Cognitive findings on child learning reveal that children are able to generalize well from few positive examples. This is unlike what is observed in purely data-driven machine learning algorithms. In order to generalize well, machine learners need a relatively large amount of training data. This raises the question of how we can make machine learners generalize better from limited training data. We investigate this problem by taking inspiration from the fields of Cognitive Development and Neuroscience. First, we observe that learning in children is aided by domain-specific inductive constraints. Second, we observe that every part of the human brain contains information about what is stored in the entire brain, that is, human memory is holographic in nature. These two observations motivate us to study the representation gap that exists between the memory of a human and that of a machine learner. We have chosen the framework of Case-Based Reasoning for our study because it is a memory-based reasoning paradigm inspired by the human way of reusing past experiences to solve new problems. Experiences are called cases in the context of Case-Based Reasoning and the collection of experiences possessed by a case-based reasoner is called case base.

We introduce a novel CBR paradigm called Holographic CBR in which we propose changes to the case representation and the case base building process of a conventional CBR system as an attempt to bridge the representation gap between a domain expert and the case base. In the proposed holographic CBR, cases are no longer passive store houses of information but become active problem-solvers in terms of being able to spawn active processes to engage in interaction with a domain expert or to do introspection for acquiring bottom-up knowledge from their neighbouring cases. The proposed design of holographic CBR has been realized on experimental datasets and the results are encouraging. An interesting fallout of holographic CBR is that it also serves as a unifying conceptual ground for many diverse implementations of CBR realized till date.