Debasmita will present Aster, which automates mixed-precision fixed-point quantization for deep feed-forward neural networks.
SPEAKER: Debasmita Lohar, Assist. Prof. at ITU.
ABSTRACT: Neural networks are increasingly becoming integral to safety-critical applications, such as controllers in embedded systems. While safety verification focuses on idealized real-valued networks, practical applications require quantization to finite precision, inevitably introducing roundoff errors. Manually optimizing precision, especially for fixed-point implementation, while ensuring safety, is complex and time-consuming.
In this talk, I will introduce Aster, the sound, fully automated, mixed-precision, fixed-point quantizer for deep feed-forward neural networks. Aster reduces the quantization problem to a mixed-integer linear programming (MILP) problem, thus efficiently determining minimal precision to guarantee predefined error bounds. Our evaluations show that Aster’s optimized code reduces machine cycles when compiled to an FPGA with a commercial HLS compiler compared to (sound) state-of-the-art tools. Furthermore, Aster handles significantly more benchmarks faster, especially for larger networks with thousands of parameters.