Extended evolutionary synthesis

From Wikipedia, the free encyclopedia

https://en.wikipedia.org/wiki/Extended_evolutionary_synthesis 

The Extended Evolutionary Synthesis (EES) consists of a set of theoretical concepts argued to be more comprehensive than the earlier modern synthesis of evolutionary biology that took place between 1918 and 1942. The extended evolutionary synthesis was called for in the 1950s by C. H. Waddington, argued for on the basis of punctuated equilibrium by Stephen Jay Gould and Niles Eldredge in the 1980s, and was reconceptualized in 2007 by Massimo Pigliucci and Gerd B. Müller.

The extended evolutionary synthesis revisits the relative importance of different factors at play, examining several assumptions of the earlier synthesis, and augmenting it with additional causative factors. It includes multilevel selection, transgenerational epigenetic inheritance, niche construction, evolvability, and several concepts from evolutionary developmental biology.

Not all biologists have agreed on the need for, or the scope of, an extended synthesis. Many have collaborated on another synthesis in evolutionary developmental biology, which concentrates on developmental molecular genetics and evolution to understand how natural selection operated on developmental processes and deep homologies between organisms at the level of highly conserved genes.

The preceding "modern synthesis"

Several major ideas about evolution came together in the population genetics of the early 20th century to form the modern synthesis, including genetic variation, natural selection, and particulate (Mendelian) inheritance. This ended the eclipse of Darwinism and supplanted a variety of non-Darwinian theories of evolution. However, it did not unify all of biology, omitting sciences such as developmental biology.

Main article: Modern synthesis (20th century)

The modern synthesis was the widely accepted early-20th-century synthesis reconciling Charles Darwin's theory of evolution by natural selection and Gregor Mendel's theory of genetics in a joint mathematical framework. It established evolution as biology's central paradigm. The 19th-century ideas of natural selection by Darwin and Mendelian genetics were united by researchers who included Ronald Fisher, J. B. S. Haldane and Sewall Wright, the three founders of population genetics, between 1918 and 1932. Julian Huxley introduced the phrase "modern synthesis" in his 1942 book, Evolution: The Modern Synthesis.

Early history

During the 1950s, English biologist C. H. Waddington called for an extended synthesis based on his research on epigenetics and genetic assimilation.

In 1978, Michael J. D. White wrote about an extension of the modern synthesis based on new research from speciation. In the 1980s, entomologist Ryuichi Matsuda coined the term "pan-environmentalism" as an extended evolutionary synthesis which he saw as a fusion of Darwinism with neo-Lamarckism. He held that heterochrony is a main mechanism for evolutionary change and that novelty in evolution can be generated by genetic assimilation. An extended synthesis was also proposed by the Austrian zoologist Rupert Riedl, with the study of evolvability.

Gordon Rattray Taylor in his 1983 book The Great Evolution Mystery called for an extended synthesis, noting that the modern synthesis is only a subsection of a more comprehensive explanation for biological evolution still to be formulated. In 1985, biologist Robert G. B. Reid authored Evolutionary Theory: The Unfinished Synthesis, which argued that the modern synthesis with its emphasis on natural selection is an incomplete picture of evolution, and emergent evolution can explain the origin of genetic variation.

In 1988, ethologist John Endler wrote about developing a newer synthesis, discussing processes of evolution that he felt had been neglected.

In 2000, Robert L. Carroll called for an "expanded evolutionary synthesis" due to new research from molecular developmental biology, systematics, geology and the fossil record.

Punctuated equilibrium

In the 1980s, the American palaeontologists Stephen Jay Gould and Niles Eldredge argued for an extended synthesis based on their idea of punctuated equilibrium, the role of species selection shaping large scale evolutionary patterns and natural selection working on multiple levels extending from genes to species.

Contributions from evolutionary developmental biology

Main article: Evolutionary developmental biology

Some researchers in the field of evolutionary developmental biology proposed another synthesis. They argue that the modern and extended syntheses should mostly center on genes and suggest an integration of embryology with molecular genetics and evolution, aiming to understand how natural selection operates on gene regulation and deep homologies between organisms at the level of highly conserved genes, transcription factors and signalling pathways. By contrast, a different strand of evo-devo following an organismal approach contributes to the extended synthesis by emphasizing (amongst others) developmental bias (both through facilitation and constraint), evolvability, and inherency of form  as primary factors in the evolution of complex structures and phenotypic novelties.

Recent history

Massimo Pigliucci, a leading proponent of the extended evolutionary synthesis in its 2007 form

The idea of an extended synthesis was relaunched in 2007 by Massimo Pigliucci, and Gerd B. Müller, with a book in 2010 titled Evolution: The Extended Synthesis, which has served as a launching point for work on the extended synthesis. This includes:

• The role of prior configurations, genomic structures, and other traits in the organism in generating evolutionary variations.
• How increasing dimensionality of fitness landscapes affects our view of speciation.
• The role of multilevel selection in the major evolutionary transitions.
• New types of inheritance, including cultural and epigenetic inheritance.
• The way that organismal development and developmental plasticity channel evolutionary pathways and generates phenotypic novelty
• How organisms modify the environments they belong to through niche construction.

Other processes such as evolvability, phenotypic plasticity, reticulate evolution, horizontal gene transfer, symbiogenesis are said by proponents to have been excluded or missed from the modern synthesis. The goal of Piglucci's and Müller's extended synthesis is to take evolution beyond the gene-centered approach of population genetics to consider more organism- and ecology-centered approaches. Many of these causes are currently considered secondary in evolutionary causation, and proponents of the extended synthesis want them to be considered first-class evolutionary causes.

Michael R. Rose and Todd Oakley have called for a postmodern synthesis, they commented that "it is now abundantly clear that living things often attain a degree of genomic complexity far beyond simple models like the "gene library" genome of the Modern Synthesis". Biologist Eugene Koonin has suggested that the gradualism of the modern synthesis is unsustainable as gene duplication, horizontal gene transfer and endosymbiosis play a pivotal role in evolution. Koonin commented that "the new developments in evolutionary biology by no account should be viewed as refutation of Darwin. On the contrary, they are widening the trails that Darwin blazed 150 years ago and reveal the extraordinary fertility of his thinking."

Arlin Stoltzfus and colleagues advocate mutational and developmental bias in the introduction of variation as an important source of orientation or direction in evolutionary change. They argue that bias in the introduction of variation was not formally recognized throughout the 20th century, due to the influence of neo-Darwinism on thinking about causation.

Organism-centered evolution

The early biologists of the organicist movement have influenced the modern extended evolutionary synthesis. Recent research has called for expanding the population genetic framework of evolutionary biology by a more organism-centered perspective. This has been described as "organism-centered evolution" which looks beyond the genome to the ways that individual organisms are participants in their own evolution. Philip Ball has written a research review on organism-centered evolution.

In her book from 2003, Developmental Plasticity and Evolution, Mary Jane West-Eberhard showed that biological theory cannot make sense of evolution if it lacks a theory of the phenotype. Only when we have understood the phenotype and its responsiveness to the environment (or phenotypic plasticity), she says, can we understand its development, selection, and evolution—and the place of genes therein: 'Why start with the phenotype and its development? Because that is where evolution starts.' Likewise, development itself depends at every step 'on the pre-existent structure of the phenotype.' Crucially for West-Eberhard, like Darwin before her, phenotypic plasticity includes the agency of organisms (as well as ontogenetic and epigenetic flexibility) because it is as agents that organisms participate in the struggle for existence from which natural selection results: "genes are usually followers, not leaders, in evolutionary change."

Rui Diogo has proposed a revision of evolutionary theory, which he has termed ONCE: Organic Nonoptimal Constrained Evolution. According to ONCE, evolution is mainly driven by the behavioural choices and persistence of organisms themselves, whilst natural selection plays a secondary role. ONCE cites examples of reciprocal causation between organism and the environment, Baldwin effect, organic selection, developmental bias and niche construction.

