Heterogeneous and Targeted Systems Seminar: Vladimir Lončar

Monday 14 Feb 2022, 12:00 → 13:00 US/Pacific

Javier Mauricio Duarte (Univ. of California San Diego (US))

Title: The hls4ml project and the future of deploying ultrafast deep learning on specialized hardware

Abstract: The hls4ml project aims to simplify deployment of deep neural networks on FPGAs within applications characterized by tight constraints in terms of latency, hardware resources and power consumption. Since its introduction, the library has grown significantly, integrating support for FPGAs from different vendors, multiple network architectures (convolutional, recurrent, graph), extreme quantization (binary and ternary networks), and multiple applications (classification, regression, anomaly detection). Thanks to a collaboration with Google, it was interfaced to QKeras to enhance network compression capabilities through quantization aware training. The vibrant and fast-expanding community of users and developers is working to expand hls4ml and apply it to new domains. In this talk we review the state of the project from it's beginnings to present day and show potential new directions it may take in the future.

Biography: Vladimir Lončar is a Senior applied fellow at CERN working on deep learning systems. He is the lead developer of the open-source hls4ml library, a compiler to build machine learning models in FPGAs. Lončar also develops a distributed deep learning and hyperparameter optimization framework (nnlo), which simplifies training and optimization tasks on multi-node systems (local clusters, HPC sites, and cloud resources). He received his Ph.D. in Computer Science in 2017 from the University of Novi Sad. 

12:00 → 12:40 The hls4ml project and the future of deploying ultrafast deep learning on specialized hardware

Speaker: Vladimir Loncar (CERN)

GMT20220214-200507_Recording_1794x934.mp4

hls4ml_future_a3d3.pdf