# Perceptual Inductive Bias Is What You Need Before Contrastive Learning

Source: [Perceptual Inductive Bias Is What You Need Before Contrastive Learning](<https://www.sota2.com/research/paper/68fa79001e209131cea7b942>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-04-15T05:00:11.244Z

Authors: Tianqin Li, Junru Zhao, Dunhan Jiang, Shenghao Wu, Alan Ramirez, Tai Sing Lee

Published: 2025-06-01T22:32:37.000Z

## Abstract

David Marr's seminal theory of human perception stipulates that visual processing is a multi-stage process, prioritizing the derivation of boundary and surface properties before forming semantic object representations. In contrast, contrastive representation learning frameworks typically bypass this explicit multi-stage approach, defining their objective as the direct learning of a semantic representation space for objects. While effective in general contexts, this approach sacrifices the inductive biases of vision, leading to slower convergence speed and learning shortcut resulting in texture bias. In this work, we demonstrate that leveraging Marr's multi-stage theory-by first constructing boundary and surface-level representations using perceptual constructs from early visual processing stages and subsequently training for object semantics-leads to 2x faster convergence on ResNet18, improved final representations on semantic segmentation, depth estimation, and object recognition, and enhanced robustness and out-of-distribution capability. Together, we propose a pretraining stage before the general contrastive representation pretraining to further enhance the final representation quality and reduce the overall convergence time via inductive bias from human vision systems.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Depth Estimation on NYU Depth v2](<https://www.sota2.com/research/sota/depth-estimation-on-nyu-depth-v2.md>): Depth Estimation; NYU Depth V2; RMSE: 0.1354.

- [Image Classification on ImageNet-100 (val)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-100-val.md>): Image Classification; ImageNet-100 (val); Top-1 Accuracy: 78.

- [Semantic Segmentation on Cityscapes (Semseg mIoU)](<https://www.sota2.com/research/sota/semantic-segmentation-on-cityscapes-semseg-miou.md>): Semantic segmentation; Cityscapes; mIoU: 68.3.

- [Semantic Segmentation on ADE20K (mIoU)](<https://www.sota2.com/research/sota/semantic-segmentation-on-ade20k-miou.md>): Semantic segmentation; ADE20K; mIoU: 32.1.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2506.01201>)

[PDF](<https://arxiv.org/pdf/2506.01201.pdf>)
