25 January 2023 in IEEE Access
JTJayaraman J. Thiagarajan
Unsupervised domain adaptation (UDA) aims to transfer and adapt knowledge from a labeled source domain to an unlabeled target domain. Traditionally, geometry-based alignment methods, e.g., Orthogonal Procrustes Alignment (OPA), formed an important class of solutions to this problem. Despite their mathematical tractability, they rarely produce effective adaptation performance with the recent benchmarks. Instead, state-of-the-art approaches rely on sophisticated distribution alignment strategies such as adversarial training. In this paper, we show that, conventional OPA, when coupled with powerful deep feature extractors and a novel bi-level optimization formulation, is indeed an effective choice for handling challenging distribution shifts. When compared to existing UDA methods, our approach offers the following benefits: computational efficiency : Through the isolation of alignment and classifier training steps during adaptation, and the use of deep OPA, our approach is computationally very effective (typically requiring only 700 K parameters more than the base feature extractor as compared to millions of extra parameters required by state-of-the-art UDA baselines); (ii) data efficiency : Our approach does not require updating our feature extractor during adaptation and hence can be effective even with limited target data; (iii) improved generalization : The resulting models are intrinsically well-regularized and demonstrate effective generalization even in the challenging partial DA setting, i.e ., target domain contains only a subset of the classes observed in the source domain.; and (iv) incremental training : Our approach allows progressive adaptation of models to novel domains (unseen during training) without requiring retraining of the model from scratch.
30 September 2020 in arXiv
JTJayaraman J. Thiagarajan
The hypothesis that sub-network initializations (lottery) exist within the initializations of over-parameterized networks, which when trained in isolation produce highly generalizable models, has led to crucial insights into network initialization and has enabled efficient inferencing. Supervised models with uncalibrated confidences tend to be overconfident even when making wrong prediction. In this paper, for the first time, we study how explicit confidence calibration in the over-parameterized network impacts the quality of the resulting lottery tickets. More specifically, we incorporate a suite of calibration strategies, ranging from mixup regularization, variance-weighted confidence calibration to the newly proposed likelihood-based calibration and normalized bin assignment strategies. Furthermore, we explore different combinations of architectures and datasets, and make a number of key findings about the role of confidence calibration. Our empirical studies reveal that including calibration mechanisms consistently lead to more effective lottery tickets, in terms of accuracy as well as empirical calibration metrics, even when retrained using data with challenging distribution shifts with respect to the source dataset.