Predictions

The extended synthesis is characterized by its additional set of predictions that differ from the standard modern synthesis theory:

1. Change in phenotype can precede change in genotype
2. Changes in phenotype are predominantly positive, rather than neutral (see: neutral theory of molecular evolution)
3. Changes in phenotype are induced in many organisms, rather than one organism
4. Revolutionary change in phenotype can occur through mutation, facilitated variation or threshold events
5. Repeated evolution in isolated populations can be by convergent evolution or developmental bias
6. Adaptation can be caused by natural selection, environmental induction, non-genetic inheritance, learning and cultural transmission (see: Baldwin effect, meme, transgenerational epigenetic inheritance, ecological inheritance, non-Mendelian inheritance)
7. Rapid evolution can result from simultaneous induction, natural selection and developmental dynamics
8. Biodiversity can be affected by features of developmental systems such as differences in evolvability
9. Heritable variation is directed towards variants that are adaptive and integrated with phenotype
10. Niche construction is biased towards environmental changes that suit the constructor's phenotype, or that of its descendants, and enhance their fitness
11. Kin selection
12. Multilevel selection
13. Self-organization
14. Symbiogenesis

Testing

From 2016 to 2019, there was an organized project entitled "Putting The Extended Evolutionary Synthesis To The Test" supported by a 7.5 million USD grant from the John Templeton Foundation, supplemented with further money from participating instutitions including Clark University, Indiana University, Lund University, Stanford University, University of Southampton and University of St Andrews.

Publications from the project include over 200 papers, a special issue, and an anthology on Evolutionary Causation. In 2019 a final report of the 2016–2019 consortium was published, Putting the Extended Evolutionary Synthesis to the Test.

The project was headed by Kevin N. Laland at the University of St Andrews and Tobias Uller at Lund University. According to Laland what the extended synthesis "really boils down to is recognition that, in addition to selection, drift, mutation and other established evolutionary processes, other factors, particularly developmental influences, shape the evolutionary process in important ways."

Status

Biologists disagree on the need for an extended synthesis. Opponents contend that the modern synthesis is able to fully account for the newer observations, whereas others criticize the extended synthesis for not being radical enough. Proponents think that the conceptions of evolution at the core of the modern synthesis are too narrow and that even when the modern synthesis allows for the ideas in the extended synthesis, using the modern synthesis affects the way that biologists think about evolution. For example, Denis Noble says that using terms and categories of the modern synthesis distorts the picture of biology that modern experimentation has discovered. Proponents therefore claim that the extended synthesis is necessary to help expand the conceptions and framework of how evolution is considered throughout the biological disciplines. In 2022, the John Templeton Foundation published a review of recent literature.

Human Brain Project

From Wikipedia, the free encyclopedia

https://en.wikipedia.org/wiki/Human_Brain_Project 

 

Human Brain Project
Type of project	Scientific research
Location	Geneva, Switzerland
Owner	European Union
Key people	Paweł Świeboda, Director General Katrin Amunts, Scientific Research Director
Established	1 October 2013; 12 years ago
Status	Ended 30 September 2023; 2 years ago
Website	humanbrainproject.eu

The Human Brain Project (HBP) was a EU scientific research project that ran for ten years from 2013 to 2023, with a total budget of €1 billion. Using high-performance exascale supercomputers it built infrastructure that allowed researchers to advance knowledge in the fields of neuroscience, computing and brain-related medicine. Its successor was the European Brain Research Infrastructures (EBRAINS) project.

The Project, which started on 1 October 2013, was a European Commission Future and Emerging Technologies Flagship. The HBP was coordinated by the École Polytechnique Fédérale de Lausanne and was largely funded by the European Union. The project coordination office was in Geneva, Switzerland.

Peer-reviewed research finds that the public discussion forum (the Human Brain Project forum) was actively utilized and showed resilience during the COVID-19 pandemic. The HBP forum has been most actively utilized and useful for solving questions related to programming issues and questions close to HBP core areas.

Strategic goals and organisation

The 2013 HBP Summit–the inauguration of the Project–took place in the EPFL Learning Centre in October 2013. It brought together scientists from over 100 Partner Institutions.

Fundamental to the HBP approach is to investigate the brain on different spatial and temporal scales (i.e. from the molecular to the large networks underlying higher cognitive processes, and from milliseconds to years). To achieve this goal, the HBP relies on the collaboration of scientists from diverse disciplines, including neuroscience, philosophy and computer science, to take advantage of the loop of experimental data, modelling theories and simulations. The idea is that empirical results are used to develop theories, which then foster modelling and simulations which result in predictions that are in turn verified by empirical results.

The primary objective of the HBP is to create an ICT-based research infrastructure for brain research, cognitive neuroscience and brain-inspired computing, which can be used by researchers world-wide.

The Project is divided into 12 Subprojects. Six of these develop ICT-based platforms (Subprojects 5-10), which consist of prototype hardware, software, databases, and programming interfaces. These tools are available to researchers worldwide via the HBP Collaboratory. Four Subprojects gather data on empirical neuroscience and establish theoretical foundations (Subprojects 1–4) and one is responsible for ethics and society (Subproject 12). Subproject 11 coordinates the project.

• SP1 Mouse Brain Organisation: Understanding the structure of the mouse brain, and its electrical and chemical functions
• SP2 Human Brain Organisation: Understanding the structure of the human brain, and its electrical and chemical functions
• SP3 Systems and Cognitive Neuroscience: Understanding how the brain performs its systems-level and cognitive functional activities
• SP4 Theoretical Neuroscience: Deriving high-level mathematical models to synthesize conclusions from research data
• SP5 Neuroinformatics Platform: Gathering, organising and making available brain data
• SP6 Brain Simulation Platform: Developing data-driven reconstructions of brain tissue and simulation capabilities to explore these reconstructions
• SP7 High-performance Analytics and Computing Platform: Providing the ICT capability to map the brain in unprecedented detail, construct complex models, run large simulations, and analyse large volumes of data
• SP8 Medical Informatics Platform: Developing the infrastructure to share hospital and medical research data for the purpose of understanding disease clusters and their respective disease signatures
• SP9 Neuromorphic Computing Platform: Developing and applying brain-inspired computing technology
• SP10 Neurorobotics Platform: Developing virtual and real robots and environments for testing brain simulations
• SP11 Management and Coordination: General coordination of the project
• SP12 Ethics and Society: Exploring the ethical and societal impact of HBP's work

The HBP is coordinated by the École Polytechnique Fédérale de Lausanne and involves researchers from over 117 partner institutions in 19 countries across Europe. Notable Partner Institutions include the University of Heidelberg, Forschungszentrum Jülich, and the University Hospital of Lausanne.

The scientific direction is provided by representatives from each of the HBP's Subprojects, which form the Science and Infrastructure Board (SIB). Katrin Amunts from Forschungszentrum Jülich is the Chair of the SIB. Alois Knoll from TU Munich is Vice Chair of the SIB for software. The Directorate steers the daily work of the HBP – it is led by Andreas Mortensen from EPFL.

Funding

The HBP is funded by the European Commission Directorate General for Communications Networks, Content and Technology (DG CONNECT) under the FP7 framework, an EU Research and Innovation funding programme. It was one of the two initial Future Emerging Technologies (FET) Flagship projects.

The project is split into five phases, each supplied with separate funding. The call for funding for the Project's initial two-and-a-half-year 'Ramp-Up Phase' of EUR 54 million closed in November 2013 and the results were announced in March 2014. Twenty-two projects from thirty-two organisations were selected for the initial funding of EUR 8.3 million. The Ramp-Up Phase ended on 31 March 2016. Funding is reassessed every two years using Specific Grant Agreements (SGA); the first of which began in April 2016 (ending in April 2018), and the second with a total EU funding of 88 Million Euro starting in April 2018 (ending in March 2020). The HBP's total costs are estimated at EUR 1.019 billion, of which EUR 500 million will be provided by the European Commission, EUR 500 million by national, public and private organisations, and EUR 19 million by the Core Project Ramp-Up Phase Partners.

Obstacles

One of the Project's primary hurdles is the unsystematic nature of the information collected from previous brain research. Neurological research data varies by biological organisation schemes, species studied, and by developmental stages, making it difficult to collectively use the data to replicate the brain in a model that acts as a single system.

Other obstacles include engineering problems involving power consumption, memory, and storage. For example, detailed neuron representations are very computationally expensive, and whole brain simulation is at the leading edge of our computational capability.

