Published in The International Conference on Learning Representations (ICLR), CSS Workshop, 2022
This paper explores how and when backgrounds help a model with object classification task
Recommended citation: Raza Hashmi, Qiang Li. (2022). "Object Recognition With Help From Background." ICLR CSS Workshop.(1) https://sites.google.com/view/cssiclr2022/home
These are active drafts—not yet submitted. I’m happy to share early write-ups or discuss collaborations: razahashmi93@gmail.com.
Pathway-Enhanced Uncertainty Estimation. Estimating predictive uncertainty from the consistency of a model’s internal pathways and prototype activations, rather than from output probabilities alone.
Diffusion-based Uncertainty. A sampling-driven approach to uncertainty estimation, with theoretical checks and comparisons against established baselines.
Difficulty Is Relative. A unified suite to measure and predict class-pair difficulty—treating “hardness” as a relationship between classes rather than a fixed per-example property.
Alongside these, I’m building evaluation tooling for distribution shift, calibration, and rejection/coverage tradeoffs.