Feedback-prop: Convolutional Neural Network Inference under Partial Evidence

TL;DR

This paper introduces feedback-prop, a novel inference method for CNNs that leverages partial evidence to improve prediction accuracy without retraining, applicable across various visual recognition tasks.

Contribution

The paper presents a general feedback-based inference procedure for CNNs that enhances predictions using partial evidence without retraining or fine-tuning.

Findings

Feedback-prop improves accuracy in multi-task CNNs with partial evidence.

The method works across different CNN architectures and tasks.

It reveals a new dynamic property of deep CNNs.

Abstract

We propose an inference procedure for deep convolutional neural networks (CNNs) when partial evidence is available. Our method consists of a general feedback-based propagation approach (feedback-prop) that boosts the prediction accuracy for an arbitrary set of unknown target labels when the values for a non-overlapping arbitrary set of target labels are known. We show that existing models trained in a multi-label or multi-task setting can readily take advantage of feedback-prop without any retraining or fine-tuning. Our feedback-prop inference procedure is general, simple, reliable, and works on different challenging visual recognition tasks. We present two variants of feedback-prop based on layer-wise and residual iterative updates. We experiment using several multi-task models and show that feedback-prop is effective in all of them. Our results unveil a previously unreported but interesting dynamic property of deep CNNs. We also present an associated technical approach that takes advantage of this property for inference under partial evidence in general visual recognition tasks.