Predictive Coding: a Theoretical and Experimental Review (2107.12979v4)

Abstract: Predictive coding offers a potentially unifying account of cortical function -- postulating that the core function of the brain is to minimize prediction errors with respect to a generative model of the world. The theory is closely related to the Bayesian brain framework and, over the last two decades, has gained substantial influence in the fields of theoretical and cognitive neuroscience. A large body of research has arisen based on both empirically testing improved and extended theoretical and mathematical models of predictive coding, as well as in evaluating their potential biological plausibility for implementation in the brain and the concrete neurophysiological and psychological predictions made by the theory. Despite this enduring popularity, however, no comprehensive review of predictive coding theory, and especially of recent developments in this field, exists. Here, we provide a comprehensive review both of the core mathematical structure and logic of predictive coding, thus complementing recent tutorials in the literature. We also review a wide range of classic and recent work within the framework, ranging from the neurobiologically realistic microcircuits that could implement predictive coding, to the close relationship between predictive coding and the widely-used backpropagation of error algorithm, as well as surveying the close relationships between predictive coding and modern machine learning techniques.

• The paper presents a comprehensive theoretical framework with rigorous variational inference derivations of predictive coding.
• It demonstrates the scalability and efficiency of local learning mechanisms, comparing them favorably to deep neural network paradigms.
• The review bridges neurobiological insights with modern AI advancements, unifying experimental findings with computational theory.

Insights into "Predictive Coding: a Theoretical and Experimental Review"

The paper "Predictive Coding: a Theoretical and Experimental Review" offers a comprehensive examination of Predictive Coding (PC) as a potentially unifying framework at the intersection of artificial intelligence and the brain sciences. Authored by Beren Millidge, Alexander Tschantz, Anil Seth, and Christopher L. Buckley, this manuscript explores the computational and neurophysiological dimensions of PC, providing both a detailed theoretical foundation and an overview of recent advancements.

Predictive Coding, which originated from foundational work by Mumford and Rao, has been progressively recognized for its theoretical promise in explaining cortical functions through the lens of information theory. The manuscript meticulously explores the mathematical underpinnings of PC, adopting a variational inference perspective over hierarchical Gaussian generative models. This mathematical rigor not only clarifies the core principles but also paves the way for understanding PC's applicability to both localized and parallelizable learning paradigms analogous to backpropagation.

Key highlights from this paper include detailed mathematical derivations, discussions on algorithmic innovations over the recent past, and comparisons of PC with established machine learning algorithms. These demonstrate the capability of PC networks to scale effectively, potentially rivaling deep neural networks in a broad spectrum of tasks. This scaling aligns with the need for highly efficient and adaptable computational models, emphasizing the relevance of PC’s local learning mechanisms.

Furthermore, this review distinguishes itself from prior literature by its breadth and depth, notably covering recent theoretical expansions and providing a comprehensive review of neurobiological arguments supporting PC. The authors aim to integrate these findings, forming a coherent narrative that underlines the synthesis between predictive coding principles and neurophysiological processes.

In addressing previous critiques and limitations, the authors highlight differences from earlier reviews such as those by Spratling, Bogacz, and others. By incorporating modern advances and connections to machine learning, this paper intends to bridge theoretical gaps and streamline the introduction of PC’s mathematical formulation, which could facilitate understanding for both newcomers and seasoned researchers within the field.

The manuscript has implications on both practical and theoretical fronts. Practically, the scalability and efficiency of PC networks suggest possible advancements in artificial intelligence applications, especially where computational resources and energy efficiency are paramount. Theoretically, the consolidation of predictive coding as a model for brain function pushes the boundaries of current understanding of neural processing, potentially informing new computational techniques and contributing to the broader field of computational neuroscience.

While the paper offers significant insight, future developments in the domain might explore how predictive coding can be integrated with or distinguished from other learning paradigms, such as reinforcement learning or unsupervised learning techniques. The evolution of PC as a dominant model for both biological and artificial systems remains a key area for ongoing research and discussion.

This review constitutes a substantial contribution to the field, providing not only an expert consolidation of predictive coding theory but also inciting further discussions on its role and impact in artificial intelligence and neurobiology. The comprehensive nature of this manuscript renders it a promising reference for understanding and advancing the discourse on predictive coding.