Starting deep learning projects without enough labelled data
summary
Starts models from shortcuts: uses a pre-trained detector class (a clock from COCO) to localize dartboards, trained Faster R-CNN on ~50 labeled images and gets usable scoring; for fruit/peak segmentation runs a generic segmentation model, adds color-space heuristics to cover ~90% of cases and ships a prototype in three weeks. Audio uses mel-spectrograms + a lightweight RNN+CRF on embedding-level synthetic mixtures, adds hard negatives from 40 shows and moves ROC AUC from 0.92 to 0.98 after smoothing. For documents uses BiGRU+CRF, fuzzy matching, HDBSCAN clustering and label propagation to reach ~70% automation.
AI-written summary · about
chapters (7)
- 0:00Start
- 7:07CV segmentation case
- 12:47Audio processing
- 16:34Financial service entity extraction
- 19:31Iterative pipeline for model training
- 20:19Modeling pipeline and inference
- 21:47Summary
related talks
28mScalable High Precision OCRs for Game Event Detections
22mDeep learning module recognizes video sounds
01h 15mTrain Object Detection with Small Datasets
01h 01mA practical guide to deep learning
01h 29mHuman-like Visual Search Application with Small Data
38mDeep Learning for detection on a phone
19mDocument Reader using Deep Learning with a Small Dataset
55mCore ML: Vision