Adaptive Tuning Curve Widths Improve Sample Efficient Learning.

Area of Science:

• Computational Neuroscience
• Machine Learning

Background:

• Natural brains excel at learning from few samples, a challenge for machine learning.
• Sample efficiency remains a key open problem in machine learning.

Purpose of the Study:

• Investigate how neural coding schemes impact sample efficiency.
• Propose and analyze a novel, gradient-free learning algorithm for sample-efficient learning.

Main Methods:

• Developed a learning algorithm using a reinforce-type plasticity mechanism without gradients.
• Incorporated bio-plausible mechanisms: population codes with bell-shaped tuning curves, continuous attractor mechanisms, and probabilistic synapses.
• Theoretically analyzed the algorithm and validated through simulations.

Main Results:

• Population codes with broadly tuned neurons enhance sample efficiency; sharply tuned neurons improve precision.
• Dynamic adaptation of tuning width during learning achieves both high sample efficiency and precision.
• The algorithm's sample efficiency guarantee is logarithmically close to the information-theoretic optimum.
• Achieved comparable sample efficiency to gradient-descent-based multi-layer perceptrons in low-dimensional tasks.

Conclusions:

• The proposed gradient-free algorithm offers a promising approach for improving sample efficiency in machine learning.
• Findings suggest adaptive tuning curve width as a potential functional role in neuroscience for sample efficiency.
• The study bridges insights between neuroscience and machine learning for developing more efficient learning systems.