Nonlinear Bayesian Filtering in Artificial Intelligence

A reliable assessment of dynamic latent characteristics, based on sensory inputs, is one of the signs of perception. This dynamic estimate can be formulated using the nonlinear Bayesian theory of filtration. Recent experimental and behavioral studies showed that animal performance in many tasks is consistent with this Bayesian statistical interpretation. However, it is currently unclear how a non-linear Bayesian filter can be effectively implemented in a network of neurons that satisfies some minimal limitations of biological certainty. Here the author offers the Neural Particle Filter (NPF), a non-linear Bayesian filter based on a sample that is independent of importance weights. The author shows that this filter can be interpreted as the dynamics of neurons in a recurrently connected neural network based on speed receiving a direct signal from sensory neurons. In addition, it covers the properties of temporal and multisensory integration, which are crucial for perception, and allows online study of parameters using the maximum likelihood method.