At Neuroloops, we study perceptual learning as we observe, model, and control the brain and behavior. Our earlier work exclusively focused on animals to shed a mechanistic light on how sensory information is translated into action and how experience alters this transformation and behavior throughout the lifespan. We have identified fundamental mechanisms through which egocentric (self-centered) and allocentric (world-centered) representations of the sensory world in the brain are formed, stored and modified in an experience-dependent manner (e.g Nature Neuro. 6:291, Nature Neuro. 7:534, Science 319:101, Cell Stem Cell 5:178, Cell Stem Cell 12:204, PLOS Comput. Biol. 11:31004386, PLOS Comput. Biol. 12: e1004984). To extend this research towards controlling behavior in freely behaving animals, we have developed new hardware, software and computational methods (e.g. J. Neurophy. 100:504, J. Neurophy. 109:3094, J. Neurophy. 110:620, J. Neural Eng. 16:065001), and training paradigms that allow precise quantification of sensory input to the brain and behavior in millisecond resolution (e.g. Science 319:101, PNAS 104:1395).
Using some of these methods, we have established quantitative behavioral, neural, and computational read-outs of perceptual learning, which we are translating into control architectures for robotic agents (Science Robotics 7:67). Our experiments have shown that within 90 ms (in rodents) to 230 ms (in humans), the brain collects the somatosensory information from the periphery, creates a percept of the stimulus, builds a tactile memory trace and the associated expectations (i.e. the prior in the Bayesian context), generates a motor plan (i.e. posterior) and executes it while keeping track of the “error” in the sensory and motor computation given its previous experiences. We are now developing methods to control every step of perceptual learning in rodents and humans. This unique bridging of animal and human experiments will help shed a mechanistic light on perceptual learning.