Showing posts with label Bayesian Inference. Show all posts
Showing posts with label Bayesian Inference. Show all posts

Artificial Intelligence - What Is Computational Neuroscience?


Computational neuroscience (CNS) is a branch of neuroscience that uses the notion of computing to the study of the brain.

Eric Schwartz coined the phrase "computational neuroscience" in 1985 to replace the words "neural modeling" and "brain theory," which were previously used to describe different forms of nervous system study.

The concept that nervous system effects may be perceived as examples of computations, since state transitions can be explained as relations between abstract attributes, is at the heart of CNS.

In other words, explanations of effects in neurological systems are descriptions of information changed, stored, and represented, rather than casual descriptions of interaction of physically distinct elements.

As a result, CNS aims to develop computational models to better understand how the nervous system works in terms of the information processing characteristics of the brain's parts.

Constructing a model of how interacting neurons might build basic components of cognition is one example.

A brain map, on the other hand, does not disclose the nervous system's computing process, but it may be utilized as a restriction for theoretical models.

Information sharing, for example, has costs in terms of physical connections between communicating areas, in that locations that make connections often (in cases of high bandwidth and low latency) would be clustered together.

The description of neural systems as carry-on computations is central to computational neuroscience, and it contradicts the claim that computational constructs are exclusive to the explanatory framework of psychology; that is, human cognitive capacities can be constructed and confirmed independently of how they are implemented in the nervous system.

For example, when it became clear in 1973 that cognitive processes could not be understood by analyzing the results of one-dimensional questions/scenarios, a popular approach in cognitive psychology at the time, Allen Newell argued that only synthesis with computer simulation could reveal the complex interactions of the proposed component's mechanism and whether the proposed component's mechanism was correct.

David Marr (1945–1980) proposed the first computational neuroscience framework.

This framework tries to give a conceptual starting point for thinking about levels in the context of computing by nervous structure.

It reflects the three-level structure utilized in computer science (abstract issue analysis, algorithm, and physical implementation).

The model, however, has drawbacks since it is made up of three poorly linked layers and uses a rigid top-down approach that ignores all neurobiological facts as instances at the implementation level.

As a result, certain events are thought to be explicable on just one or two levels.

As a result, the Marr levels framework does not correspond to the levels of nervous system structure (molecules, synapses, neurons, nuclei, circuits, networks layers, maps, and systems), nor does it explain nervous system emergent type features.

Computational neuroscience takes a bottom-up approach, beginning with neurons and illustrating how computational functions and their implementations with neurons result in dynamic interactions between neurons.

Models of connectivity and dynamics, decoding models, and representational models are the three kinds of models that try to get computational understanding from brain-activity data.

The correlation matrix, which displays pairwise functional connectivity between places and establishes the features of related areas, is used in connection models.

Because they are generative models, they can generate data at the level of the measurements and are models of brain dynamics, analyses of effective connectivity and large-scale brain dynamics go beyond generic statistical models that are linear models used in action and information-based brain mapping.

The goal of the decoding models is to figure out what information is stored in each brain area.

When an area is designated as a "knowledge representing" one, its data becomes a functional entity that informs regions that receive these signals about the content.

In the simplest scenario, decoding identifies which of the two stimuli elicited a recorded response pattern.

The representation's content might be the sensory stimulus's identity, a stimulus feature (such as orientation), or an abstract variable required for a cognitive operation or action.

Decoding and multivariate pattern analysis were utilized to determine the components that must be included in the brain computational model.

Decoding, on the other hand, does not provide models for brain computing; rather, it discloses some elements without requiring brain calculation.

Because they strive to characterize areas' reactions to arbitrary stimuli, representation models go beyond decoding.

Encoding models, pattern component models, and representational similarity analysis are three forms of representational model analysis that have been presented.

All three studies are based on multivariate descriptions of the experimental circumstances and test assumptions about representational space.

In encoding models, the activity profile of each voxel across stimuli is predicted as a linear combination of the model's properties.

