A survey on neuro-mimetic deep learning via predictive coding

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Current AI predominantly uses deep neural networks trained with error backpropagation, a method questioned for biological plausibility.
• Neuroscience-inspired learning algorithms are emerging as alternatives to traditional AI training methods.
• Predictive coding (PC) is a neuroscience theory with potential for AI applications.

Purpose of the Study:

• To survey recent advancements in predictive coding (PC) inspired algorithms for deep neural networks.
• To provide a historical overview of PC to establish a foundation for understanding current developments.
• To discuss the implications and future directions of PC in machine learning and AI.

Main Methods:

• Review of existing literature on predictive coding (PC) and its application in AI.
• Analysis of PC's properties, including its biological plausibility, mathematical foundation in variational inference, and asynchronous computation.
• Synthesis of current research efforts and results in PC-based machine learning algorithms.

Main Results:

• Predictive coding (PC) demonstrates potential for modeling brain information processing.
• PC algorithms show promise in control, robotics, and various machine learning sub-fields.
• Novel PC-like algorithms are increasingly being developed and applied across AI.

Conclusions:

• Predictive coding (PC) presents a biologically plausible and mathematically robust framework for AI.
• The continued development of PC-inspired algorithms could significantly advance machine learning and artificial intelligence.
• Further research into PC holds promise for future AI innovations and applications.