Principled Transfer Learning for Autonomic Systems: A Neuro-Symbolic Vision

[Context] Transfer learning techniques are commonly employed to address the problem of distribution shifts and concept drifts in learning-enabled autonomic systems. Reusing previous knowledge introduces the risk of invalidating correct behavior while updating the system for a new environment. Recent approaches addressed this via specifying and learning thresholds that allow the detection of a distribution shift that induces a misprediction of an adaptive action. [Problem] These thresholds typically allow only for a low-dimensional form of reasoning. This is not sufficient for making decisions in more complex scenarios of knowledge reuse, where the different partitions of the prediction models present distinct requirements concerning knowledge transfer, e.g., reuse, update, extend, and forget. Without more nuanced reasoning about knowledge transfer, there are increased risks of catastrophic forgetting and negative transfer, i.e., losing valid knowledge after retraining. [Vision] We propose a principled approach to transfer learning, where the learning element is steered by a reasoning engine. We envision a framework that enables different levels of reasoning about the transference of the system knowledge across environments. For that, we endow the reasoning engine with the joint capability of manipulating both learned and rule-based knowledge, in other words, in a neuro-symbolic way.