A Brain-inspired Memory Transformation based Differentiable Neural Computer for Reasoning-based Question Answering

Yao Liang^1,2*Yuwei Wang^1,3Hongjian Fang^1,4Feifei Zhao¹Yi Zeng^1,2,3,4,5

• ¹Brain-inspired Cognitive Intelligence Lab, Institute of Automation, Chinese Academy of Sciences, Beijing, China
• ²School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China
• ³Center for Long-term Artificial Intelligence, Beijing, China
• ⁴School of Future Technology, University of Chinese Academy of Sciences, Beijing, China
• ⁵Key Laboratory of Brain Cognition and Brain-inspired Intelligence Technology, Chinese Academy of Sciences, Shanghai, China

Reasoning and question answering, as fundamental cognitive functions in humans, remain significant hurdles for artificial intelligence. While large language models (LLMs) have achieved notable success, integrating explicit memory with structured reasoning capabilities remains a persistent difficulty. The Differentiable Neural Computer (DNC) model, despite addressing these issues to some extent, still faces challenges such as algorithmic complexity, slow convergence, and limited robustness. Inspired by the brain's learning and memory mechanisms, this paper proposes a Memory Transformation based Differentiable Neural Computer (MT-DNC) model. The MT-DNC integrates working memory-a cognitive system temporarily holding and processing information relevant to immediate tasks-and long-term memory, which stores frequently accessed and enduring information, within the DNC framework, enabling the autonomous transformation of acquired experiences between these memory systems. This facilitates efficient knowledge extraction and enhances reasoning capabilities. Experimental results on the bAbI question answering task demonstrate that the proposed method outperforms existing Deep Neural Network (DNN) and DNC models, achieving faster convergence and superior performance.Ablation studies further confirm that the transformation of memory from working memory to long-term memory is critical for improving the robustness and stability of reasoning. This work offers new insights into incorporating brain-inspired memory mechanisms into dialogue and reasoning systems.