Accelerating the Static Analysis of Neural Networks by Batch Representation of(…)
summary
They convert affine forms into three tensors — one for centers, one for noise symbols and one for residuals — so linear, convolutional and many nonlinear abstract-operations become PyTorch tensor ops and run on GPUs or CPU clusters. Concrete scale: images produce >150,000 initial noise symbols, and batching plus data and task parallelism yields about 80% runtime reduction in experiments. The method keeps sound over-approximations, reuses deep learning frameworks, and trades memory layout and symbol merging against precision for scalable verification.
AI-written summary · about
chapters (10)
- 0:00Static analysis intro
- 2:12Abstract interpretation challenges
- 6:33Affine forms and zonotopes
- 12:49Scalability with affine forms
- 15:42Batch representation as tensors
- 18:39Abstract inference algorithm
- 19:22GPU acceleration gains
- 20:33Parallelizing computations
- 21:59Parallel execution results
- 22:41Conclusion and future
related talks
41mUnveiling the simple math that powers AI
43mThe Linear Algebra of Deep Learning
51mTips & Tricks for Fast Neural Net Inference in Production
27mAccelerating Generative AI
27mDemystifying Deep Neural Networks
44mReversing Large Deep Learning AI Models
34mThe physics of deep learning using tensor networks
25mProof transfer for fast certification of multiple approximate neural networks