AI at Stitch Fix
summary
Shows how interpretability comes from concrete choices: replacing point embeddings with distributions so vectors carry uncertainty ellipses, using SVD for word vectors to get deep‑learning quality in minutes, and extracting human‑readable directions with K‑SVD atoms (tank‑top, exposed‑shoulder, jewelry). Demonstrations include variational t‑SNE that reports confidence (large ellipses where the model is unsure) and applied sparse factorization/lda2vec pipelines; code examples use PyTorch/Chainer and scikit‑learn and reportedly produce usable, trustworthy outputs.
AI-written summary · about
chapters (9)
- 0:00Introduction and motivation
- 2:47Simple co-occurrence model
- 8:26Understanding t-SNE
- 11:56KSVD for interpretability
- 14:27Variational methods overview
- 20:47Variational word vectors
- 21:07Variational t-SNE demo
- 23:39Key takeaways and conclusion
- 24:24Q&A
related talks
34mInterpretable AI or How I Learned to Stop Worrying and Trust AI
01h 00mData, visualization, and designing AI
49mlda2vec
34mAugmenting Human Decision Making with Data Science
40mEnabling simple ML ops at Stitch Fix
47mML: Explainable Artificial Intelligence explained
18mSynthesizing human and machine capabilities
39mKeynote 3: How did you know? Explaining Black Box Model Predictions in Python