The distribution of the activity profiles that define the representational space is treated as a multivariate normal distribution in pattern component models.

The representational space is defined by the representational dissimilarities of the activity patterns evoked by the stimuli in representational similarity analysis.

The qualities that indicate how the information processing cognitive function could operate are not tested in the brain models.

Task performance models are used to describe cognitive processes in terms of algorithms.

These models are put to the test using experimental data and, in certain cases, data from brain activity.

Neural network models and cognitive models are the two basic types of models.

Models of neural networks are created using varying degrees of biological information, ranging from neurons to maps.

Multiple steps of linear-nonlinear signal modification are supported by neural networks, which embody the parallel distributed processing paradigm.

To enhance job performance, models often incorporate millions of parameters (connection weights).

Simple models will not be able to describe complex cognitive processes, hence a high number of parameters is required.

The implementations of deep convolutional neural network models have been used to predict brain representations of new pictures in the ventral visual stream of primates.

The representations in the first few layers of neural networks are comparable to those in the early visual cortex.

Higher layers are similar to the inferior temporal cortical representation in that they both allow for the decoding of object location, size, and posture, as well as the object's categorization.

Various research have shown that deep convolutional neural networks' internal representations provide the best current models of visual picture representations in the inferior temporal cortex in humans and animals.

When a wide number of models were compared, those that were optimized for object categorization described the cortical representation the best.

Cognitive models are artificial intelligence applications in computational neuroscience that target information processing that do not include any neurological components (neurons, axons, etc.).

Production systems, reinforcement learning, and Bayesian cognitive models are the three kinds of models.

They use logic and predicates, and they work with symbols rather than signals.

There are various advantages of employing artificial intelligence in computational neuroscience research.

  1. First, although a vast quantity of information on the brain has accumulated through time, the true knowledge of how the brain functions remains unknown.
  2. Second, there are embedded effects created by networks of neurons, but how these networks of neurons operate is yet unknown.
  3. Third, although the brain has been crudely mapped, as has understanding of what distinct brain areas (mostly sensory and motor functions) perform, a precise map is still lacking.

Furthermore, some of the information gathered via experiments or observations may be useless; the link between synaptic learning principles and computing is mostly unclear.

The models of a production system are the first models for explaining reasoning and problem resolution.

A "production" is a cognitive activity that occurs as a consequence of the "if-then" rule, in which "if" defines the set of circumstances under which the range of productions ("then" clause) may be carried out.

When the prerequisites for numerous rules are satisfied, the model uses a conflict resolution algorithm to choose the best production.

The production models provide a sequence of predictions that seem like a conscious stream of brain activity.

The same approach is now being used to predict the regional mean fMRI (functional Magnetic Resonance Imaging) activation time in new applications.

Reinforcement models are used in a variety of areas to simulate the accomplishment of optimum decision-making.

The implementation in neurobiological systems is a basal ganglia in neurobiochemical systems.

The agent might learn a "value function" that links each state to the predicted total reward.

The agent may pick the most promising action if it can forecast which state each action will lead to and understands the values of those states.

The agent could additionally pick up a "policy" that links each state to promised actions.

Exploitation (which provides immediate gratification) and exploration must be balanced (which benefits learning and brings long-term reward).

The Bayesian models show what the brain should really calculate in order to perform at its best.

These models enable inductive inference, which is beyond the capability of neural network models and requires previous knowledge.

The models have been used to explain cognitive biases as the result of past beliefs, as well as to comprehend fundamental sensory and motor processes.

The representation of the probability distribution of neurons, for example, has been investigated theoretically using Bayesian models and compared to actual evidence.

These practices illustrate that connecting Bayesian inference to real brain implementation is still difficult since the brain "cuts corners" in trying to be efficient, therefore approximations may explain departures from statistical optimality.

The concept of a brain doing computations is central to computational neuroscience, so researchers are using modeling and analysis of information processing properties of nervous system elements to try to figure out how complex brain functions work.

