Seminar "Accelerating deep learning with approximate computing".

Summary

Rapid investment in AI has led to a remarkable increase in data, models and infrastructure capacity. Simultaneously, we find ourselves at the dawn of a paradigm shift marked by basic models that can be generalized and adapted on a large scale. However, this self-propelled expansion of AI has resulted in larger, more complex AI models, exascale computing demands and an increased carbon footprint.
To meet these challenges, hardware specialization and acceleration are crucial to satisfy the computational demands of DNNs, which require synergistic design between the different layers of the compute stack. In this talk, we describe a holistic approach to the design of specialized AI systems, pioneered by IBM Research. It highlights a series of techniques for designing and building efficient deep learning systems. This involves approximate computing principles and non-Von-Neumann Analog In-memory approaches to unlock exponential gains in AI computation making AI faster, more efficient and more sustainable.

Biography

Kaoutar El Maghraoui is a senior research scientist at the IBM T.J Watson Research Center and assistant professor of computer science at Columbia University. Kaoutar's work lies at the intersection of systems and artificial intelligence (AI). She leads the AI Testbed at the IBM Research AI Hardware Center, a global research facility that enables efficient next-generation gas pedals and systems for AI workloads. She is responsible for the open-source development and implementation of an AI management system.