Turaco: Complexity-Guided Data Sampling for Training Neural Surrogates of Progr...
summary
Method proposes Turaco, a tiny language plus analysis that computes per-path sample complexity from program source by forming a power series tilde representation, taking absolute coefficients, and using forward-mode AD to evaluate its derivative at 1 to get Zeta. It allocates an overall sample budget across strata proportional to path frequency times complexity, finds daytime path is about 3.7× harder than night, trains separate stratified neural surrogates, and yields about a 15% improvement on a shader case and ~5% average accuracy gain.
AI-written summary · about
chapters (9)
- 0:00Turaco and neural surrogates
- 2:28Program surrogates and applications
- 5:35Sample complexity for program regions
- 8:46Stratified function analysis foundations
- 11:54Turaco language for complexity
- 13:39Sampling ratios evaluation
- 15:43Renderer surrogates results
- 16:23Summary and implications
- 16:46Q&A on loops and constraints
related talks
15mEnabling Tensor Language Model to Assist in Generating High-Performance Tensor...
37mFast-Forward(ing) Models with AI Emulators
01h 02mScheme as a framework for Deep Learning
40mAn Occam's Razor Principle for Transformers?
39mSaaaS - Sampling as an Algorithm Service
01h 04mRandomized Embeddings and Neural Networks
19mGuided Tensor Lifting
18mLearning-Enhanced High-Throughput Pattern Matching Based on Programmable Data Plane