Source-linked AI summary
Cognitive computational neuroscience
Nikolaus Kriegeskorte, Pamela K. Douglas
TL;DR
The paper addresses the gap between cognitive models that perform tasks and neural models that explain biological computation. It reviews task-performing computational models tested against brain and behavioral data, finding emerging evidence that biologically plausible models can connect cognition with brain information processing. The paper also notes that such models face substantial complexity and modeling limitations.
Problem
Cognitive science lacks neurobiological grounding, while computational neuroscience has not explained how neural components interact to produce human cognition and behavior.
Method
The paper reviews computational models that perform cognitive tasks and are evaluated with brain-activity and behavioral data across cognitive science, computational neuroscience, and AI.
Results
Recent neural network studies provide the best current models of visual-image representations in inferior temporal cortex, with task-optimized object-classification models explaining those representations better.
Takeaways & Limitations
Understanding the brain requires theory and experiment in tandem, using rich measurements to adjudicate among task-performing brain-computational models.
Takeaways & Limitations
Bayesian inference can be statistically optimal but computationally challenging, and inevitable generative-model misspecification can make inference non-optimal.
Abstract
from arXiv · showhide
To learn how cognition is implemented in the brain, we must build computational models that can perform cognitive tasks, and test such models with brain and behavioral experiments. Cognitive science has developed computational models of human cognition, decomposing task performance into computational components. However, its algorithms still fall short of human intelligence and are not grounded in neurobiology. Computational neuroscience has investigated how interacting neurons can implement component functions of brain computation. However, it has yet to explain how those components interact to explain human cognition and behavior. Modern technologies enable us to measure and manipulate brain activity in unprecedentedly rich ways in animals and humans. However, experiments will yield theoretical insight only when employed to test brain-computational models. It is time to assemble the pieces of the puzzle of brain computation. Here we review recent work in the intersection of cognitive science, computational neuroscience, and artificial intelligence. Computational models that mimic brain information processing during perceptual, cognitive, and control tasks are beginning to be developed and tested with brain and behavioral data.
From experiment toward theory
Connectivity, decoding, and representational analyses extract increasingly informative structure from brain-activity data, but they stop short of specifying the mechanistic computations underlying cognition.
- Connectivity models: Connectivity models characterize interactions among brain regions across scales, using anatomical or functional graphs and generative models of dynamics.Effective-connectivity methods instead compare candidate causal-interaction models for task-relevant regions.
- Connectivity models: Whole-brain and effective-connectivity models provide high-level descriptions of interactions and can track dynamical-state variation within and across individuals.Examples include changes across states of consciousness and disorder-associated differences.
- Decoding models: Decoding models read out stimulus, object, face, belief, attention, and working-memory information from regional activity patterns.Decoding establishes that particular information is present, but functional claims about how it informs other regions or behavior require further analysis.
- Representational models: Representational models test hypotheses about representational spaces using encoding models, pattern component models, and representational similarity analysis.When models generalize to novel stimuli, they can provide stronger constraints and help adjudicate among brain-computational models.
- Limitations: These analyses provide theoretically motivated evidence, but without task-performing mechanistic models they do not explain precisely how cognitive information processing works.Large-scale dynamics models also fail to capture the information exchanged or processing occurring in the brain.
Box 1: The many meanings of model
The paper distinguishes statistical data-analysis models from computational models that specify information processing and perform tasks, while emphasizing trade-offs between cognitive and biological fidelity.
- Meanings of model: Data-analysis models describe relationships among measured variables but are not models of brain information processing.Examples include correlation, regression, linear decoding, and effective-connectivity analyses.
- Meanings of model: Box-and-arrow and verbal models sketch cognitive information processing, but their mechanisms are respectively ill-defined or vaguely specified.They represent component functions and information flow without fully implementing the proposed computation.
- Computational models: A computational model mathematically specifies a theory and implements it so that the theory can be tested as a task-performing system.Models can occupy different descriptive levels, trading off cognitive fidelity against biological fidelity.
- Neural network models: Neural network models connect computational neuroscience, cognitive science, and AI by implementing task performance through biologically inspired network computations.They can contain feedforward and recurrent computations and multiple linear-nonlinear transformation stages.
- Model complexity: Large models with many parameters are difficult to understand, but their internal representations can be probed extensively and evaluated for overfitting.The paper argues that complex intelligence requires sufficient parametric complexity to store world knowledge.
Box 2: Neural network models
Neural network models abstract biological details while learning task-relevant representations, providing models of brain information processing and behavior. Their successes motivate integrating richer biological dynamics and connecting cognitive-level models with neural and behavioral data.
- Model abstraction: Neural network models process inputs in parallel through units that abstract from detailed neuronal biology, while still explaining some cognitive functions.A typical unit combines inputs and applies a static nonlinearity; deep networks can efficiently represent complex functions.
- Learning: Neural networks learn parameters by adjusting connection weights to reduce output error, commonly using gradient descent.Both feedforward and recurrent networks are defined by their architecture and connection weights.
- Biological plausibility: Whether the brain uses backpropagation remains controversial, although biologically plausible implementations and internally generated supervision signals have been proposed.Candidate supervision signals may arise from multisensory context, evolving representations, memory, or reinforcement signals.
- Future directions: Integrating biological dynamics such as action potentials, microcircuits, dendritic dynamics, and oscillations into task-performing models may reveal their computational functions.Cognitive-level models meanwhile provide abstractions for higher cognition that current neural network models do not yet capture well.
- Brain-data tests: Deep convolutional networks trained for object recognition currently provide the best models of image representations in inferior temporal cortex in humans and monkeys.Among many models, those optimized for object classification better explained IT representations.
- Brain-data tests: Representations in successive network layers resemble progressively later stages of the ventral visual stream and predict behavioral responses involving shape and object similarity.Higher layers also support decoding object position, size, pose, and category, paralleling IT cortex.
Box 3: Bayesian cognitive models
Bayesian cognitive models frame perception as inference over causes using prior knowledge and sensory evidence, while highlighting the computational challenges of realistic generative models. The broader framework argues that explaining cognition requires integrating cognitive, computational, and neurobiological approaches through task-performing models and shared empirical tests.
- Bayesian inference: Bayesian cognitive models assume that the brain approximates statistically optimal inference by combining current evidence with prior knowledge.For vision, the relevant inference concerns causes in the world that could have generated sensory data.
- Bayesian inference: A generative model represents the joint distribution p(d, c) as the product of a prior p(c) and likelihood p(d|c).The prior captures knowledge about possible causes, while the likelihood describes images produced by those causes.
- Computational constraints: Realistic Bayesian inference is computationally challenging because exact models may require more neurons or time than an animal can use.Model misspecification is inevitable in real-world tasks, and posterior computation may require iterative algorithms such as MCMC, belief propagation, or variational inference.
- Multidisciplinary integration: Cognitive science, computational neuroscience, and AI are presented as mutually necessary for connecting cognitive functions, neural dynamics, and intelligent behavior.Their integration is intended to support neurobiologically plausible models that perform cognitive tasks and can be tested against brain and behavioral data.
- Multidisciplinary integration: Progress requires theory and experiment in tandem, with brain measurements used to adjudicate among models that can perform behaviorally relevant functions.The proposed research culture also emphasizes shared tasks, data, models, and quantitative tests across disciplines and laboratories.