Implications

The Human Brain Project moved to Campus Biotech in 2014.

Technologies generated by the HBP and other similar projects offer several possibilities to other fields of research. For instance, a brain model can be used to investigate signatures of disease in the brain and the impact of certain drugs, enabling the development of better diagnosis and treatment methods. Ultimately, these technologies will likely lead to more advanced medical options available to patients at a lower cost.

In addition, detailed brain simulation requires significant computing power, leading to developments in supercomputing and energy-efficient, brain-inspired computing techniques. Computational developments can be extended into areas such as data mining, telecommunications, appliances, and other industrial uses.

The long-term ethical consequences of the Project are also considered. The Project follows a policy of Responsible Research and Innovation, and its Ethics Advisory Board is responsible for monitoring the use of human volunteers, animal subjects, and the data collected. Implications on European society, industry, and economy are investigated by the HBP Ethics and Society Programme's Foresight Lab.

Criticism

An open letter was sent on 7 July 2015 to the European Commission by 154 European researchers (750 signatures as of 3 September 2014) complaining of the HBP's overly narrow approach, and threatening to boycott the project. Central to this controversy was an internal dispute about funding for cognitive scientists who study high level brain functions, such as thought and behaviour. However, the HBP stated that there is "no question that cognition and behaviour are vital to the HBP", explaining that cognitive neuroscience research was repositioned in the project to allow the core project to focus on building the platforms. In addition, the open letter called on the EC to "reallocate the funding currently allocated to the HBP core and partnering projects to broad neuroscience-directed funding to meet the original goals of the HBP—understanding brain function and its effect on society". In its response, the HBP said that "while neuroscience research generates a vast amount of valuable data, there is currently no technology for sharing, organising, analysing or integrating this information, beyond papers and even databases. The HBP will provide the critical missing layer to move towards a multi-level reconstruction and simulation of the brain". It added that "cognitive and behavioural neuroscience will become the most significant component of neuroscience in HBP over the course of the project. However, for this to happen the platforms have to be in place first".

Peter Dayan, director of computational neuroscience at University College London, argued that the goal of a large-scale simulation of the brain is radically premature, and Geoffrey Hinton said that "the real problem with that project is they have no clue how to get a large system like that to learn". Similar concerns as to the project's methodology were raised by Robert Epstein.

The HBP has said that its members share the uncertainty surrounding large-scale simulation, but that "reconstructing and simulating the human brain is a vision, a target; the benefits will come from the technology needed to get there. That technology, developed by the HBP, will benefit all of neuroscience as well as related fields".

In 2015 the project underwent a review process and the three-member executive committee, led by Henry Markram, was dissolved and replaced by a 22-member governing board.

According to a 2019 article, "In 2013, the European Commission awarded his initiative—the Human Brain Project (HBP)—a staggering 1 billion euro grant (worth about $1.42 billion at the time)...the people I contacted struggled to name a major contribution that the HBP has made in the past decade." Another article concluded that "Ultimately, the mega-project did create communities of scientists focused on some common goals, he says. "That's an enduring legacy.""

Legacy

The European Brain Research Infrastructures (EBRAINS) is a research infrastructure created through HBP "that gathers an extensive range of data and tools for brain related research". EBRAINS consists of a set of infrastructure initiatives (such as brain atlases), tools and services (such as the sPyNNaker software suite for SpiNNaker hardware), and community projects. It is an international non-profit association headquartered in Brussels, Belgium, and a member of the European Open Science Cloud association.

Neural network (machine learning)

From Wikipedia, the free encyclopedia

https://en.wikipedia.org/wiki/Neural_network_(machine_learning)

An artificial neural network is an interconnected group of nodes, inspired by a simplification of neurons in a brain. Here, each circular node represents an artificial neuron and an arrow represents a connection from the output of one artificial neuron to the input of another.

In machine learning, a neural network (NN) or neural net, also known as an artificial neural network (ANN), is a computational model inspired by the structure and functions of biological neural networks.

A neural network consists of connected units or nodes called artificial neurons, which loosely model the neurons in the brain. Artificial neuron models that mimic biological neurons more closely have also been recently investigated and shown to significantly improve performance. These are connected by edges, which model the synapses in the brain. Each artificial neuron receives signals from connected neurons, then processes them and sends a signal to other connected neurons. The "signal" is a real number, and the output of each neuron is computed by some non-linear function of the totality of its inputs, called the activation function. The strength of the signal at each connection is determined by a weight, which adjusts during the learning process.

Typically, neurons are aggregated into layers. Different layers may perform different transformations on their inputs. Signals travel from the first layer (the input layer) to the last layer (the output layer), possibly passing through multiple intermediate layers (hidden layers). A network is typically called a deep neural network if it has at least two hidden layers.

Artificial neural networks are used for various tasks, including predictive modeling, adaptive control, and solving problems in artificial intelligence. They can learn from experience, and can derive conclusions from a complex and seemingly unrelated set of information.

Training

Neural networks are typically trained through empirical risk minimization, which is based on the idea of optimizing the network's parameters to minimize the difference, or empirical risk, between the predicted output and the actual target values in a given dataset. Gradient-based methods such as backpropagation are usually used to estimate the parameters of the network. During the training phase, ANNs learn from labeled training data by iteratively updating their parameters to minimize a defined loss function. This method allows the network to generalize to unseen data.

Simplified example of training a neural network in object detection: The network is trained by multiple images that are known to depict starfish and sea urchins, which are correlated with "nodes" that represent visual features. The starfish match with a ringed texture and a star outline, whereas most sea urchins match with a striped texture and oval shape. However, the instance of a ring textured sea urchin creates a weakly weighted association between them.

 

Subsequent run of the network on an input image: The network correctly detects the starfish. However, the weakly weighted association between ringed texture and sea urchin also confers a weak signal to the latter from one of two intermediate nodes. In addition, a shell that was not included in the training gives a weak signal for the oval shape, also resulting in a weak signal for the sea urchin output. These weak signals may result in a false positive result for sea urchin.
In reality, textures and outlines would not be represented by single nodes, but rather by associated weight patterns of multiple nodes.

History

Main article: History of artificial neural networks

Early work

Today's deep neural networks are based on early work in statistics over 200 years ago. The simplest kind of feedforward neural network (FNN) is a linear network, which consists of a single layer of output nodes with linear activation functions; the inputs are fed directly to the outputs via a series of weights. The sum of the products of the weights and the inputs is calculated at each node. The mean squared errors between these calculated outputs and the given target values are minimized by creating an adjustment to the weights. This technique has been known for over two centuries as the method of least squares or linear regression. It was used as a means of finding a good rough linear fit to a set of points by Legendre (1805) and Gauss (1795) for the prediction of planetary movement.

Historically, digital computers such as the von Neumann model operate via the execution of explicit instructions with access to memory by a number of processors. Some neural networks, on the other hand, originated from efforts to model information processing in biological systems through the framework of connectionism. Unlike the von Neumann model, connectionist computing does not separate memory and processing.

Warren McCulloch and Walter Pitts (1943) considered a non-learning computational model for neural networks. This model paved the way for research to split into two approaches. One approach focused on biological processes while the other focused on the application of neural networks to artificial intelligence.

In the late 1940s, D. O. Hebb proposed a learning hypothesis based on the mechanism of neural plasticity that became known as Hebbian learning. It was used in many early neural networks, such as Rosenblatt's perceptron and the Hopfield network. Farley and Clark (1954) used computational machines to simulate a Hebbian network. Other neural network computational machines were created by Rochester, Holland, Habit and Duda (1956).

In 1958, psychologist Frank Rosenblatt described the perceptron, one of the first implemented artificial neural networks, funded by the United States Office of Naval Research. R. D. Joseph (1960) mentions an even earlier perceptron-like device by Farley and Clark: "Farley and Clark of MIT Lincoln Laboratory actually preceded Rosenblatt in the development of a perceptron-like device." However, "they dropped the subject." The perceptron raised public excitement for research in Artificial Neural Networks, causing the US government to drastically increase funding. This contributed to "the Golden Age of AI" fueled by the optimistic claims made by computer scientists regarding the ability of perceptrons to emulate human intelligence.

