hand, several variants of the backpropagation algorithm have been proposed in order to increase its processing realism. Other researchers have argued that unsupervised forms of deep learning, such as those based on hierarchical generative models and deep belief networks, may be closer to biological reality. In this respect, generative neural network models have been related to neurobiological evidence about sampling-based processing in the cerebral cortex.

Although a systematic comparison between the human brain organization and the neuronal encoding in deep networks has not yet been established, several analogies have been reported. For example, the computations performed by deep learning units could be similar to those of actual neurons and neural populations. Similarly, the representations developed by deep learning models are similar to those measured in the primate visual system both at the single-unit and at the population levels.

Commercial activity

Facebook's AI lab performs tasks such as automatically tagging uploaded pictures with the names of the people in them.

Google's DeepMind Technologies developed a system capable of learning how to play Atari video games using only pixels as data input. In 2015 they demonstrated their AlphaGo system, which learned the game of Go well enough to beat a professional Go player. Google Translate uses a neural network to translate between more than 100 languages.

In 2017, Covariant.ai was launched, which focuses on integrating deep learning into factories.

As of 2008, researchers at The University of Texas at Austin (UT) developed a machine learning framework called Training an Agent Manually via Evaluative Reinforcement, or TAMER, which proposed new methods for robots or computer programs to learn how to perform tasks by interacting with a human instructor. First developed as TAMER, a new algorithm called Deep TAMER was later introduced in 2018 during a collaboration between U.S. Army Research Laboratory (ARL) and UT researchers. Deep TAMER used deep learning to provide a robot the ability to learn new tasks through observation. Using Deep TAMER, a robot learned a task with a human trainer, watching video streams or observing a human perform a task in-person. The robot later practiced the task with the help of some coaching from the trainer, who provided feedback such as “good job” and “bad job.”

Criticism and comment

Deep learning has attracted both criticism and comment, in some cases from outside the field of computer science.

Theory

A main criticism concerns the lack of theory surrounding some methods. Learning in the most common deep architectures is implemented using well-understood gradient descent. However, the theory surrounding other algorithms, such as contrastive divergence is less clear. (e.g., Does it converge? If so, how fast? What is it approximating?) Deep learning methods are often looked at as a black box, with most confirmations done empirically, rather than theoretically.

Others point out that deep learning should be looked at as a step towards realizing strong AI, not as an all-encompassing solution. Despite the power of deep learning methods, they still lack much of the functionality needed for realizing this goal entirely. Research psychologist Gary Marcus noted:

"Realistically, deep learning is only part of the larger challenge of building intelligent machines. Such techniques lack ways of representing causal relationships (...) have no obvious ways of performing logical inferences, and they are also still a long way from integrating abstract knowledge, such as information about what objects are, what they are for, and how they are typically used. The most powerful A.I. systems, like Watson (...) use techniques like deep learning as just one element in a very complicated ensemble of techniques, ranging from the statistical technique of Bayesian inference to deductive reasoning."

In further reference to the idea that artistic sensitivity might be inherent in relatively low levels of the cognitive hierarchy, a published series of graphic representations of the internal states of deep (20-30 layers) neural networks attempting to discern within essentially random data the images on which they were trained demonstrate a visual appeal: the original research notice received well over 1,000 comments, and was the subject of what was for a time the most frequently accessed article on The Guardian's website.

Errors

Some deep learning architectures display problematic behaviors, such as confidently classifying unrecognizable images as belonging to a familiar category of ordinary images (2014) and misclassifying minuscule perturbations of correctly classified images (2013). Goertzel hypothesized that these behaviors are due to limitations in their internal representations and that these limitations would inhibit integration into heterogeneous multi-component artificial general intelligence (AGI) architectures. These issues may possibly be addressed by deep learning architectures that internally form states homologous to image-grammar decompositions of observed entities and events. Learning a grammar (visual or linguistic) from training data would be equivalent to restricting the system to commonsense reasoning that operates on concepts in terms of grammatical production rules and is a basic goal of both human language acquisition and artificial intelligence (AI).

Cyber threat

As deep learning moves from the lab into the world, research and experience show that artificial neural networks are vulnerable to hacks and deception. By identifying patterns that these systems use to function, attackers can modify inputs to ANNs in such a way that the ANN finds a match that human observers would not recognize. For example, an attacker can make subtle changes to an image such that the ANN finds a match even though the image looks to a human nothing like the search target. Such manipulation is termed an “adversarial attack.”

In 2016 researchers used one ANN to doctor images in trial and error fashion, identify another's focal points and thereby generate images that deceived it. The modified images looked no different to human eyes. Another group showed that printouts of doctored images then photographed successfully tricked an image classification system. One defense is reverse image search, in which a possible fake image is submitted to a site such as TinEye that can then find other instances of it. A refinement is to search using only parts of the image, to identify images from which that piece may have been taken.

Another group showed that certain psychedelic spectacles could fool a facial recognition system into thinking ordinary people were celebrities, potentially allowing one person to impersonate another. In 2017 researchers added stickers to stop signs and caused an ANN to misclassify them.

ANNs can however be further trained to detect attempts at deception, potentially leading attackers and defenders into an arms race similar to the kind that already defines the malware defense industry. ANNs have been trained to defeat ANN-based anti-malware software by repeatedly attacking a defense with malware that was continually altered by a genetic algorithm until it tricked the anti-malware while retaining its ability to damage the target.

In 2016, another group demonstrated that certain sounds could make the Google Now voice command system open a particular web address, and hypothesized that this could "serve as a stepping stone for further attacks (e.g., opening a web page hosting drive-by malware)."

In “data poisoning,” false data is continually smuggled into a machine learning system's training set to prevent it from achieving mastery.

Reliance on human microwork

Most Deep Learning systems rely on training and verification data that is generated and/or annotated by humans. It has been argued in media philosophy that not only low-paid clickwork (e.g. on Amazon Mechanical Turk) is regularly deployed for this purpose, but also implicit forms of human microwork that are often not recognized as such. The philosopher Rainer Mühlhoff distinguishes five types of "machinic capture" of human microwork to generate training data: (1) gamification (the embedding of annotation or computation tasks in the flow of a game), (2) "trapping and tracking" (e.g. CAPTCHAs for image recognition or click-tracking on Google search results pages), (3) exploitation of social motivations (e.g. tagging faces on Facebook to obtain labeled facial images), (4) information mining (e.g. by leveraging quantified-self devices such as activity trackers) and (5) clickwork.

Mühlhoff argues that in most commercial end-user applications of Deep Learning such as Facebook's face recognition system, the need for training data does not stop once an ANN is trained. Rather, there is a continued demand for human-generated verification data to constantly calibrate and update the ANN. For this purpose Facebook introduced the feature that once a user is automatically recognized in an image, they receive a notification. They can choose whether of not they like to be publicly labeled on the image, or tell Facebook that it is not them in the picture. This user interface is a mechanism to generate "a constant stream of verification data" to further train the network in real-time. As Mühlhoff argues, involvement of human users to generate training and verification data is so typical for most commercial end-user applications of Deep Learning that such systems may be referred to as "human-aided artificial intelligence".

Delayed-choice quantum eraser

From Wikipedia, the free encyclopedia https://en.wikipedia.org/wiki/Delayed-choice_quantum_eraser A delayed-cho...