Neural Network Architecture for Forming Automatic Skills. Part 1

The paper introduces a neural architecture inspired by the human mechanism of habit formation, as described in James Clear's 'Atomic Habits,' where actions become automatic after repetition. The system has four main components: a novelty classifier, a conscious network, a generative network, and a subconscious network, plus an external verifier. Input images of geometric shapes are first classified as novel or familiar. Novel images go to the conscious network, which learns via reinforcement. After a correct answer, the generative network creates variations of the successful example, which are used to fine-tune the subconscious network, making the skill automatic. The experiment involves recognizing circles, squares, triangles, crosses, and stars, with the latter two initially unseen. The subconscious network is pretrained on the first three shapes, and the novelty classifier is trained to consider them familiar. When a cross appears, it is directed to the conscious network, which must classify it correctly, then the generative network produces many synthetic examples to train the subconscious. After training, the novelty classifier is updated so crosses become familiar, and the subconscious handles them directly. The paper outlines success criteria: reduced calls to the conscious network, improved subconscious accuracy on new variants, retention of old skills, and reduced novelty for learned shapes. The architecture is intentionally simple for initial testing, with future plans for more complex tasks and an agent.