Lecture Slides

Lecture Notes

Zero-shot Learning

Zero-shot learning aims to bridge the gap between visual features and semantic attributes. By leveraging neural networks, we can learn effective mappings between these different types of information, allowing models to recognize objects they've never seen during training based on their described attributes.

Bridging visual and semantic information

Zero-shot Learning

Generalized zero-shot learning

Different Datasets

Revisiting DeViSE

ConSE

Some Challenges

1. Hubness

   

   The hubness problem refers to a phenomenon in high-dimensional spaces where certain data points (called "hubs") become unusually close to many other points. This creates challenges for zero-shot learning because:

   • It can lead to biased predictions where the same classes are repeatedly predicted
   • The semantic embedding space may become distorted, reducing model accuracy
   • Distance metrics become less reliable in high dimensions

2. Bias towards seen classes

   Model tend to favor predicting classes they were trained on (seen classes) over novel classes they haven't encountered before.

   

   Some Network Architectures