HOME News & Events Forthcoming events Friday Lunch Seminar: Yuji Kawai: "Oscillations as the key to learn time series: A computational approach from simple timing to complex rhythms" (On-line & In-person : Sign-up required)

Friday Lunch Seminar: Yuji Kawai: "Oscillations as the key to learn time series: A computational approach from simple timing to complex rhythms" (On-line & In-person : Sign-up required)

Friday Lunch Seminar (English)

September 27, 2024
12:15 〜 13:00 (JST)

Apply for participation
on-line:
or you can come to the conference room on the 1st floor of the CiNet bldg.

Sign up by noon on September 26.
When we cannot identify your affiliation etc., we may have to turn down your application.
You will be notified of participation details by e-mail on September 26.

Talk Title: Oscillations as the key to learn time series: A computational approach from simple timing to complex rhythms

Yuji Kawai
Associate Professor
Symbiotic Intelligent Systems Research Center, Institute for Open and Transdisciplinary Research Initiatives
Osaka University

Host PI: Minoru Asada

Abstract:
Understanding how the brain learns, generates, and generalizes time series, ranging from simple timing to complex rhythms, is a fundamental question in neuroscience. This talk explores the computational mechanisms underlying these processes, with a focus on how reservoir computing, a type of artificial recurrent neural network, replicates and generalizes long-term temporal patterns. The perception and generation of timing and rhythms involve some brain areas including the basal ganglia and cerebellum. We propose oscillation-driven reservoir computing (ODRC) as a principal computational model for these areas, where oscillatory signals are fed into a random recurrent neural network to stabilize network activity and induce complex neural dynamics. These stable and complex dynamics enable the ODRC to learn long-term motor timing. The ODRC not only replicates target time series but also generalizes them. For example, when the ODRC learns chaotic Lorenz time series for a specific period, it can replicate the series during that period and generate similar time series afterward. This capability of the ODRC was applied to the learning of complex rhythms. Professional drumming performances were encoded into time series and learned by the ODRC. The results showed that the ODRC not only reproduced the performances but also generated similar performances, potentially including improvisations.