Making Use of All Your Data, Labelled or Otherwise
summary
Unlabeled fruit images are used to train embeddings via contrastive methods (augment an anchor into a positive, sample a negative) using triplet loss with margin alpha and online hard-negative mining; embeddings project into 3D and visibly cluster by true label. A small labelled set (even one-shot/few-shot (m=1)) fine-tunes a lightweight classifier or initializes end-to-end training, cutting annotation time by batch-confirming model-proposed groups. Implementation uses TensorFlow and highlights ML‑Ops needs: fast retrain cycles, embedding compatibility tricks to avoid full redeploys.
AI-written summary · about
chapters (11)
- 0:00Intro
- 3:34Motivation
- 5:20Quick explainer
- 8:24Semi-supervised learning An overview of end-to-end process
- 9:25Representation learning High level concept
- 13:34Triplet loss
- 18:09How do we know if the ML is working? Example: How do we sense check an embeddings?
- 21:54Few shot learning: method 1
- 27:17Practical considerations
- 30:50Iterative improving
- 33:48Summary
related talks
33mBulk Labelling Techniques
30mStarting deep learning projects without enough labelled data
30mVisual data abundant relevant labelled cheap Pick two
41mBetter Machine Learning with Less Data
20mGreenfield vs. Brownfield Data Labeling
41mModels Schm-odels: Why you Should Care about Data-Centric AI
20mPragmatic Deep Learning for image labelling
26mMaximizing Multimodal: Exploring the search frontier of text-to-image models