Unbiased Gradient Estimation for Event Binning via Functional Backpropagation

Jinze Chen, Wei Zhai, Han Han, Tiankai Ma, Yang Cao, Bin Li, Zheng-Jun Zha · 2026

Read Unbiased Gradient Estimation for Event Binning via Functional Backpropagation by Jinze Chen, Wei Zhai, Han Han, Tiankai Ma, Yang Cao, Bin Li, Zheng-Jun Zha (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-2602-12590.txt.

Description

Event cameras encode dynamic scenes as asynchronous spikes called events, but conventional pipelines need frames, and binning events into frames breaks gradients. Jinze Chen led this work with six co-authors, noting that binning functions are discontinuous, truncating gradients at frame level and forcing most event algorithms to rely solely on frame features. Learning directly from raw events avoids that restriction but suffers biased gradient estimation from binning discontinuities, limiting learning efficiency.

Their frame gives unbiased gradient estimation for arbitrary binning functions by synthesising weak derivatives during backpropagation while keeping forward output unchanged. Integration by parts lifts target functions to functionals, yielding an integral form of the binning derivative where the cotangent function naturally arises. Reconstructing that cotangent from sampled vectors gives weak derivatives matching long-range finite differences for smooth and non-smooth targets. Tests show lower errors and faster convergence on egomotion, optical flow, and SLAM tasks.

Catalog record

CREATORJinze Chen, Wei Zhai, Han Han, Tiankai Ma, Yang Cao, Bin Li, Zheng-Jun Zha
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-12590

Browse the full library · About the Tower of Babel