~ Jai Krishna Ponnappan

You may also want to read more about Artificial Intelligence here.

See also: 

Bayesian Inference; Cognitive Computing.

Further Reading

Kaplan, David M. 2011. “Explanation and Description in Computational Neuroscience.” Synthese 183, no. 3: 339–73.

Kriegeskorte, Nikolaus, and Pamela K. Douglas. 2018. “Cognitive Computational Neuroscience.” Nature Neuroscience 21, no. 9: 1148–60.

Schwartz, Eric L., ed. 1993. Computational Neuroscience. Cambridge, MA: Massachusetts Institute of Technology.

Trappenberg, Thomas. 2009. Fundamentals of Computational Neuroscience. New York: Oxford University Press.

Artificial Intelligence - What Is Bayesian Inference?


Bayesian inference is a method of calculating the likelihood of a proposition's validity based on a previous estimate of its likelihood plus any new and relevant facts.

In the twentieth century, Bayes' Theorem, from which Bayesian statistics are derived, was a prominent mathematical technique employed in expert systems.

The Bayesian theorem has been used to issues such as robot locomotion, weather forecasting, juri metry (the application of quantitative approaches to legislation), phylogenetics (the evolutionary links among animals), and pattern recognition in the twenty-first century.

It's also used in email spam filters and can be used to solve the famous Monty Hall issue.

The mathematical theorem was derived by Reverend Thomas Bayes (1702–1761) of England and published posthumously in the Philosophical Transactions of the Royal Society of London in 1763 as "An Essay Towards Solving a Problem in the Doctrine of Chances." Bayes' Theorem of Inverse Probabilities is another name for it.

A classic article titled "Reasoning Foundations of Medical Diagnosis," written by George Washington University electrical engineer Robert Ledley and Rochester School of Medicine radiologist Lee Lusted and published by Science in 1959, was the first notable discussion of Bayes' Theorem as applied to the field of medical artificial intelligence.

Medical information in the mid-twentieth century was frequently given as symptoms connected with an illness, rather than diseases associated with a symptom, as Lusted subsequently recalled.

They came up with the notion of expressing medical knowledge as the likelihood of a disease given the patient's symptoms using Bayesian reasoning.

Bayesian statistics are conditional, allowing one to determine the likelihood that a specific disease is present based on a specific symptom, but only with prior knowledge of how frequently the disease and symptom are correlated, as well as how frequently the symptom is present in the absence of the disease.

It's pretty similar to what Alan Turing called the evidence-based element in support of the hypothesis.

The symptom-disease complex, which involves several symptoms in a patient, may also be resolved using Bayes' Theorem.

In computer-aided diagnosis, Bayesian statistics analyzes the likelihood of each illness manifesting in a population with the chance of each symptom manifesting given each disease to determine the probability of all possible diseases given each patient's symptom-disease complex.

All induction, according to Bayes' Theorem, is statistical.

In 1960, the theory was used to generate the posterior probability of certain illnesses for the first time.

In that year, University of Utah cardiologist Homer Warner, Jr.

used Bayesian statistics to detect well-defined congenital heart problems at Salt Lake's Latter-Day Saints Hospital, thanks to his access to a Burroughs 205 digital computer.

The theory was used by Warner and his team to calculate the chances that an undiscovered patient having identifiable symptoms, signs, or laboratory data would fall into previously recognized illness categories.

As additional information became available, the computer software could be employed again and again, creating or rating diagnoses via serial observation.

The Burroughs computer outperformed any professional cardiologist in applying Bayesian conditional-probability algorithms to a symptom-disease matrix of thirty-five cardiac diseases, according to Warner.

John Overall, Clyde Williams, and Lawrence Fitzgerald for thyroid problems; Charles Nugent for Cushing's illness; Gwilym Lodwick for primary bone tumors; Martin Lipkin for hematological diseases; and Tim de Dombal for acute abdominal discomfort were among the early supporters of Bayesian estimation.

