[paper reading] active class incremental learning for balanced datasets

Address of thesis ：https://arxiv.org/abs/2008.10968
Published in ：ECCV 20 Workshop

Abstract

Incremental learning (IL) So that the artificial intelligence system can adapt to streaming data . Most existing algorithms propose two strong assumptions , Reduce the reality of incremental scheme ：(1) It is assumed that new data can be easily marked during streaming ;(2) Test with a balanced data set , Most real-life data sets are unbalanced . These assumptions have been abandoned , The resulting challenges will be solved through the combination of active learning and unbalanced learning . We introduce a sample acquisition function that solves the imbalance problem and is compatible with incremental learning constraints . We also regard incremental learning as an unbalanced learning problem , Rather than the established usage of knowledge extraction for catastrophic forgetting . ad locum , The imbalance effect is reduced by scaling the category prediction in the reasoning process . Four visual data sets were evaluated , The existing and proposed sample acquisition functions are compared . It turns out that , The contribution made has a positive effect , And reduces the gap between active and standard incremental learning performance .

I. Introduction

This paper is the first work that combines quasi incremental learning with active learning . The current class incremental learning has two problems ：1) Data tagging is simple ;2) Data set equalization . In practice , These two requirements are not always met , The task of active learning is to select the most valuable samples , Suitable for use while maintaining performance as much as possible , Reduce the amount of labels and solve the problem of data set imbalance . therefore , Active learning can be combined with incremental learning .

The algorithm flow of this paper is as follows ：

In essence, active learning is added to the class incremental learning method , therefore , The initialization method of the model and class incremental learning , It's all about selecting some classes ( Such as 50% Class ) All samples are marked , Then, on this basis, a fully supervised training is carried out to obtain an initial model ( In the picture $M_0$). after , If you follow the standard class incremental learning process , Is to constantly select some new classes ( Such as 10% Class ) Of all sample , On this basis finetune, And keep the performance of old and new classes as much as possible . however , Since it is active learning , Then it becomes to select some new classes ( Such as 10% Class ) Of part sample .

And as for these part Selection of samples , Using the idea of active learning . For example , It is assumed that the marked budget for this batch of data is B, Then choose... Every time 1/5 B Data and sample (exemplar) Together finetune, Not common in active learning retrain. From this perspective , It can also be considered that class incremental learning improves a classical dilemma in active learning ( Need to repeat retrain).

II. Classical Sample Acquisition Phase

The active learning of this paper adopts a two-stage strategy . In the first phase , Some classical active learning methods are used for initialization . In this paper, coreset、random、entropy、margin sampling These four methods ( Note that there are even random). These methods do not consider the problem of category imbalance , The assumption of this paper is class disequilibrium , So , The second stage was born .

III. Balancing-Driven Sample Acquisition

The second stage is to solve the class imbalance problem . however , The solution is also quite primitive , That is, the classic oversample Strategy ： Which classes have fewer labeled samples , The following sequence will mark more of these classes , This heuristic strategy is called "Poorest Class First".

IV. Experiment

Note that the goal of this article is to compare with fully supervised class increment ( That is, in the penultimate column sIL). However, the experimental results are also difficult to say , And sIL The performance gap is still large , The optimal active learning method is basically random Occupied by , Basically equal to not doing .