Neural Reasoning

신경망 추론(Neural Reasoning)은 딥러닝을 범용 인공지능(AGI) 수준으로 향상시키는데 필수적인 추론 능력에 대한 연구를 의미합니다. MADE Lab은 신경망 기반 Multimodal QA에 대해 연구 개발을 진행하고 있고, 특히 Knowledge Graph를 활용하여 LLM의 환각 현상을 최소화하는 연구를 진행하고 있습니다.

Neural reasoning is a research field of answering a question based on neural networks. Due to the great success of LLMs, recent reasoning techniques are mostly based on LLMs for factual decision. Similarly, we focus on how to make LLMs cooperate with various external data like text and large-scale knowledge graphs. Especially, we are devising an effective way of providing information to LLMs an making them generate answers from the external data with high priority. This technique would be essential for reducing hallucination which hinder a model’s ability for getting accurate answers. For reference, we are solving this problem in context of visual question anwering and commonsense-based QA.