[Paper Reading] Learning Loss for Active Learning

Paper address: https://arxiv.org/abs/1905.03677
Published in: CVPR'19

Abstract

The performance of deep neural networks improves with more labeled data. The problem is that the budget for annotation is limited. One way to address this problem is active learning, in which the model requires humans to label data that it deems uncertain. Various approaches have been proposed recently to apply active learning to deep networks, but most of them are either designed for their target task or are computationally inefficient for large networks. In this paper, we propose a new active learning method that is simple and task-agnostic, yet works effectively with deep networks. We attach a small parametric module named "loss prediction module" to the target network and learn it to predict the target loss for unlabeled inputs. This module can then alert the target model to samples that are likely to produce incorrect predictions. This approach is task-agnostic because the network learns from a single loss regardless of the target task. We rigorously validate our approach with recent network architectures through image classification, object detection, and human pose estimation. The results show that our method consistently outperforms previous methods in these tasks.

I. Motivation

The idea of ​​this article is very interesting. For active learning, valuable samples are selected. Ideally, the simplest (and most effective) method is to compare the predicted result with the true value. If the model predicts the sample incorrectly, then the sample is naturally valuable; the problem is that there is no labeling. of.

The approach of this article is that since all we want is the difference between the predicted result and the real result, it is good to directly predict the gap itself (loss), which is the "Learning Loss" mentioned in the title:

from a more macro perspective In the past, we had to design various heuristic or non-heuristic algorithms to judge the amount of information of a sample. Now we can directly design a small deep network module alone.

Next, the design of the loss prediction module will be introduced.

II. Loss Prediction Module

From a purely structural point of view, "very simple", the general idea is to extract the features of the middle layer of the model for further processing. Specifically, for the intermediate feature map, global average pooling (GAP) is used to extract its core features, and then it is learned through FC. Finally, the processed intermediate features are concat and the final prediction score is obtained through FC.

The real difficulty is what to use to supervise this loss prediction module.

III. Learning Loss

In this section, we talk about the supervision problem mentioned in the previous section: the

first is "what to use to supervise". In fact, this is also very intuitive, because we have real Loss when we train, so this Loss is the true value. Now to solve the last problem, how to design the "Loss loss function".

The easiest way is to go to MSE. However, as the training progresses, the true value of Loss itself will drop rapidly, causing the gradient to fluctuate violently. At this time, what the model learns is actually the fluctuation of Loss, so it cannot be used directly. Specifically, the loss of this paper is as follows: $$\begin{array}{r}{L}_{\mathrm{l}\mathrm{o}\mathrm{s}\mathrm{s}}\left(\widehat{l^p},l^p\right)=\max \left(0,-\mathbb{A}\left(l_i,l_j\right)\cdot \left({\hat{l}}_i-{\hat{l}}_j\right)+\xi \right)\\ \text{ s.t. }\mathbb{A}\left(l_i,l_j\right)=\{\begin{array}{ll}+1,& \text{ if }{l}_i>l_j\\ -1,& \text{ otherwise }\end{array}\end{array}$$The idea is slightly complicated, and interested readers can read the corresponding part of the original text to understand.