Title: [017][Robustness and Efficiency][Quantization and Robustness] Error-Silenced Quantization: Bridging Robustness and Compactness Authors: Zhicong Tang , Yinpeng Dong and Hang Su Affiliations: Tsinghua University Venue: IJCAI 2020 Link: https://ceur-ws.org/Vol-2640/paper_8.pdf Code Link: None Summary: Problem: Quantization is found to weaken the robustness. But the adversarial training depends on the larger network capacity which contradicts with quantization. How to design an adversarial method for a quantization network to achieve both robustness and compactness? Key ideas: Silencing the Error Amplification Effect is the key to a robustness-aware quantization. Solutions: Combine the quantization loss and robustness loss, for the quantization, the method just apply a progressive quantization during the training, and for robustness, pairing activation is proposed. Strengths & Weaknesses: Strengths: It's another form of AT... Weaknesses: A model that behaves closely on clean and adversarial inputs is supposed to gain close prediction accuracy on both. Is it correct??? Takeaways&How can I do better: It's not reasonable.