Enhancing Case-Based Reasoning with Neural Networks

Abstract: Neuro-symbolic systems enable leveraging strengths of both neural and symbolic ap- proaches [1]. Case-based reasoning (CBR) is a knowledge-based reasoning and learning methodology that exploits a memory of prior cases--records of prior instances or experiences--by adapting their lessons to solve new problems [2, 3, 4, 5]. The integration of CBR with neural systems is appealing because of their complementary character- istics. Neural models have achieved impressive performance for tasks such as categorization, but reasoning and synthesis of structured solutions (e.g., for planning and design) remain a challenge, as does incorporating existing knowledge into network processes. Training network models--and retraining them for additional data--can be costly. In addition, network models are opaque, functioning as "black boxes." Much research has addressed explanation of black box systems [6] but there are strong arguments for benefits of applying inherently interpretable models, especially as AI is increasingly used in critical domains [7]. Case-based reasoning can be effective with few examples, is amenable to reasoning and synthesis tasks, and supports addition of expert knowledge; it also provides inertia-free learning and is naturally interpretable [8]. However, its success depends on knowledge--not only case knowledge, but also the indexing, similarity and case adaptation knowledge required to retrieve and apply cases--which may be difficult to obtain. This chapter presents research on combining neural network methods with CBR, illustrating how neural system components can reduce the knowledge engineering burden for CBR and improve system performance. It also describes how the use of neural components can benefit CBR system performance by enabling optimization of retrieval and adaptation components in context of each other [9].