An introduction to out-of-distribution detection for AI safety, covering anomaly, novelty, open-set and drift detection along with classification, density and distance methods.
Paper review of GAIA, a gradient-based attribution framework that detects out-of-distribution samples through channel-wise and zero-deflation abnormality scores.
Our BACS paper tackles background ambiguity in continual semantic segmentation with a transformer decoder, a Mahalanobis-based shift detector and masked knowledge distillation.
A survey of uncertainty estimation in deep learning covering aleatoric vs epistemic uncertainty, Bayesian inference, Fisher information and Gaussian processes.
Mentoring high-school students in Ghana through the MISE program on machine learning fundamentals and OOD detection research, with one mentee admitted to MIT.
Our 6th-place CVPR 2021 continual learning challenge submission using Dark Experience Replay with a 6000-sample memory buffer to fight catastrophic forgetting.
Training a Duckiebot for lane following and obstacle avoidance with DAgger imitation learning, transferring a pure-pursuit expert policy from simulation to real hardware.
Paper review showing the generalization gap of a CNN can be predicted from layer-wise weight statistics alone, even across architectures and datasets via domain shift.
Paper review on predicting the deep network generalization gap from margin-distribution statistics across hidden layers, fed into a simple linear estimator.
