Cognitive Computational Neuroscience

Cognitive science has developed computational models that decompose cognition into functional components. Computational neuroscience has modeled how interacting neurons can implement elementary components of cognition.
It is time to assemble the pieces of the puzzle of brain computation and to better integrate these separate disciplines. 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.
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.

Cognitive neuroscience has mapped the global functional layout of the human and nonhuman primate brain. However, it has not achieved a full computational account of brain information processing.
The challenge ahead is to build computational models of brain information processing that are consistent with brain structure and function and perform complex cognitive tasks. The following recent developments in cognitive science, computational neuroscience and artificial intelligence suggest that this may be achievable.

1. Cognitive science has proceeded from the top down, decomposing complex cognitive processes into their computational components.
   Unencumbered by the need to make sense of brain data, it has developed task-performing computational models at the cognitive level. One success story is that of Bayesian cognitive models, which optimally combine prior knowledge about the world with sensory evidence. Initially applied to basic sensory and motor processes, Bayesian models have begun to engage complex cognition, including the way our minds model the physical and social world. These developments occurred in interaction with statistics and machine learning, where a unified perspective on probabilistic empirical inference has emerged.
2. Computational neuroscience has taken a bottom-up approach, demonstrating how dynamic interactions between biological neurons can implement computational component functions. In the past two decades, the field developed mathematical models of elementary computational components and their implementation with biological neurons. These include components for sensory coding, normalization, working memory, evidence accumulation and decision mechanisms, and motor control. Most of these component functions are computationally simple, but they provide
   building blocks for cognition.
   Computational neuroscience has also begun to test complex computational models that can explain high level sensory and cognitive brain representations.
3. Artificial intelligence has shown how component functions can be combined to create intelligent behavior. Early AI failed to live up to its promise because the rich world knowledge required for feats of intelligence could not be either engineered or automatically learned. Recent advances in machine learning, boosted by growing computational power and larger datasets from which to learn, have brought progress at perceptual, cognitive and control challenges.
   Many advances were driven by cognitive-level symbolic models.
   Some of the most important recent advances are driven by deep neural network models, composed of units that compute linear combinations of their inputs, followed by static nonlinearities. These models employ only a small subset of the dynamic capabilities of biological neurons, abstracting from fundamental features such as action potentials. However, their functionality is inspired by brains and could be implemented with biological neurons.

---

The three disciplines contribute complementary elements to biologically plausible computational models that perform cognitive tasks and explain brain information processing and behavior. If computational models are to explain animal and human cognition, they will have to perform feats of intelligence. AI, and in particular machine learning, is therefore a key discipline that provides the theoretical and technological foundation for cognitive computational neuroscience.

The word “model” has many meanings in the brain and behavioral sciences.
Data-analysis models are generic statistical models that help establish relationships between measured variables. Examples include linear correlation, univariate multiple linear regression for brain mapping, and linear decoding analysis.
Effective connectivity and causal-interaction models are, similarly,
data-analysis models. They help us infer causal influences and interactions between brain regions. Data-analysis models can serve the purpose of testing hypotheses about relationships among variables (for example, correlation, information, causal influence). They are not models of brain information processing.
A box-and-arrow model, by contrast, is an information processing model in the form of labeled boxes that represent cognitive component functions and arrows that represent information flow. In cognitive psychology, such models provided
useful, albeit ill-defined, sketches for theories of brain computation.
A word model, similarly, is a sketch for a theory about brain information processing that is defined vaguely by a verbal description. While these are models of information processing, they do not perform the information processing thought to occur in the brain.
An oracle model is a model of brain responses (often instantiated in a data-analysis model) that relies on information not available to the animal whose brain is being modeled. For example, a model of ventral temporal visual responses as a function of an abstract shape description, or as a function of category labels or continuous semantic features, constitutes an oracle model if the model is not capable of computing the shape, category or semantic features from images. An oracle model may provide a useful characterization of the information present in a region and its representational format, without specifying any theory as to how the representation is computed by the brain.
A brain-computational model (BCM), by contrast, is a model that mimics the brain information processing underlying the performance of some task at some level of abstraction. In visual neuroscience, for example, an image-computable model is a BCM of visual processing that takes image bitmaps as inputs and predicts brain activity and/or behavioral responses. Deep neural nets provide image-computable models of visual processing. However, deep neural nets trained by supervision rely on category-labeled images for training. Because labeled examples are not available (in comparable quantities) during biological development and learning, these models are BCMs of visual processing, but they are not BCMs of development and learning.
Reinforcement learning models use environmental feedback that is more realistic in quality and can provide BCMs of learning processes.
A sensory encoding model is a BCM of the computations that transform sensory input to some stage of internal representation.
An internal-transformation model is a BCM of the transformation of representations between two stages of processing.
A behavioral decoding model is a BCM of the transformation from some internal representation to a behavioral output. Note that the label BCM indicates merely that the model is intended to capture brain computations at some level of abstraction. A BCM may abstract from biological detail to an arbitrary degree, but must predict some aspect of brain activity and/or behavior.
Psychophysical models that predict behavioral outputs from sensory input and cognitive models that perform cognitive tasks are BCMs formulated at a high level of description. The label BCM does not imply that the model is either plausible or consistent with empirical data. Progress is made by rejecting candidate BCMs on empirical grounds. Like microscale biophysical models, which capture biological processes that underlie brain computations, and macroscale brain-dynamical and causal-interaction models, BCMs are models of processes occurring in the brain. However, unlike the other types of process model, BCMs perform the information processing that is thought to be the function of brain dynamics.
Finally, the term “model” is used to refer to models of the world employed by the brain, as in model-based reinforcement learning and model-based cognition.

