Production of Neural Receptive Field Due to Unsupervised Learning in Images and Audio

Abstract

The efficient computing required for sensory processing is fueled by a blend of high level task dependent learning and lowlevel unsupervised statistical structural learning. Sparse and independent coding techniques can simulate brain functioning at the earliest stages of sensory processing utilising the identical coding mettthod with just a modification in input. The authors offer a comprehensive discussion on Autonomous Component Analysis (ACA), a neural coding mechanism that is effective in simulating fast auditory and visual neural processing. Using a standardised five phase process, we developed an auto included, approachable Jupyter notebook in python to show how to efficiently code for various modalities. The comparison of derived receptive field models for each modality shows how neural codes do not form when inputs adequetly differ from those that organisms were adapted to exercise. The presentation also demonstrates that ACA generates receptive field models that are more neurally appropriate than those based on conventional squeezing techniques, such as Chief Component Analysis (CCA). The five phase approach not only creates models that resemble neurons, but also encourages code reprocess to highlight the input sceptic feature of the approach, which enables every modality to be modelled with a single modification in inputs. This notebook makes it simple to see the connections between unsupervised machine learning techniques and fast sensory neuroscience, which advances our knowledge of how adaptable data driven neural networks form and their potential uses in the future.