Are Associations All You Need to Solve the Dimension Change Card Sort and N-bit Parity Task

When problem-solving, humans can cycle between learned rules to solve tasks. Yet, in artificial neural networks, this cognitive strategy is replaced by learning the entire solution space, making it far less effective. This work aimed to emulate the basis of this human strategy by using a recurrent neural associative memory model. To achieve this, two networks interacted; one served as a task Identifier and the other as a memory Extractor, giving the desired behavior influenced by the Identifier. Each network was trained on sets of interacting associations to represent behavior, such as recognizing shape, color, parity, and when a task started and ended. Once learned, the proposed model was subject to the dimension change card sort (DCCS) and the N-bit parity task. Results showed that the model could switch between behaviors to solve both tasks in linear time once the associations were learned. Moreover, this was possible with 93.3% fewer inputs and no retraining.