In the previous half-century, the Bayesian model has been expanded and changed several times to account for or compensate for sequential diagnosis and conditional independence, as well as to weight other elements.

Poor prediction of rare diseases, insufficient discrimination between diseases with similar symptom complexes, inability to quantify qualitative evidence, troubling conditional dependence between evidence and hypotheses, and the enormous amount of manual labor required to maintain the requisite joint probability distribution tables are all criticisms leveled at Bayesian computer-aided diagnosis.

Outside of the populations for which they were intended, Bayesian diagnostic helpers have been chastised for their shortcomings.

When rule-based decision support algorithms became more prominent in the mid-1970s, the application of Bayesian statistics in differential diagnosis reached a low.

In the 1980s, Bayesian approaches resurfaced and are now extensively employed in the area of machine learning.

From the concept of Bayesian inference, artificial intelligence researchers have developed robust techniques for supervised learning, hidden Markov models, and mixed approaches for unsupervised learning.

Bayesian inference has been controversially utilized in artificial intelligence algorithms that aim to calculate the conditional chance of a crime being committed, to screen welfare recipients for drug use, and to identify prospective mass shooters and terrorists in the real world.

The method has come under fire once again, especially when screening includes infrequent or severe incidents, where the AI system might act arbitrarily and flag too many people as being at danger of partaking in the unwanted behavior.

In the United Kingdom, Bayesian inference has also been used into the courtroom.

The defense team in Regina v.

Adams (1996) offered jurors the Bayesian approach to aid them in forming an unbiased mechanism for combining introduced evidence, which included a DNA profile and varying match probability calculations, as well as constructing a personal threshold for convicting the accused "beyond a reasonable doubt." Before Ledley, Lusted, and Warner revived Bayes' theorem in the 1950s, it had previously been "rediscovered" multiple times.

Pierre-Simon Laplace, the Marquis de Condorcet, and George Boole were among the historical figures who saw merit in the Bayesian approach to probability.

The Monty Hall dilemma, named after the presenter of the famous game show Let's Make a Deal, involves a contestant selecting whether to continue with the door they've chosen or swap to another unopened door when Monty Hall (who knows where the reward is) opens one to reveal a goat.

Switching doors, contrary to popular belief, doubles your odds of winning under conditional probability.

~ Jai Krishna Ponnappan

You may also want to read more about Artificial Intelligence here.

See also: 

Computational Neuroscience; Computer-Assisted Diagnosis.

Further Reading

Ashley, Kevin D., and Stefanie BrĂ¼ninghaus. 2006. “Computer Models for Legal Prediction.” Jurimetrics 46, no. 3 (Spring): 309–52.

Barnett, G. Octo. 1968. “Computers in Patient Care.” New England Journal of Medicine
279 (December): 1321–27.

Bayes, Thomas. 1763. “An Essay Towards Solving a Problem in the Doctrine of Chances.” 
Philosophical Transactions 53 (December): 370–418.

Donnelly, Peter. 2005. “Appealing Statistics.” Significance 2, no. 1 (February): 46–48.
Fox, John, D. Barber, and K. D. Bardhan. 1980. “Alternatives to Bayes: A Quantitative 
Comparison with Rule-Based Diagnosis.” Methods of Information in Medicine 19, 
no. 4 (October): 210–15.

Ledley, Robert S., and Lee B. Lusted. 1959. “Reasoning Foundations of Medical Diagnosis.” Science 130, no. 3366 (July): 9–21.

Lusted, Lee B. 1991. “A Clearing ‘Haze’: A View from My Window.” Medical Decision 
Making 11, no. 2 (April–June): 76–87.

Warner, Homer R., Jr., A. F. Toronto, and L. G. Veasey. 1964. “Experience with Bayes’ 
Theorem for Computer Diagnosis of Congenital Heart Disease.” Annals of the 
New York Academy of Sciences 115: 558–67.

What Is Artificial General Intelligence?

Artificial General Intelligence (AGI) is defined as the software representation of generalized human cognitive capacities that enables the ...