On Thursday, May 7, 2020 at 10 am Pacific Time, Alliance Member company Hailo will present the free webinar “Quantization of Neural Networks – High Accuracy at Low Precision”. From the event page:
Quantization is a key component of efficient deployment of deep neural networks. 8-bit quantization holds the promise of 4x reduction in model size and an x16 reduction in both compute and power consumption but can result in severe accuracy degradation.
At its heart, quantization is a trade-off between dynamic range and precision. Finding the local optimum for each layer is simple, however, the complex way in which the output of individual layers affect the output of the network is what makes quantization of neural networks tricky. One simple way to address this is by using algorithms which perform greedy global optimization by iteratively applying local optimizations. While conceptually simple, these methods have excellent performance well and are cheap to implement.
In this webinar, we give a brief overview of the principles behind neural network quantization, followed by a review of two techniques recently developed at Hailo: Equalization by inversely proportional factorization (presented at ICML2019) and bias-correction (presented at ECV workshop, CVPR2019). When used in combination, these methods enable fast post-training quantization to 8-bit while achieving state-of-the-art results.
For more information and to register, see the event page.
