From Walruses to Code: A Guide to Tackling Unique ML Problems
summary
Argues that niche vision problems reach usable performance without huge corpora by combining transfer learning with synthetic data generation and lightweight classical models; shows a pipeline that augments a few dozen to a few hundred labeled images (typically 50–200) and fine-tunes a pretrained CNN to cut error by about 20–30% versus naive training; demonstrates active sampling and simple heuristics for label efficiency, builds an end-to-end detection proof of concept on a small walrus dataset that generalizes well to held‑out footage.
AI-written summary · about
chapters (9)
- 0:00Unique ML challenges
- 4:14Biological inspiration
- 18:50Project complexity
- 29:08Environmental factors
- 36:08Visualization and production
- 47:25Model refinement
- 50:17Resolution adjustments
- 1:01:20Theory to practice
- 1:08:07Q&A
related talks
04mUsing Computer Vision to Understand Big Visual Data
41mLarge Scale Representation Learning In-the-wild
30mData-Centric AI Cookbook-let's prep that data
33mRunning Machine Learning in Production - a Journey to Success
28m🧼 From GPU-poor to data-rich: data quality practices for LLM fine-tuning
53mDeep Learning in the world of little ponies
55mCore ML: Vision
01h 18mComputer Vision after the Victory of Data