The first perceptrons did not have adaptive hidden units. However, Joseph (1960) also discussed multilayer perceptrons with an adaptive hidden layer. Rosenblatt (1962) cited and adopted these ideas, also crediting work by H. D. Block and B. W. Knight. Unfortunately, these early efforts did not lead to a working learning algorithm for hidden units, i.e., deep learning.

Deep learning breakthroughs in the 1960s and 1970s

Fundamental research was conducted on ANNs in the 1960s and 1970s. The first working deep learning algorithm was the Group method of data handling, a method to train arbitrarily deep neural networks, published by Alexey Ivakhnenko and Lapa in the Soviet Union (1965). They regarded it as a form of polynomial regression, or a generalization of Rosenblatt's perceptron. A 1971 paper described a deep network with eight layers trained by this method, which is based on layer by layer training through regression analysis. Superfluous hidden units are pruned using a separate validation set. Since the activation functions of the nodes are Kolmogorov-Gabor polynomials, these were also the first deep networks with multiplicative units or "gates."

The first deep learning multilayer perceptron (MLP) trained by stochastic gradient descent was published in 1967 by Shun'ichi Amari. In computer experiments conducted by Amari's student Saito, a five layer MLP with two modifiable layers learned internal representations to classify non-linearily separable pattern classes. Subsequent developments in hardware and hyperparameter tunings have made end-to-end stochastic gradient descent the currently dominant training technique.

In 1969, Kunihiko Fukushima introduced the ReLU (rectified linear unit) activation function. The rectifier has become the most popular activation function for deep learning.

Nevertheless, research stagnated in the United States following the work of Minsky and Papert (1969), who emphasized that basic perceptrons were incapable of processing the exclusive-or circuit. This insight was irrelevant for the deep networks of Ivakhnenko (1965) and Amari (1967).

In 1976 transfer learning was introduced in neural networks learning.

Deep learning architectures for convolutional neural networks (CNNs) with convolutional layers and downsampling layers and weight replication began with the neocognitron introduced by Kunihiko Fukushima in 1979, though not trained by backpropagation.

Backpropagation

Backpropagation is an efficient application of the chain rule derived by Gottfried Wilhelm Leibniz in 1673 to networks of differentiable nodes. The terminology "back-propagating errors" was actually introduced in 1962 by Rosenblatt, but he did not know how to implement this, although Henry J. Kelley had a continuous precursor of backpropagation in 1960 in the context of control theory. In 1970, Seppo Linnainmaa published the modern form of backpropagation in his Master's thesis (1970). G.M. Ostrovski et al. republished it in 1971. Paul Werbos applied backpropagation to neural networks in 1982 (his 1974 PhD thesis, reprinted in a 1994 book, did not yet describe the algorithm). In 1986, David E. Rumelhart et al. popularised backpropagation but did not cite the original work.

Convolutional neural networks

Kunihiko Fukushima's convolutional neural network (CNN) architecture of 1979 also introduced max pooling, a popular downsampling procedure for CNNs. CNNs have become an essential tool for computer vision.

The time delay neural network (TDNN) was introduced in 1987 by Alex Waibel to apply CNN to phoneme recognition. It used convolutions, weight sharing, and backpropagation. In 1988, Wei Zhang applied a backpropagation-trained CNN to alphabet recognition. In 1989, Yann LeCun et al. created a CNN called LeNet for recognizing handwritten ZIP codes on mail. Training required 3 days. In 1990, Wei Zhang implemented a CNN on optical computing hardware. In 1991, a CNN was applied to medical image object segmentation and breast cancer detection in mammograms. LeNet-5 (1998), a 7-level CNN by Yann LeCun et al., that classifies digits, was applied by several banks to recognize hand-written numbers on checks digitized in 32×32 pixel images.

From 1988 onward, the use of neural networks transformed the field of protein structure prediction, in particular when the first cascading networks were trained on profiles (matrices) produced by multiple sequence alignments.

Recurrent neural networks

One origin of RNN was statistical mechanics. In 1972, Shun'ichi Amari proposed to modify the weights of an Ising model by Hebbian learning rule as a model of associative memory, adding in the component of learning. This was popularized as the Hopfield network by John Hopfield (1982). Another origin of RNN was neuroscience. The word "recurrent" is used to describe loop-like structures in anatomy. In 1901, Cajal observed "recurrent semicircles" in the cerebellar cortex. Hebb considered "reverberating circuit" as an explanation for short-term memory. The McCulloch and Pitts paper (1943) considered neural networks that contain cycles, and noted that the current activity of such networks can be affected by activity indefinitely far in the past.

In 1982 a recurrent neural network with an array architecture (rather than a multilayer perceptron architecture), namely a Crossbar Adaptive Array, used direct recurrent connections from the output to the supervisor (teaching) inputs. In addition of computing actions (decisions), it computed internal state evaluations (emotions) of the consequence situations. Eliminating the external supervisor, it introduced the self-learning method in neural networks.

In cognitive psychology, the journal American Psychologist in early 1980's carried out a debate on the relation between cognition and emotion. Zajonc in 1980 stated that emotion is computed first and is independent from cognition, while Lazarus in 1982 stated that cognition is computed first and is inseparable from emotion. In 1982 the Crossbar Adaptive Array gave a neural network model of cognition-emotion relation. It was an example of a debate where an AI system, a recurrent neural network, contributed to an issue in the same time addressed by cognitive psychology.

Two early influential works were the Jordan network (1986) and the Elman network (1990), which applied RNN to study cognitive psychology.

In the 1980s, backpropagation did not work well for deep RNNs. To overcome this problem, in 1991, Jürgen Schmidhuber proposed the "neural sequence chunker" or "neural history compressor" which introduced the important concepts of self-supervised pre-training (the "P" in ChatGPT) and neural knowledge distillation. In 1993, a neural history compressor system solved a "Very Deep Learning" task that required more than 1000 subsequent layers in an RNN unfolded in time.

In 1991, Sepp Hochreiter's diploma thesis identified and analyzed the vanishing gradient problem and proposed recurrent residual connections to solve it. He and Schmidhuber introduced long short-term memory (LSTM), which set accuracy records in multiple applications domains. This was not yet the modern version of LSTM, which required the forget gate, which was introduced in 1999. It became the default choice for RNN architecture.

During 1985–1995, inspired by statistical mechanics, several architectures and methods were developed by Terry Sejnowski, Peter Dayan, Geoffrey Hinton, etc., including the Boltzmann machine, restricted Boltzmann machine, Helmholtz machine, and the wake-sleep algorithm. These were designed for unsupervised learning of deep generative models.

Deep learning

Between 2009 and 2012, ANNs began winning prizes in image recognition contests, approaching human level performance on various tasks, initially in pattern recognition and handwriting recognition. In 2011, a CNN named DanNet by Dan Ciresan, Ueli Meier, Jonathan Masci, Luca Maria Gambardella, and Jürgen Schmidhuber achieved for the first time superhuman performance in a visual pattern recognition contest, outperforming traditional methods by a factor of 3. It then won more contests. They also showed how max-pooling CNNs on GPU improved performance significantly.

In October 2012, AlexNet by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton won the large-scale ImageNet competition by a significant margin over shallow machine learning methods. Further incremental improvements included the VGG-16 network by Karen Simonyan and Andrew Zisserman and Google's Inceptionv3.

In 2012, Ng and Dean created a network that learned to recognize higher-level concepts, such as cats, only from watching unlabeled images. Unsupervised pre-training and increased computing power from GPUs and distributed computing allowed the use of larger networks, particularly in image and visual recognition problems, which became known as "deep learning".

Radial basis function and wavelet networks were introduced in 2013. These can be shown to offer best approximation properties and have been applied in nonlinear system identification and classification applications.

Generative adversarial network (GAN) (Ian Goodfellow et al., 2014) became state of the art in generative modeling during 2014–2018 period. The GAN principle was originally published in 1991 by Jürgen Schmidhuber who called it "artificial curiosity": two neural networks contest with each other in the form of a zero-sum game, where one network's gain is the other network's loss. The first network is a generative model that models a probability distribution over output patterns. The second network learns by gradient descent to predict the reactions of the environment to these patterns. Excellent image quality is achieved by Nvidia's StyleGAN (2018) based on the Progressive GAN by Tero Karras et al. Here, the GAN generator is grown from small to large scale in a pyramidal fashion. Image generation by GAN reached popular success, and provoked discussions concerning deepfakes. Diffusion models (2015) eclipsed GANs in generative modeling since then, with systems such as DALL·E 2 (2022) and Stable Diffusion (2022).

