We propose a temporal reweighting approach for training models under slow concept drift. A meta-model scores each instance, and its age, according to the value it provides for future predictions. We outperform a range of other robust reweighting schemes by upto 8% relative, on a longitudinal dataset (9 years), and on a range of other nonstationary learning benchmarks. To our knowledge, this is the first proposal to leverage instance characteristics and data age for forward transfer.

Instance-conditional timescales of decay for nonstationary learning

N. Jain, P. Shenoy. AAAI 2024.

We propose a feature sieve--a novel method for automatically mediating between potential predictive features in a deep network based on their generalization capability.  Our method identifies and suppresses features with spurious label correlations, without access to definitions or other characterizations of potential features. We report significant gains (upto 11% relative) on real-world datasets with spurious feature-label correlations such as BAR, NICO, CelebA, Imagenet-9/Imagenet-A.

Overcoming simplicity bias in deep networks using a feature sieve

R. Tiwari, P. Shenoy. ICML 2023.

We improve accuracy and across-group fairness of student models in distillation. We show that early readouts (linear decoding from earlier layers of the network) indicate featural bias through overconfident errors on underrepresented instances. By reweighting teacher loss as a function of early-layer error confidence, we show gains not only in worst-group accuracy but also overall accuracy over other distillation approaches on fairness benchmark datasets.

Using early readouts to mediate featural bias in distillation.

R. Tiwari, D. Sivasubramanian, A. Reddy, G. Ramakrishnan, P. Shenoy. WACV 2024.

We propose a novel , cost-optimized approach for human-AI cooperative prediction. Our model refines its predictions by iteratively querying the human for supporting information at test time, while minimizing overall querying cost.  Our learned querying policies improve accuracy rapidly, with fewer human inputs at test time (top) compared to baselines.

Interactive concept bottleneck models

K. Chauhan, R. Tiwari, J. Freyberg, P. Shenoy, K. Dvijotham. AAAI 2023.

We propose simple inductive biases for multi-object multi-part segmentation: a) separate predictions for objects & parts,  pooling part data across categories, b) separate predictions for parts & attributes, generalizing attributes to whole objects. Both inductive biases increase available training data, improve generalization, and reduce the output label space significantly. Our approach shows 5%-20% relative gains over SOTA on the PASCAL-Parts dataset. 

FLOAT: Factorized learning of object attributes for improved multi-object multi-part scene parsing

R. Singh, P. Gupta, P. Shenoy, R. Sarvadevabhatla. CVPR 2022.