Challenges in Learning-Based Vision

Learning-based vision comes with a unique set of challenges. This approach to vision derives rules from data, and therefore requires tools for understanding where data can go wrong. This part of the book first presents several failure modes of data-driven methods and then provides tools for mitigating these failures.

Outline

• Chapter 35  Data Bias and Shift introduces the problem of dataset bias and distribution shift. We also encounter the issue of adversarial examples. These problems all come down to a gap between how the model is trained and how it will be used.

• Chapter 36  Training for Robustness and Generality presents one way of dealing with this gap: train on a more diverse distribution of data.

• Chapter 37  Transfer Learning and Adaptation presents a second way of dealing with the gap: adapt your models to bridge the gap.