Computer Vision after the Victory of Data
summary
Argues that scale and concrete data choices — not clever algorithms alone — explain many breakthroughs: 80 million 32×32 images let simple nearest-neighbor retrieval colorize and geolocate; a jump from 20k to 2 million images turns a hole‑filling demo from failure into plausible riverbanks; face detectors all succeed because someone added negative examples. Shows human visual memory retains thousands of exemplars (10k → ~83% repeat recognition) but texture patterns fail. Demonstrates a pipeline that uses a few hundred hand‑written text edit pairs expanded by GPT and Prompt‑to‑Prompt synthesis to produce supervised image-editing training data, yielding realistic edits with layout and language limitations and a warning about synthetic‑data drift.
AI-written summary · about
chapters (9)
- 0:00The power of simple data
- 25:06Human visual memory capacity
- 33:07Psychology of recognition
- 39:16Auditory and textual memory
- 44:22Modern generative models
- 51:53Image editing with AI
- 58:58Composability in generation
- 1:02:05The future of data science
- 1:15:03Workshop closing remarks
related talks
04mUsing Computer Vision to Understand Big Visual Data
26mComputer Vision in the Era of Deep Learning
40mMachines are Learning
41mDeep Learning to Solve Challenging Problems
03mRethinking software engineering in the AI era
42mState of the Art of Innovation in Computer Vision
40mVisual concept learning from few images
34mNeural networks: insane in the membrane