Promotionsvortrag Physik: „Unlocking Optoacoustics for Photonic Machine Learning“

Ankündigung des Promotionsvortrags von: Herrn Steven Becker

Artificial intelligence models have not only become more powerful but also more power hungry. One reason for it, is the used von Neumann computing architecture, which separates computing units and memory.
The connection between the two forms the socalled von Neumann bottleneck. Especially for ma-chine learning, this bottleneck creates a huge energy overhead because millions of model parameters must pass it. Photonic machine learning is a new way to solve the von Neumann bottleneck by means of low-loss, fast, and broadband optical signal processing.
The talk demonstrates how stimulated Brillouin scattering (SBS) can accelerate photonic machine learning. Firstly, we show an optoacoustic activation function that is all-optical, trainable, and compatible with resource-efficient multi-frequency operation. Secondly, we experimentally extend the storage time of an optoacoustic memory by one order of magnitude compared to the status quo to 123 ns by using cryogenic temperatures. Finally, we exploit sound waves as a memory to experimentally implement an optoacoustic recurrent operator to realize an information flow in both space and time.
In conclusion, this talk contributes to the research landscape in two ways. Firstly, it extends the methodology and knowledge of multi-frequency SBS. Secondly, it provides a new toolbox for photonic machine learning and shows that the different time scales of light and sound could serve as a new dimension for optical computing.

(Vortrag auf Englisch)

Dem Vortrag schließt sich eine Diskussion von 15 Minuten an. Vortrag und Diskussion sind öffentlich. Diesen Verfahrensteilen folgt ein nicht öffentliches Rigorosum von 45 Minuten.