Dual sampling neural network: Learning without explicit optimization

Artificial intelligence using neural networks has achieved remarkable success. However, optimization procedures of the learning algorithms require global and synchronous operations variables, making it difficult to realize neuromorphic hardware, a promising candidate low-cost energy-efficient artificial intelligence. The also fails explain recently observed criticality brain. Cortical neurons show critical power law implying best balance between expressivity robustness code. gives less robust codes without criticality. To solve these two problems simultaneously, we propose model network, dual sampling in which both synapses are commonly represented as probabilistic bit like network can learn external signals explicit stably retain memories while all entities stochastic because seemingly optimized macroscopic behavior emerges from microscopic stochasticity. reproduces various experimental results, including law. Providing conceptual framework for computation by stochasticity optimization, will be fundamental tool developing scalable devices revealing learning.