Figure 1.

Inference model and its response properties. A schematic for the generative and inference models. Time-dependent features x_i in the environment are encoded in the activity of feature detector neurons p_i through sensory receptors, s_i. Feature detectors receive input through feedforward (black solid) connections q_ij (PF), which are also used to recurrently predict the stimulus (magenta connections). Feedforward and recurrent activity form the RF of neurons, which is different from the underlying selectivity (PF).