Two connected ideas. Transfer learning is the practical insight that a network trained on millions of images already knows a lot about edges, textures, and shapes — you can freeze those layers and just retrain the final classifier on your own data in minutes. Time series is the first look at data where order matters, setting up the sequence models that follow.

Concepts You’ll Learn About

  • Transfer learning — pre-trained weights as a feature extractor; frozen vs. fine-tuned layers
  • Fine-tuning — unfreezing the top layers and retraining at a low learning rate to adapt to a new domain
  • Audio as an image — spectrograms; applying image CNNs to sound
  • Time series — sequences where the past predicts the future; sliding windows; stationarity
  • Recurrent processing (preview) — why feed-forward networks struggle with sequences; motivating the RNNs of Unit 12

Topics

  • Bird call classification — audio spectrograms treated as images; transfer learning applied to an unusual domain (a “snow day” notebook with a fun origin story).

  • Fine-tuning ResNet / MobileNet — freeze all layers except the final classifier, retrain on a new task; measure what fine-tuning gains over pure feature extraction.

  • Collect-your-own-images project — photograph objects around the school, label them, train a transfer-learning classifier on your own data. Assignment: submit your trained model and a reflection on what worked.

  • Guest speaker.

  • AAPL stock price prediction — a time-series regression: predict tomorrow’s close from a window of recent prices (AAPL.csv). A deliberate setup for “what would an RNN do better?” (reference prediction)

What’s next

Unit 12 answers the time-series question properly with recurrent neural networks — and takes them all the way to neural machine translation with attention.