In 2014, the state of the art was training "very deep neural network" with 20 to 30 layers. Stacking too many layers led to a steep reduction in training accuracy, known as the "degradation" problem. In 2015, two techniques were developed to train very deep networks: the highway network was published in May 2015, and the residual neural network (ResNet) in December 2015. ResNet behaves like an open-gated Highway Net.

Main article: Transformer (deep learning architecture) § History

During the 2010s, the seq2seq model was developed, and attention mechanisms were added. It led to the modern Transformer architecture in 2017 in Attention Is All You Need. It requires computation time that is quadratic in the size of the context window. Jürgen Schmidhuber's fast weight controller (1992) scales linearly and was later shown to be equivalent to the unnormalized linear Transformer. Transformers have increasingly become the model of choice for natural language processing. Many modern large language models such as ChatGPT, GPT-4, and BERT use this architecture.

Models

Further information: Mathematics of artificial neural networks

Neuron and myelinated axon, with signal flow from inputs at dendrites to outputs at axon terminals

ANNs began as an attempt to exploit the architecture of the human brain to perform tasks that conventional algorithms had little success with. They soon reoriented towards improving empirical results, abandoning attempts to remain true to their biological precursors. ANNs have the ability to learn and model non-linearities and complex relationships. This is achieved by neurons being connected in various patterns, allowing the output of some neurons to become the input of others. The network forms a directed, weighted graph.

An artificial neural network consists of simulated neurons. Each neuron is connected to other nodes via links like a biological axon-synapse-dendrite connection. All the nodes connected by links take in some data and use it to perform specific operations and tasks on the data. Each link has a weight, determining the strength of one node's influence on another, allowing weights to choose the signal between neurons.

Artificial neurons

Main article: Artificial neuron

ANNs are composed of artificial neurons which are conceptually derived from biological neurons. Each artificial neuron has inputs and produces a single output which can be sent to multiple other neurons.^[114] The inputs can be the feature values of a sample of external data, such as images or documents, or they can be the outputs of other neurons. The outputs of the final output neurons of the neural net accomplish the task, such as recognizing an object in an image.

To find the output of the neuron we take the weighted sum of all the inputs, weighted by the weights of the connections from the inputs to the neuron. We add a bias term to this sum. This weighted sum is sometimes called the activation. This weighted sum is then passed through a (usually nonlinear) activation function to produce the output. The initial inputs are external data, such as images and documents. The ultimate outputs accomplish the task, such as recognizing an object in an image.

Organization

The neurons are typically organized into multiple layers, especially in deep learning. Neurons of one layer connect only to neurons of the immediately preceding and immediately following layers. The layer that receives external data is the input layer. The layer that produces the ultimate result is the output layer. In between them are zero or more hidden layers. Single layer and unlayered networks are also used. Between two layers, multiple connection patterns are possible. They can be 'fully connected', with every neuron in one layer connecting to every neuron in the next layer. They can be pooling, where a group of neurons in one layer connects to a single neuron in the next layer, thereby reducing the number of neurons in that layer. Neurons with only such connections form a directed acyclic graph and are known as feedforward networks. Alternatively, networks that allow connections between neurons in the same or previous layers are known as recurrent networks.

Hyperparameter

Main article: Hyperparameter (machine learning)

A hyperparameter is a constant parameter defining any configurable part of the learning process, whose value is set prior to training. Examples of hyperparameters include learning rate, batch size and regularization parameters. The performance of a neural network is strongly influenced by the choice of hyperparameter values, and thus the hyperparameters are often optimized as part of the training process, a process called hyperparameter tuning or hyperparameter optimization.

Learning

See also: Mathematical optimization, Estimation theory, and Machine learning

Learning is the adaptation of the network to better handle a task by considering sample observations. Learning involves adjusting the weights (and optional thresholds) of the network to improve the accuracy of the result. This is done by minimizing the observed errors. Learning is complete when examining additional observations does not usefully reduce the error rate. Even after learning, the error rate typically does not reach 0. If after learning, the error rate is too high, the network typically must be redesigned. Practically this is done by defining a cost function that is evaluated periodically during learning. As long as its output continues to decline, learning continues. The cost is frequently defined as a statistic whose value can only be approximated. The outputs are actually numbers, so when the error is low, the difference between the output (almost certainly a cat) and the correct answer (cat) is small. Learning attempts to reduce the total of the differences across the observations. Most learning models can be viewed as a straightforward application of optimization theory and statistical estimation.

Learning rate

Main article: Learning rate

The learning rate defines the size of the corrective steps that the model takes to adjust for errors in each observation. A high learning rate shortens the training time, but with lower ultimate accuracy, while a lower learning rate takes longer, but with the potential for greater accuracy. Optimizations such as Quickprop are primarily aimed at speeding up error minimization, while other improvements mainly try to increase reliability. In order to avoid oscillation inside the network such as alternating connection weights, and to improve the rate of convergence, refinements use an adaptive learning rate that increases or decreases as appropriate. The concept of momentum allows the balance between the gradient and the previous change to be weighted such that the weight adjustment depends to some degree on the previous change. A momentum close to 0 emphasizes the gradient, while a value close to 1 emphasizes the last change.

Cost function

While it is possible to define a cost function ad hoc, frequently the choice is determined by the function's desirable properties (such as convexity) because it arises from the model (e.g. in a probabilistic model, the model's posterior probability can be used as an inverse cost).

Backpropagation

Main article: Backpropagation

Backpropagation is a method used to adjust the connection weights to compensate for each error found during learning. The error amount is effectively divided among the connections. Technically, backpropagation calculates the gradient (the derivative) of the cost function associated with a given state with respect to the weights. The weight updates can be done via stochastic gradient descent or other methods, such as extreme learning machines, "no-prop" networks, training without backtracking, "weightless" networks, and non-connectionist neural networks.

Learning paradigms

Machine learning is commonly separated into three main learning paradigms, supervised learning, unsupervised learning and reinforcement learning. Each corresponds to a particular learning task.

Supervised learning

Supervised learning uses a set of paired inputs and desired outputs. The learning task is to produce the desired output for each input. In this case, the cost function is related to eliminating incorrect deductions. A commonly used cost is the mean-squared error, which tries to minimize the average squared error between the network's output and the desired output. Tasks suited for supervised learning are pattern recognition (also known as classification) and regression (also known as function approximation). Supervised learning is also applicable to sequential data (e.g., for handwriting, speech and gesture recognition). This can be thought of as learning with a "teacher", in the form of a function that provides continuous feedback on the quality of solutions obtained thus far.

Unsupervised learning

In unsupervised learning, input data is given along with the cost function, some function of the data ${\displaystyle x}$ and the network's output. The cost function is dependent on the task (the model domain) and any a priori assumptions (the implicit properties of the model, its parameters and the observed variables). As a trivial example, consider the model ${\displaystyle f(x)=a}$ where ${\displaystyle a}$ is a constant and the cost ${\displaystyle C=E[(x-f(x){)}^2]}$. Minimizing this cost produces a value of ${\displaystyle a}$ that is equal to the mean of the data. The cost function can be much more complicated. Its form depends on the application: for example, in compression it could be related to the mutual information between ${\displaystyle x}$ and ${\displaystyle f(x)}$, whereas in statistical modeling, it could be related to the posterior probability of the model given the data (note that in both of those examples, those quantities would be maximized rather than minimized). Tasks that fall within the paradigm of unsupervised learning are in general estimation problems; the applications include clustering, the estimation of statistical distributions, compression and filtering.

Reinforcement learning

Main article: Reinforcement learning

See also: Stochastic control

In applications such as playing video games, an actor takes a string of actions, receiving a generally unpredictable response from the environment after each one. The goal is to win the game, i.e., generate the most positive (lowest cost) responses. In reinforcement learning, the aim is to weight the network (devise a policy) to perform actions that minimize long-term (expected cumulative) cost. At each point in time the agent performs an action and the environment generates an observation and an instantaneous cost, according to some (usually unknown) rules. The rules and the long-term cost usually only can be estimated. At any juncture, the agent decides whether to explore new actions to uncover their costs or to exploit prior learning to proceed more quickly.

