Bio-inspired problems in rate-distortion theory
Designed and evolved sensors are faced with a difficult optimization problem: storing more information about the environment and reaping greater “rewards” costs time, energy, and material. Quantifying the ben- efits and costs associated with storing environmental information will aid design of more efficient sensors and allow other researchers to test whether or not a particular biosensor is, in fact, efficient. One principled approach to this quantification comes from rate-distortion theory. We discuss three biologically inspired problems in rate-distortion theory. First, we validate a new approach to perceptually compressing natural image patches; sec- ond, we show that sensing tradeoffs in large fluctuating environments depend only on coarse environmental statistics; and third, we describe a new method for calculating predictive features and memory-prediction trade- offs in stochastic environments. For the last of these problems, we suggest a new class of stimuli that might be used to profitably test whether a par- ticular biosensor is an efficient predictor of its environment.