作者: Eric Jonas , Vikash K. Mansinghka
DOI:
关键词:
摘要: The brain interprets ambiguous sensory information faster and more reliably than modern computers, using neurons that are slower less reliable logic gates. But Bayesian inference, which underpins many computational models of perception cognition, appears computationally challenging even given transistor speeds energy budgets. principles structures needed to narrow this gap unknown. Here we show how build fast computing machines intentionally stochastic, digital parts, narrowing efficiency by multiple orders magnitude. We find connecting stochastic components according simple mathematical rules, one can massively parallel, low precision circuits solve inference problems compatible with the Poisson firing statistics cortical neurons. evaluate for depth motion perception, perceptual learning causal reasoning, each performing over 10,000+ latent variables in real time - a 1,000x speed advantage commodity microprocessors. These results suggest new role randomness engineering reverse-engineering intelligent computation.