Open-Set Learning: Metric Learning, Zero-Shot and Re-Identification

For teams whose models must handle classes, individuals or items they have not seen.

What you will be able to do

Build a Siamese network and explain what it learns.
Explain the difference between closed-set and open-set classification, and evaluate a model on classes it never saw in training.
Classify zero-shot with CLIP, and measure what prompt wording changes.
Train an embedding with metric learning and score the resulting metric space.
Build a retrieval pipeline, such as re-identification: split data honestly, match queries against a database, reject individuals who are not in it, and cluster.

Experience training a neural network, for example from Introduction to Deep Learning.

The worked examples use images; the same ideas apply wherever you can learn an embedding.

Duration: 2 days, on-site or remote, with exercises.

Siamese networks and open sets: learning a similarity, open-set classification, and evaluating on unseen classes
Zero-shot classification with CLIP, and metric learning: training and scoring an embedding
Retrieval and re-identification: honest splits, and matching queries against a database
Retrieval and re-identification, continued: rejecting unknown individuals, and clustering

Tell me about your team and what you would like them to be able to do. A few lines are enough to start; I will reply to set up a scoping call.

Email Yoav yoav@yoavram.com