Transfer Learning: Leveraging Pretrained Models with Limited Data
summary
Claims that transfer learning finds laughter in very small labeled sets using pretrained audio encoders and simple classifiers; dataset is ~5s clips with ~1,500 training and ~400 test samples and severe class imbalance. Off‑the‑shelf YAMNet yields precision under 15% and recall around 30%, a ResNet laughter model improves but remains weak, while taking Whisper encoder embeddings plus a small feed‑forward head outperforms both. Work uses feature extraction, plans adapter/finetune steps to scale.
AI-written summary · about
chapters (7)
- 0:00Transfer learning project intro
- 1:07What is transfer learning
- 3:13Pre-trained model types
- 4:37Core transfer strategies
- 7:25Laughter detection case study
- 11:59Key takeaways
- 12:41Q&A
related talks
44mTransfer learning in NLP
41mBetter Machine Learning with Less Data
45mTransfer Learning - Entering a new era in NLP
32mAn Introduction to Transfer Learning in NLP and HuggingFace
43mThe Evolution of Domain Transfer in the Era of Pre-trained Language Models
35mTricks Learned from Scaling WhisperSpeech Models to 80k+ Hours of Speech
30mAI on a Microbudget - Methods of Machine Learning Miniaturization
33mWhat if a Robot Could Laugh at Your Jokes? (Laughter Prediction with Natural Language Processing)