Making Our Models Robust to Changing Visual Environments

Explore the challenges and solutions for making machine learning models robust to changing visual environments in this conference talk from ROB 2018. Delve into classic domain adaptation techniques, deep domain adaptation methods, and adversarial approaches. Learn about discrepancies between source and target domains, domain adversarial optimization, and the application of GANs for adaptation. Examine real-world examples of adaptation in digit recognition, semantic segmentation, cross-city scenarios, and cross-season environments. Investigate continuous learning strategies, including unsupervised adaptation and replay mechanisms. Analyze experiments with MNIST rotations to understand the balance between adaptation and memory retention. Gain insights into batch and continuous adaptation techniques for improving model performance across diverse visual contexts.

Intro
Benchmark Performance
Dataset Bias
Classic Domain Adaptation
Deep Domain Adaptation
Discrepancy Between Source and Target
Domain Adversarial Optimization
Domain Adversarial Adaptation
Standard GAN Model
CycleGAN for Domain Adaptation
Failures of Image to Image Translation
Adaptation Results: Digit Recognition
Adaptation of Semantic Segmentation
Cross-city Adaptation
Cross Season Adaptation
Cross Season Pixel Adaptation
Synthetic to Real Pixel Adaptation
Summary: Adversarial Domain Adaptation
Continuous Learning
Continuous Unsupervised Adaptation
Experiment: MNIST Rotations
Replay to Remember: MNIST Rotations
Adapt vs Remember: MNIST Rotations
Evaluate MNIST 135 after all rotations
Summary Batch Adaptation
Summary Continuous Adaptation