SAFE-KD: Risk-Controlled Early-Exit Distillation for Vision Backbones

Salim Khazem · 2026

Read SAFE-KD: Risk-Controlled Early-Exit Distillation for Vision Backbones by Salim Khazem (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-03043.txt.

Description

Salim Khazem's solo 2026 paper makes early-exit vision networks safer to deploy. Early-exit networks let easy inputs stop at shallow layers, saving computation. The practical question is knowing when stopping early is safe, since a wrong confident exit wastes the savings and harms accuracy on exactly the inputs the system claims are easy.

His wrapper, SAFE-KD, attaches lightweight exit heads at intermediate depths of a backbone. A strong teacher is distilled into every exit, with consistency enforced from deep to shallow exits. At inference, per-exit thresholds are calibrated on held-out data with conformal risk control. That guarantees a user-chosen misclassification risk among early exits, under standard assumptions. Tests across datasets and architectures showed better accuracy-compute trade-offs and calibration.

Catalog record

CREATORSalim Khazem
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE27 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-03043

Browse the full library · About the Tower of Babel