Quant-Noisier : Second-Order Quantization Noise

Memory and compute constraints associated with deploying AI models on the "edge" have motivated development of compression methods designed to reduce larger models into compact forms. We aim to improve upon a state-of-the-art method developed by Facebook AI Research, Quant-Noise [1]; our novel method, Quant-Noisier, utilizes second-order noise to improve performance on compressed models. For all three datasets we test, our best Quant-Noisier variant, random jitter, outperforms Quant-Noise on two out of three quantization schemes. Key Information: Custom project mentored by Angelica Sun; Lyron is sharing with CS224S.