Randomization and Selection of Self-Calibrating Binary Encoders in Neural Systems

Abstract Zebrafish larvae are able to track and capture prey five days post-fertilization, consistent with a genetic specification of pattern recognition by the nervous system, yet inconsistent with a process that involves tweaking of individual neural connections by processes similar to those currently employed to train artificial neural nets (ANN). Biological neural networks appear extremely efficient and adaptable when compared to ANN. Here I describe self-calibrating randomly initiated binary encoders (SCRIBEs) that produce sparse output and are highly scalable. The self-calibrating nets (SCN) they generate are capable of classifying and tracking objects with high accuracy and fluency by a process based on selection of SCRIBE arrays optimizing survival. Calibration of SCRIBEs depends on array thresholds that vary with their size but can happen without external input during development and sleep. In neural systems, synchronization likely depends on neuromodulator and brainwave status.