Machine learning

We aim at the emergence of "intelligence" through a deep understanding of “memory, thought, reasoning, and bias” in machine learning.
We aim to explore the concepts corresponding to these processes in machine learning and to conduct empirical research and construct theories on them.

"Without A Theory The Facts Are Silent"
　　　　　　　　　　　　—Friedrich August von Hayek.

Recent work

Elucidation of Memorization Mechanisms:
We aim to uncover how deep models, such as language models, reconstruct and store information within computers, what constitutes an appropriate internal representation, and how such representations can be effectively learned.

Exploration of Reasoning and Thought:
Our objective is to analyze in detail the mechanisms that enable logical reasoning processes and flexible thinking for complex problems—using techniques such as Chain of Thought (CoT), Inference/Test Time Compute algorithms, and fine tuning via reinforcement learning—to enhance the transparency of decision-making processes in systems like LLMs and to achieve more precise inference models.

Analysis of Bias:
Focusing on the inductive bias that emerges during the learning process, we examine the types of biases present, with the goal of constructing AI systems that embody both fairness and reliability.

Through a deep understanding of memory, thought, reasoning, and bias, we pursue research themes aimed at the emergence of intelligence.

In machine learning, the process of encoding concrete “data” such as images or text into real-valued vectors suitable for computation is indispensable. This encoding forms part of the memorization mechanism that acquires appropriate internal representations from vast amounts of training data and stores them as geometric configurations within a memory space. The stored information can then be flexibly repurposed for prediction or generation of unknown data through reasoning via test time compute, reinforcement learning, and thinking processes such as Chain of Thought. Furthermore, by analyzing in detail the impact of the inductive bias inherent in the learning process on generation and reasoning outcomes, we aim to realize AI systems that embody fairness and reliability. This suite of research endeavors contributes to constructing a theoretical foundation that paves the way for the emergence of intelligence through a deep understanding of memory, thinking, reasoning, and bias.