Formally, the environment is modeled as a Markov decision process (MDP) with states ${\displaystyle s_1,...,s_n\in S}$ and actions ${\displaystyle a_1,...,a_m\in A}$. Because the state transitions are not known, probability distributions are used instead: the instantaneous cost distribution ${\displaystyle P(c_t|s_t)}$, the observation distribution ${\displaystyle P(x_t|s_t)}$ and the transition distribution ${\displaystyle P(s_{t+1}|s_t,a_t)}$, while a policy is defined as the conditional distribution over actions given the observations. Taken together, the two define a Markov chain (MC). The aim is to discover the lowest-cost MC.

ANNs serve as the learning component in such applications. Dynamic programming coupled with ANNs (giving neurodynamic programming) has been applied to problems such as those involved in vehicle routing, video games, natural resource management and medicine because of ANNs ability to mitigate losses of accuracy even when reducing the discretization grid density for numerically approximating the solution of control problems. Tasks that fall within the paradigm of reinforcement learning are control problems, games and other sequential decision making tasks.

Self-learning

Self-learning in neural networks was introduced in 1982 along with a neural network capable of self-learning named crossbar adaptive array (CAA). It is a system with only one input, situation s, and only one output, action (or behavior) a. It has neither external advice input nor external reinforcement input from the environment. The CAA computes, in a crossbar fashion, both decisions about actions and emotions (feelings) about encountered situations. The system is driven by the interaction between cognition and emotion. Given the memory matrix, W =||w(a,s)||, the crossbar self-learning algorithm in each iteration performs the following computation:

 In situation s perform action a;
 Receive consequence situation s';
 Compute emotion of being in consequence situation v(s');
 Update crossbar memory w'(a,s) = w(a,s) + v(s').

The backpropagated value (secondary reinforcement) is the emotion toward the consequence situation. The CAA exists in two environments, one is behavioral environment where it behaves, and the other is genetic environment, where from it receives initial emotions (only once) about to be encountered situations in the behavioral environment. Having received the genome vector (species vector) from the genetic environment, the CAA will learn a goal-seeking behavior, in the behavioral environment that contains both desirable and undesirable situations.

Neuroevolution

Main article: Neuroevolution

Neuroevolution can create neural network topologies and weights using evolutionary computation. It is competitive with sophisticated gradient descent approaches. One advantage of neuroevolution is that it may be less prone to get caught in "dead ends".

Stochastic neural network

Stochastic neural networks originating from Sherrington–Kirkpatrick models are a type of artificial neural network built by introducing random variations into the network, either by giving the network's artificial neurons stochastic transfer functions, or by giving them stochastic weights. This makes them useful tools for optimization problems, since the random fluctuations help the network escape from local minima. Stochastic neural networks trained using a Bayesian approach are known as Bayesian neural networks.

Topological deep learning

Topological deep learning, first introduced in 2017, is an emerging approach in machine learning that integrates topology with deep neural networks to address highly intricate and high-order data. Initially rooted in algebraic topology, TDL has since evolved into a versatile framework incorporating tools from other mathematical disciplines, such as differential topology and geometric topology. As a successful example of mathematical deep learning, TDL continues to inspire advancements in mathematical artificial intelligence, fostering a mutually beneficial relationship between AI and mathematics.

Other

In a Bayesian framework, a distribution over the set of allowed models is chosen to minimize the cost. Evolutionary methods, gene expression programming, simulated annealing, expectation–maximization, non-parametric methods and particle swarm optimization are other learning algorithms. Convergent recursion is a learning algorithm for cerebellar model articulation controller (CMAC) neural networks.

Modes

Two modes of learning are available: stochastic and batch. In stochastic learning, each input creates a weight adjustment. In batch learning, weights are adjusted based on a batch of inputs, accumulating errors over the batch. Stochastic learning introduces "noise" into the process, using the local gradient calculated from one data point; this reduces the chance of the network getting stuck in local minima. However, batch learning typically yields a faster, more stable descent to a local minimum, since each update is performed in the direction of the batch's average error. A common compromise is to use "mini-batches", small batches with samples in each batch selected stochastically from the entire data set.

Types

Main article: Types of artificial neural networks

ANNs have evolved into a broad family of techniques that have advanced the state of the art across multiple domains. The simplest types have one or more static components, including number of units, number of layers, unit weights and topology. Dynamic types allow one or more of these to evolve via learning. The latter is much more complicated but can shorten learning periods and produce better results. Some types allow/require learning to be "supervised" by the operator, while others operate independently. Some types operate purely in hardware, while others are purely software and run on general purpose computers.

Some of the main breakthroughs include:

• Convolutional neural networks (CNN) that have proven particularly successful in processing visual and other two-dimensional data; where long short-term memory avoids the vanishing gradient problem and can handle signals that have a mix of low and high frequency components aiding large-vocabulary speech recognition, text-to-speech synthesis, and photo-real talking heads. Convolutional Neural Networks have also been applied to fraud detection.
• Competitive networks such as generative adversarial networks in which multiple networks (of varying structure) compete with each other, on tasks such as winning a game or on deceiving the opponent about the authenticity of an input.

Network design

Using artificial neural networks requires an understanding of their characteristics.

• Choice of model: This depends on the data representation and the application. Model parameters include the number, type, and connectedness of network layers, as well as the size of each and the connection type (full, pooling, etc.). Overly complex models learn slowly.
• Learning algorithm: Numerous trade-offs exist between learning algorithms. Almost any algorithm will work well with the correct hyperparameters for training on a particular data set. However, selecting and tuning an algorithm for training on unseen data requires significant experimentation.
• Robustness: If the model, cost function and learning algorithm are selected appropriately, the resulting ANN can become robust.

Neural architecture search (NAS) uses machine learning to automate ANN design. Various approaches to NAS have designed networks that compare well with hand-designed systems. The basic search algorithm is to propose a candidate model, evaluate it against a dataset, and use the results as feedback to teach the NAS network. Available systems include AutoML and AutoKeras. scikit-learn library provides functions to help with building a deep network from scratch. We can then implement a deep network with TensorFlow or Keras.

Hyperparameters must also be defined as part of the design (they are not learned), governing matters such as how many neurons are in each layer, learning rate, step, stride, depth, receptive field and padding (for CNNs), etc. The Python code snippet provides an overview of the training function, which uses the training dataset, number of hidden layer units, learning rate, and number of iterations as parameters:

def train(X, y, n_hidden, learning_rate, n_iter):
    """Training function.

    Args:
      X: Argument X.
      y: Argument y.
      n_hidden: The number of hidden layer units.
      learning_rate: The learning rate.
      n_iter: The number of iterations.

    Returns:
      dict: A dictionary.
    """
    m, n_input = X.shape

    # 1. random initialize weights and biases
    w1 = np.random.randn(n_input, n_hidden)
    b1 = np.zeros((1, n_hidden))
    w2 = np.random.randn(n_hidden, 1)
    b2 = np.zeros((1, 1))

    # 2. in each iteration, feed all layers with the latest weights and biases
    for i in range(n_iter + 1):
        z2 = np.dot(X, w1) + b1
        a2 = sigmoid(z2)
        z3 = np.dot(a2, w2) + b2
        a3 = z3
        dz3 = a3 - y
        dw2 = np.dot(a2.T, dz3)
        db2 = np.sum(dz3, axis=0, keepdims=True)
        dz2 = np.dot(dz3, w2.T) * sigmoid_derivative(z2)
        dw1 = np.dot(X.Y, dz2)
        db1 = np.sum(dz2, axis=0)

        # 3. update weights and biases with gradients
        w1 -= learning_rate * dw1 / m
        w2 -= learning_rate * dw2 / m
        b1 -= learning_rate * db1 / m
        b2 -= learning_rate * db2 / m

        if i % 1000 == 0:
            print("Epoch", i, "loss: ", np.mean(np.square(dz3)))

    model = {"w1": w1, "b1": b1, "w2": w2, "b2": b2}
    return model

Monitoring and concept drift detection of ANNs

When neural networks are deployed in real-world applications, the statistical properties of the input data may change over time, a phenomenon known as concept drift or non-stationarity. Drift can reduce predictive accuracy and lead to unreliable or biased decisions if it is not detected and corrected. In practice, this means that the model's accuracy in deployment may differ substantially from the levels observed during training or cross-validation.

