Inferring neural activity before plasticity as a foundation for learning beyond backpropagation

Area of Science:

• Neuroscience
• Machine Learning
• Cognitive Science

Background:

• Credit assignment is crucial for learning in both biological and artificial systems, identifying error sources in information processing.
• Backpropagation is the dominant method for credit assignment in modern machine learning and is widely assumed to be biologically plausible.

Purpose of the Study:

• To introduce and explore a novel principle for credit assignment termed 'prospective configuration'.
• To investigate the biological plausibility and effectiveness of prospective configuration compared to backpropagation.

Main Methods:

• Proposed a new credit assignment mechanism: prospective configuration, where neural activity patterns are inferred first, followed by synaptic weight adjustment.
• Evaluated prospective configuration's performance in established models of cortical circuits.
• Compared prospective configuration with backpropagation in simulated learning scenarios relevant to biological organisms.

Main Results:

• Demonstrated that prospective configuration underlies learning in models of cortical circuits.
• Showed that prospective configuration learning is more efficient and effective in various contexts relevant to biological organisms.
• Observed that prospective configuration reproduces surprising neural activity and behavioral patterns seen in human and rat learning experiments.

Conclusions:

• Prospective configuration presents a fundamentally different and potentially more effective principle for credit assignment than backpropagation.
• This mechanism offers a plausible explanation for observed learning phenomena in biological systems, suggesting a new direction for neuroscience and AI.