Machine and Deep Learning
Existing machine and deep learning models lack allostatic strategies. In addition, current systems offer great performance in narrow environments. This means deep reinforcement learning or evolutionary strategies fail to integrate responses across multiple non-related environments. In fact, there is no viable long-term memory approach for multienvironment integration. Training these systems remains very expensive, even for approaches that mimic natural selection such as genetic algorithms.
Interestingly, biological systems have developed solutions for integrating massive tons of data with genetically encoded responses. These responses can last many generations and can in fact be used in different environments. Until now, we know of many molecular solutions that have been shaped over extensive period of time and offer better evolutionary outcomes. Some of the more impactful adaptations include comparmentalization of specific tasks, development of low entropy models and transcriptional encoding of information.
In line with this, we aim at integrating evolutionary adaptations observed in biological systems with machine and deep learning models. We believe this scientific journey is a step closer to the generation of allostatic learning systems. These systems could be the foundation for novel agents with the ability to adapt in multiple environments by emulating biological adjustments.