Several strategies have been developed to monitor neural networks for drift and degradation:

• Error-based monitoring: comparing current predictions against ground-truth labels when they become available. This approach directly quantifies predictive performance but may be impractical when labels are delayed or costly to obtain.
• Data distribution monitoring: detecting changes in the input data distribution using statistical tests, divergence measures, or density-ratio estimation.
• Representation monitoring: tracking the distribution of internal embeddings or hidden-layer features. Shifts in the latent representation can indicate nonstationarity even when labels are unavailable. Statistical methods such as statistical process control charts have been adapted for this purpose.

Applications

Because of their ability to model and reproduce nonlinear processes, artificial neural networks have found applications in many disciplines. These include:

• Function approximation, or regression analysis, (including time series prediction, fitness approximation, and modeling)
• Data processing (including filtering, clustering, blind source separation, and compression)
• Nonlinear system identification and control (including vehicle control, trajectory prediction, adaptive control, process control, and natural resource management)
• Pattern recognition (including radar systems, face identification, signal classification, novelty detection, 3D reconstruction, object recognition, and sequential decision making)
• Sequence recognition (including gesture, speech, and handwritten and printed text recognition)
• Sensor data analysis (including image analysis)
• Robotics (including directing manipulators and prostheses)
• Data mining (including knowledge discovery in databases)
• Finance (such as ex-ante models for specific financial long-run forecasts and artificial financial markets)
• Quantum chemistry
• General game playing
• Generative AI
• Data visualization
• Machine translation
• Social network filtering
• E-mail spam filtering
• Medical diagnosis

ANNs have been used to diagnose several types of cancers and to distinguish highly invasive cancer cell lines from less invasive lines using only cell shape information.

ANNs have been used to accelerate reliability analysis of infrastructures subject to natural disasters and to predict foundation settlements. It can also be useful to mitigate flood by the use of ANNs for modelling rainfall-runoff. ANNs have also been used for building black-box models in geoscience: hydrology, ocean modelling and coastal engineering, and geomorphology. ANNs have been employed in cybersecurity, with the objective to discriminate between legitimate activities and malicious ones. For example, machine learning has been used for classifying Android malware, for identifying domains belonging to threat actors and for detecting URLs posing a security risk. Research is underway on ANN systems designed for penetration testing, for detecting botnets, credit cards frauds and network intrusions.

ANNs have been proposed as a tool to solve partial differential equations in physics and simulate the properties of many-body open quantum systems. In brain research ANNs have studied short-term behavior of individual neurons, the dynamics of neural circuitry arise from interactions between individual neurons and how behavior can arise from abstract neural modules that represent complete subsystems. Studies considered long-and short-term plasticity of neural systems and their relation to learning and memory from the individual neuron to the system level.

It is possible to create a profile of a user's interests from pictures, using artificial neural networks trained for object recognition.

Beyond their traditional applications, artificial neural networks are increasingly being utilized in interdisciplinary research, such as materials science. For instance, graph neural networks (GNNs) have demonstrated their capability in scaling deep learning for the discovery of new stable materials by efficiently predicting the total energy of crystals. This application underscores the adaptability and potential of ANNs in tackling complex problems beyond the realms of predictive modeling and artificial intelligence, opening new pathways for scientific discovery and innovation.

Theoretical properties

Computational power

The multilayer perceptron is a universal function approximator, as proven by the universal approximation theorem. However, the proof is not constructive regarding the number of neurons required, the network topology, the weights and the learning parameters.

A specific recurrent architecture with rational-valued weights (as opposed to full precision real number-valued weights) has the power of a universal Turing machine, using a finite number of neurons and standard linear connections. Further, the use of irrational values for weights results in a machine with super-Turing power.

Capacity

A model's "capacity" property corresponds to its ability to model any given function. It is related to the amount of information that can be stored in the network and to the notion of complexity. Two notions of capacity are known by the community. The information capacity and the VC Dimension. The information capacity of a perceptron is intensively discussed in Sir David MacKay's book which summarizes work by Thomas Cover. The capacity of a network of standard neurons (not convolutional) can be derived by four rules that derive from understanding a neuron as an electrical element. The information capacity captures the functions modelable by the network given any data as input. The second notion, is the VC dimension. VC Dimension uses the principles of measure theory and finds the maximum capacity under the best possible circumstances. This is, given input data in a specific form. As noted in, the VC Dimension for arbitrary inputs is half the information capacity of a perceptron. The VC Dimension for arbitrary points is sometimes referred to as Memory Capacity.

Convergence

Models may not consistently converge on a single solution, firstly because local minima may exist, depending on the cost function and the model. Secondly, the optimization method used might not guarantee to converge when it begins far from any local minimum. Thirdly, for sufficiently large data or parameters, some methods become impractical.

Another issue worthy to mention is that training may cross some saddle point which may lead the convergence to the wrong direction.

The convergence behavior of certain types of ANN architectures are more understood than others. When the width of network approaches to infinity, the ANN is well described by its first order Taylor expansion throughout training, and so inherits the convergence behavior of affine models. Another example is when parameters are small, it is observed that ANNs often fit target functions from low to high frequencies. This behavior is referred to as the spectral bias, or frequency principle, of neural networks. This phenomenon is the opposite to the behavior of some well studied iterative numerical schemes such as Jacobi method. Deeper neural networks have been observed to be more biased towards low frequency functions.

Generalization and statistics

Applications whose goal is to create a system that generalizes well to unseen examples, face the possibility of over-training. This arises in convoluted or over-specified systems when the network capacity significantly exceeds the needed free parameters.

Two approaches address over-training. The first is to use cross-validation and similar techniques to check for the presence of over-training and to select hyperparameters to minimize the generalization error. The second is to use some form of regularization. This concept emerges in a probabilistic (Bayesian) framework, where regularization can be performed by selecting a larger prior probability over simpler models; but also in statistical learning theory, where the goal is to minimize over two quantities: the 'empirical risk' and the 'structural risk', which roughly corresponds to the error over the training set and the predicted error in unseen data due to overfitting.

Confidence analysis of a neural network

Supervised neural networks that use a mean squared error (MSE) cost function can use formal statistical methods to determine the confidence of the trained model. The MSE on a validation set can be used as an estimate for variance. This value can then be used to calculate the confidence interval of network output, assuming a normal distribution. A confidence analysis made this way is statistically valid as long as the output probability distribution stays the same and the network is not modified.

By assigning a softmax activation function, a generalization of the logistic function, on the output layer of the neural network (or a softmax component in a component-based network) for categorical target variables, the outputs can be interpreted as posterior probabilities. This is useful in classification as it gives a certainty measure on classifications.

The softmax activation function is:

${\displaystyle y_i=\frac{e^{x_i}}{\sum _{j=1}^c{e}^{x_j}}}$

Criticism

Training

A common criticism of neural networks, particularly in robotics, is that they require too many training samples for real-world operation. Any learning machine needs sufficient representative examples in order to capture the underlying structure that allows it to generalize to new cases. Potential solutions include randomly shuffling training examples, by using a numerical optimization algorithm that does not take too large steps when changing the network connections following an example, grouping examples in so-called mini-batches and/or introducing a recursive least squares algorithm for CMAC. Dean Pomerleau uses a neural network to train a robotic vehicle to drive on multiple types of roads (single lane, multi-lane, dirt, etc.), and a large amount of his research is devoted to extrapolating multiple training scenarios from a single training experience, and preserving past training diversity so that the system does not become overtrained (if, for example, it is presented with a series of right turns—it should not learn to always turn right).

