Supervised Generative Reconstruction: An Efficient Way To Flexibly Store and Recognize Patterns

TL;DR

This paper introduces supervised generative models that emulate traditional recognition algorithms while avoiding catastrophic interference, offering a brain-like approach to pattern recognition and storage.

Contribution

It demonstrates how generative models with feedback can prevent forgetting and interference, providing a novel framework for flexible and stable recognition systems.

Findings

Generative models store information as fixed points, reducing interference.

Mathematical analysis shows emulation of feedforward algorithms without catastrophic forgetting.

Brain-like capabilities and limitations are exhibited by the proposed models.

Abstract

Matching animal-like flexibility in recognition and the ability to quickly incorporate new information remains difficult. Limits are yet to be adequately addressed in neural models and recognition algorithms. This work proposes a configuration for recognition that maintains the same function of conventional algorithms but avoids combinatorial problems. Feedforward recognition algorithms such as classical artificial neural networks and machine learning algorithms are known to be subject to catastrophic interference and forgetting. Modifying or learning new information (associations between patterns and labels) causes loss of previously learned information. I demonstrate using mathematical analysis how supervised generative models, with feedforward and feedback connections, can emulate feedforward algorithms yet avoid catastrophic interference and forgetting. Learned information in generative models is stored in a more intuitive form that represents the fixed points or solutions of the network and moreover displays similar difficulties as cognitive phenomena. Brain-like capabilities and limits associated with generative models suggest the brain may perform recognition and store information using a similar approach. Because of the central role of recognition, progress understanding the underlying principles may reveal significant insight on how to better study and integrate with the brain.