RELEVANCE AMPLIFICATION FOR FEW-SHOT LEARNING WITH LARGE REPRESENTATIONS

The brain comprises billions of neurons encoding a large diversity of information. This enables humans to navigate multifarious situations but also causes a "needle in the haystack" problem: When encountering a new context, behavioral learning requires identifying the neurons encoding relevant sensory information from within a vast neural population. Surprisingly, humans also achieve remarkably efficient learning from limited experience compared to artificial neural networks. We propose that this conundrum can be explained by the presence of top-down gain modulation in biological neural networks (e.g. via dendrites). Selectively amplifying task-relevant neurons increases the signal-to-noise ratio, thereby enabling rapid credit assignment and accelerating learning. To test this, we compared different algorithms for neuron-specific relevance amplification, including error backpropagation of multiplicative gain parameters, biological models (e.g. Mackintosh), and a novel biologically inspired amplification model based on unsigned reward (|R|) correlation. Our simulations demonstrate that relevance-based amplification substantially improves few-shot reinforcement learning efficacy in a visual two-alternative forced choice task based on a large sensory representation that was computed by a deep convolutional neural network. In conclusion, these findings support the hypothesis that relevance-based amplification of task-relevant neurons through top-down gain modulation constitutes a critical computational principle underlying the exceptional sample efficiency of the brain. By dynamically gating which neurons contribute to learning and behavior, the brain may effectively circumvent the curse of dimensionality inherent to large neural populations. Our findings bridge computational neuroscience and machine learning, suggesting that relevance amplification mechanisms may help close the sample-efficiency gap between biological and artificial intelligence.