A computational deep learning approach for establishing long-term declarative episodic memory through one-shot learning

Abstract

Researchers have long been captivated by the intricate workings of the human brain, an enduring enigma. Extensive efforts have been devoted to unraveling its complexities, with disciplines like psychology employing experimentation and analysis to scrutinize and formulate models of brain function. Comprising billions of interconnected neurons, the human brain has inspired experts in deep learning to construct artificial neural networks capable of tasks akin to human brain functions, such as pattern and speech recognition. Despite substantial progress in artificial intelligence, advancements in memory storage capabilities have been relatively constrained. This study aims to investigate mechanisms for simulating long-term declarative episodic memory, reminiscent of human cognition, using one-shot deep-learning neural networks. The proposed deep learning architecture extends to Rosenblatt's C-system memory model, and experiments were conducted to assess the effectiveness of various adaptations of the C-system storage mechanism. The fashion MNIST dataset is used in the experiments, and the results indicate that these models exhibit proficient recall abilities, even when faced with a large number of input images. Furthermore, the study delves into emulating the forgetting process of the human brain. The experiment demonstrates that as the units in the C-system increase, the corresponding results also increase. Specifically, when employing 40,000 units, the system maintains an accuracy exceeding 92% for the sequence of images.