[Succinct Information Processing Team Seminar] Geometry of Brain Representation Space and AI Latent Space: Factors Determining the Performance of Brain–AI Integration

Format: Hybrid

Online venue: Open to all registered participants. The seminar will be streamed via Zoom. The URL will be visible only to registered participants.

Offline venue: RIKEN members only. RIKEN AIP Nihonbashi Office Open Space

Talk Information

Speaker: Takufumi Yanagisawa (Professor, Department of Neuroscience, Graduate School of Medicine, Osaka University)

Title: Geometry of Brain Representation Space and AI Latent Space: Factors Determining the Performance of Brain–AI Integration

Abstract :

In recent years, advances in deep learning have rapidly accelerated research on decoding perceptual, motor, and linguistic information from brain activity with high accuracy. In particular, by embedding diverse external information such as images and language into a unified AI latent representation space, zero-shot brain decoding—which enables the decoding of previously unseen stimuli—has begun to emerge.

Furthermore, by using a closed-loop Brain–Machine Interface (BMI), it has become possible to map visual images imagined by humans into representations in an AI latent space estimated from brain activity, and then retrieve and present corresponding images from large-scale databases. Such BMIs represent an example of Brain–AI integration, where neural information and AI representations are utilized in an integrated manner.

At the same time, it has become clear that the performance of Brain–AI integration depends on the properties of both the AI latent space and the brain’s internal representation space. In particular, their representation geometry and information structure strongly influence the accuracy and generalization ability of brain decoding. We have demonstrated that the representation structure of AI models affects whether brain decoding trained on perceived images can generalize to visual imagery.

In this talk, we focus on the relationship between brain representation space and AI latent space, and discuss how their representation structures influence the performance of Brain–AI integration. Furthermore, by combining these technologies with implantable devices, medical applications such as communication support and motor function restoration for patients with severe motor disabilities are expected. We will introduce our efforts toward realizing clinical applications of BMI by integrating technologies including device development, brain decoding, output generation, and external device control.