Proj CDeepFuzz Paper Reading: Detecting numerical bugs in neural network architectures

Abstract

背景：在architecture level detecting bugs获利更高
本文：DEBAR
Github: https://github.com/ForeverZyh/DEBAR
Task: static analysis of neural architecture to detect bugs based on abstract interpretation
Method: 2 kinds of abstraction(tensors, numerical values), 为了保证精度，tensor partitioning and elementwise affine relation analysis + interval analysis

实验:
dataset:

1. neural architectures with known bugs(The buggy architectures come from two studies: 8 architectures were collected by a previous empirical study on TensorFlow bugs and 1 architecture was obtained from a study conducted to evaluate TensorFuzz.)
2. real-world neural architecture(48 architectures from a large collection of research projects)

效果：

1. outperforms other tensor and numerical abstraction techs on accuracy without losing scalability
2. 在1.7-2.3s中在每个架构中都检测到了全部bugs，无false positive
3. on real-world architecture, +529 warnings(299 true positive, 230 false positive?) with 2.6-135.4 seconds per architecture