Congratulations! Our paper titled “Squeezing Accumulators in Binary Neural Networks for Extremely Resource-Constrained Applications,” authored by Azat and Jaewoo has been accepted to the 41st International Conference on Computer-Aided Design (ICCAD 2022), which is held in San Diego, California, in October 30 – November 3.
Unlike the previous papers trying to reduce the multiplication overhead of neural network hardware, this paper asks a different question that is, in binarized neural networks and extremely low-precision quantized neural networks, what is the real bottleneck in hardware implementation? It turns out that accumulators now take a lion’s share in terms of not only area but more power dissipation, and we propose a novel method to minimize accumulator overhead.
Congratulations! Our paper titled “Non-Uniform Step Size Quantization for Post-Training Quantization” authored by Sangyun and Jounghyun as well as our graduate, Hyeonuk, has been accepted to the European Conference on Computer Vision (ECCV) 2022, which is held in Tel Aviv, Israel, in October 23-27.
Unlike the previous papers focusing on better training for quantized neural networks, this paper proposes a radically new concept called subset quantizer, which is based on the idea that by selecting the best subset of quantization levels from a given set of predefined levels, we can increase the representation capability of a quantizer while ensuring the hardware friendliness of arithmetic operations. The concept of the subset quantizer itself was developed by Dr. Hyeonuk Sim together with his advisor, Dr. Jongeun Lee, during the last year of his Ph.D. program.