---

Why do cognitive science, computational neuroscience and AI need one another?

• Cognitive science needs computational neuroscience, not merely to explain the implementation of cognitive models in the brain, but also to discover the algorithms. For example, the dominant models of sensory processing and object recognition are brain-inspired neural networks, whose computations are not easily captured at a cognitive level. Recent successes with Bayesian nonparametric models do not yet in general scale to real-world cognition. Explaining the computational efficiency of human cognition and predicting detailed cognitive dynamics and behavior could benefit from studying brain-activity dynamics.
  Explaining behavior is essential, but behavioral data alone provide insufficient constraints for complex models. Brain data can provide rich constraints for cognitive algorithms if leveraged appropriately. Cognitive science has always progressed in close interaction with artificial intelligence. The disciplines share the goal of building task-performing models and thus rely on common mathematical theory and programming environments.
• Computational neuroscience needs cognitive science to challenge it to engage higher-level cognition. At the experimental level, the tasks of cognitive science enable computational neuroscience to bring cognition into the lab. At the level of theory, cognitive science challenges computational neuroscience to explain how the neurobiological dynamical components it studies contribute to cognition and behavior. Computational neuroscience needs AI, and in particular machine learning, to provide the theoretical and technological basis for modeling cognitive functions with biologically plausibledynamical components.
• Artificial intelligence needs cognitive science to guide the engineering of intelligence. Cognitive science’s tasks can serve as benchmarks for AI systems, building up from elementary cognitive abilities to artificial general intelligence. The literatures on human development and learning provide an essential guide to what is possible for a learner to achieve and what kinds of interaction with the world can support the acquisition of intelligence. AI needs computational neuroscience for algorithmic inspiration. Neural network models are an example of a brain-inspired technology that is unrivalled in several domains of AI. Taking further inspiration from the neurobiological dynamical components (for example, spiking neurons, dendritic dynamics, the canonical cortical microcircuit, oscillations, neuromodulatory processes) and the global functional layout of the human brain (for example, subsystems specialized for distinct functions, including sensory modalities, memory, planning and motor control) might lead to further AI breakthroughs. Machine learning draws from separate traditions in statistics and computer science, which have optimized statistical and computational efficiency, respectively. The integration of computational and statistical efficiency is an essential challenge in the age of big data. The brain appears to combine computational and statistical efficiency, and understanding its algorithm might boost machine learning.

Bottom up and top down. The brain seamlessly merges bottom up discriminative and top-down generative computations in perceptual inference, and model-free and model-based control. Brain science likewise needs to integrate its levels of description and to progress both bottom-up and top-down, so as to explain task performance on the basis of neuronal dynamics and provide a mechanistic account of how the brain gives rise to the mind.
Bottom-up visions, proceeding from detailed measurements toward an understanding of brain computation, have been prominent and have driven the most important recent funding initiatives. The European Human Brain Project and the US BRAIN Initiative are both motivated by bottom-up visions, in which an understanding of brain computation is achieved by measuring and modeling brain dynamics with a focus on the circuit level.
Understanding the brain requires that we develop theory and experiment in tandem and complement the bottom-up, data driven approach by a top-down, theory-driven approach that starts with behavioral functions to be explained. Unprecedentedly rich measurements and manipulations of brain activity will drive theoretical insight when they are used to adjudicate between brain-computational models that pass the first test of being able to perform a function that contributes to the behavioral fitness of the organism.
The top-down approach, therefore, is an essential complement to the bottom-up approach toward understanding the brain.