Theory

A central claim of ANNs is that they embody new and powerful general principles for processing information. These principles are ill-defined. This allows simple statistical association (the basic function of artificial neural networks) to be described as learning or recognition. In 1997, Alexander Dewdney, a former Scientific American columnist, commented that as a result, artificial neural networks have a

something-for-nothing quality, one that imparts a peculiar aura of laziness and a distinct lack of curiosity about just how good these computing systems are. No human hand (or mind) intervenes; solutions are found as if by magic; and no one, it seems, has learned anything.

One response to Dewdney is that neural networks have been successfully used to handle many complex and diverse tasks, ranging from autonomously flying aircraft to detecting credit card fraud to mastering the game of Go.

Technology writer Roger Bridgman commented:

Neural networks, for instance, are in the dock not only because they have been hyped to high heaven, (what hasn't?) but also because you could create a successful net without understanding how it worked: the bunch of numbers that captures its behaviour would in all probability be "an opaque, unreadable table...valueless as a scientific resource".

In spite of his emphatic declaration that science is not technology, Dewdney seems here to pillory neural nets as bad science when most of those devising them are just trying to be good engineers. An unreadable table that a useful machine could read would still be well worth having.

Although it is true that analyzing what has been learned by an artificial neural network is difficult, it is much easier to do so than to analyze what has been learned by a biological neural network. Moreover, recent emphasis on the explainability of AI has contributed towards the development of methods, notably those based on attention mechanisms, for visualizing and explaining learned neural networks. Furthermore, researchers involved in exploring learning algorithms for neural networks are gradually uncovering generic principles that allow a learning machine to be successful. For example, Bengio and LeCun (2007) wrote an article regarding local vs non-local learning, as well as shallow vs deep architecture.

Biological brains use both shallow and deep circuits as reported by brain anatomy, displaying a wide variety of invariance. Weng argued that the brain self-wires largely according to signal statistics and therefore, a serial cascade cannot catch all major statistical dependencies.

Hardware

Large and effective neural networks require considerable computing resources. While the brain has hardware tailored to the task of processing signals through a graph of neurons, simulating even a simplified neuron on von Neumann architecture may consume vast amounts of memory and storage. Furthermore, the designer often needs to transmit signals through many of these connections and their associated neurons – which require enormous CPU power and time.

Some argue that the resurgence of neural networks in the twenty-first century is largely attributable to advances in hardware: from 1991 to 2015, computing power, especially as delivered by GPGPUs (on GPUs), has increased around a million-fold, making the standard backpropagation algorithm feasible for training networks that are several layers deeper than before. The use of accelerators such as FPGAs and GPUs can reduce training times from months to days.

Neuromorphic engineering or a physical neural network addresses the hardware difficulty directly, by constructing non-von-Neumann chips to directly implement neural networks in circuitry. Another type of chip optimized for neural network processing is called a Tensor Processing Unit, or TPU.

Practical counterexamples

Analyzing what has been learned by an ANN is much easier than analyzing what has been learned by a biological neural network. Furthermore, researchers involved in exploring learning algorithms for neural networks are gradually uncovering general principles that allow a learning machine to be successful. For example, local vs. non-local learning and shallow vs. deep architecture.

Hybrid approaches

Advocates of hybrid models (combining neural networks and symbolic approaches) say that such a mixture can better capture the mechanisms of the human mind.

Dataset bias

Neural networks are dependent on the quality of the data they are trained on, thus low quality data with imbalanced representativeness can lead to the model learning and perpetuating societal biases. These inherited biases become especially critical when the ANNs are integrated into real-world scenarios where the training data may be imbalanced due to the scarcity of data for a specific race, gender or other attribute. This imbalance can result in the model having inadequate representation and understanding of underrepresented groups, leading to discriminatory outcomes that exacerbate societal inequalities, especially in applications like facial recognition, hiring processes, and law enforcement. For example, in 2018, Amazon had to scrap a recruiting tool because the model favored men over women for jobs in software engineering due to the higher number of male workers in the field. The program would penalize any resume with the word "woman" or the name of any women's college. However, the use of synthetic data can help reduce dataset bias and increase representation in datasets.

Gallery

• 
  A single-layer feedforward artificial neural network. Arrows originating from ${\displaystyle x_2}$ are omitted for clarity. There are p inputs to this network and q outputs. In this system, the value of the qth output, ${\displaystyle y_q}$, is calculated as ${\displaystyle y_q=K\cdot \left(\sum _i(x_i{w}_{iq})-b_q\right).}$
• 
  A two-layer feedforward artificial neural network
  
  
• 
  An artificial neural network
   
• 
  An ANN dependency graph
   
• 
  A single-layer feedforward artificial neural network with 4 inputs, 6 hidden nodes and 2 outputs. Given position state and direction, it outputs wheel based control values.
   
• 
  A two-layer feedforward artificial neural network with 8 inputs, 2x8 hidden nodes and 2 outputs. Given position state, direction and other environment values, it outputs thruster based control values.
   
• 
  Parallel pipeline structure of CMAC neural network. This learning algorithm can converge in one step.

Recent advancements and future directions

Artificial neural networks (ANNs) have undergone significant advancements, particularly in their ability to model complex systems, handle large data sets, and adapt to various types of applications. Their evolution over the past few decades has been marked by a broad range of applications in fields such as image processing, speech recognition, natural language processing, finance, and medicine.

Image processing

In the realm of image processing, ANNs are employed in tasks such as image classification, object recognition, and image segmentation. For instance, deep convolutional neural networks (CNNs) have been important in handwritten digit recognition, achieving state-of-the-art performance. This demonstrates the ability of ANNs to effectively process and interpret complex visual information, leading to advancements in fields ranging from automated surveillance to medical imaging.

Speech recognition

By modeling speech signals, ANNs are used for tasks like speaker identification and speech-to-text conversion. Deep neural network architectures have introduced significant improvements in large vocabulary continuous speech recognition, outperforming traditional techniques. These advancements have enabled the development of more accurate and efficient voice-activated systems, enhancing user interfaces in technology products.

Natural language processing

In natural language processing, ANNs are used for tasks such as text classification, sentiment analysis, and machine translation. They have enabled the development of models that can accurately translate between languages, understand the context and sentiment in textual data, and categorize text based on content. This has implications for automated customer service, content moderation, and language understanding technologies.

Control systems

In the domain of control systems, ANNs are used to model dynamic systems for tasks such as system identification, control design, and optimization. For instance, deep feedforward neural networks are important in system identification and control applications.

Finance

Further information: Applications of artificial intelligence § Trading and investment

ANNs are used for stock market prediction and credit scoring:

• In investing, ANNs can process vast amounts of financial data, recognize complex patterns, and forecast stock market trends, aiding investors and risk managers in making informed decisions.
• In credit scoring, ANNs offer data-driven, personalized assessments of creditworthiness, improving the accuracy of default predictions and automating the lending process.

ANNs require high-quality data and careful tuning, and their "black-box" nature can pose challenges in interpretation. Nevertheless, ongoing advancements suggest that ANNs continue to play a role in finance, offering valuable insights and enhancing risk management strategies.

Medicine

ANNs are able to process and analyze vast medical datasets. They enhance diagnostic accuracy, especially by interpreting complex medical imaging for early disease detection, and by predicting patient outcomes for personalized treatment planning. In drug discovery, ANNs speed up the identification of potential drug candidates and predict their efficacy and safety, significantly reducing development time and costs. Additionally, their application in personalized medicine and healthcare data analysis allows tailored therapies and efficient patient care management. Ongoing research is aimed at addressing remaining challenges such as data privacy and model interpretability, as well as expanding the scope of ANN applications in medicine.

Content creation

ANNs such as generative adversarial networks (GAN) and transformers are used for content creation across numerous industries. This is because deep learning models are able to learn the style of an artist or musician from huge datasets and generate completely new artworks and music compositions. For instance, DALL-E is a deep neural network trained on 650 million pairs of images and texts across the internet that can create artworks based on text entered by the user. In the field of music, transformers are used to create original music for commercials and documentaries through companies such as AIVA and Jukedeck. In the marketing industry, generative models are used to create personalized advertisements for consumers. Additionally, major film companies are partnering with technology companies to analyze the financial success of a film, such as the partnership between Warner Bros and technology company Cinelytic established in 2020. Furthermore, neural networks have found uses in video game creation, where non-player characters (NPCs) can make decisions based on all the characters currently in the game.