Effective receptive field in DiTs mirrors 2017 CNN finding, CS professor notes
CSProfKGD · x · 2026-10-22
CS professor CSProfKGD (88k followers) highlights new research showing that the effective receptive field in diffusion transformers (DiTs) is significantly smaller than the theoretical one during later denoising stages — echoing the 2017 paper "Understanding the Effective Receptive Field in Deep Convolutional Neural Networks" (Luo, Li, Urtasun, Zemel, NIPS 2016). That paper introduced the effective receptive field concept, showed it is Gaussian-distributed and occupies only a fraction of the theoretical field, and analyzed how activations, dropout, subsampling, and skip connections affect it — suggesting the gap between theoretical and effective receptive field is a cross-architecture phenomenon.
Related event: Study Finds DiT Effective Receptive Field Far Smaller Than Theoretical(3 posts)→
More from Research
- Researcher Steers Model Attention Heads Toward Honesty, Flips Nihilism Traits — voooooogel · 2026-10-22
- We need an L0.5 for continual learning: agent state stacks up in 4 layers — willcb · 2026-10-22
- JOLT trains the teacher too, with penalties when its predictions diverge from the student — perplexity_ai · 2026-10-22
- Perplexity's JOLT matches GRPO with 6.5x fewer training tokens, JOLT+ wins on all five benchmarks — perplexity_ai · 2026-10-22
- Parallel Decoding Distillation Adopted by Alibaba, NVIDIA, Amazon Weeks After Release — ArashVahdat · 2026-10-22
- RFD4-Proteina hits 3.5% DNA-binder pass rate vs 0.8% for BoltzGen — AllThingsApx · 2026-10-22