Iclr2022: how does AI recognize "things I haven't seen"?

Line early From the Aofei temple
qubits | official account QbitAI

This time Foreign object detection Direction out of a new model VOS, The team is from the University of Wisconsin Madison , The paper has been included in ICLR 2022 in .

This model achieves the best performance in target detection and image classification ,FPR95 The index is lower than the previous best effect 7.87% As much as .

It is always a difficult problem for deep networks to deal with unknown situations .

For example, in autonomous driving , Identify known objects （ Like cars 、 Stop sign ） Our detection model often “ deliberately misrepresent ”, For extraterritorial objects （OOD） Will produce a high confidence prediction .

Like a moose in the picture below , stay Faster-RCNN It is recognized as a pedestrian under the model , also 89% The degree of confidence .

Therefore, the detection of extraterritorial objects undoubtedly becomes AI Safety is a very important topic .

Let's take a look at how this model judges extraterritorial objects .

VOS How to detect foreign objects

Understanding VOS Before , I have to mention the reason why it is difficult to detect foreign objects .

It's easy to understand , After all, neural networks are just data for learning, training and testing , I don't know when I meet something I haven't seen .

To solve this problem , We have to find a way to let the network know “ Unknown ” Things of . What should I do ？

VOS The idea is , Simulate an extraterritorial object for the model to learn .

For example, the detection in the figure below , Three of them are our goals . When there is no simulation of extraterritorial objects （ Left ）, The model can only hold the target in a large range .

After training with simulated extraterritorial objects （ Right ）, The model can lock the target compactly and accurately , Form a more reasonable decision-making boundary .

And once the target is locked, it is more accurate , As long as it is outside this range , Other objects can be judged as extraterritorial objects .

Based on this idea ,VOS Our team built such a framework ：

With a Faster-RCNN Based on the Internet , Add some data simulating objects outside the domain to the classification header , Put it together with the data in the training set , Jointly build a standardized uncertainty loss function .

And where does the data of these simulated extraterritorial objects come from ？ It can be seen in the structure diagram , These points come from the target area （ Blue dots 、 Yellow square dots and green triangle dots ） Around , That is, the low likelihood region .

Finally, according to the calculation of confidence , Blue represents the target detection data , Green represents extraterritorial objects .

To judge the car and moose in the image .

Compare it with many other foreign object detection methods , We can see that VOS The advantages of .

The arrow down in each indicator indicates that the smaller the data, the better , On the contrary, it means that the larger the item, the better .

among FPR95 This is the most prominent , Describe the OOD The accuracy of sample classification is 95% when ,OOD The sample was wrongly assigned to ID Probability in the sample .

This result is lower than the previous best result 7.87%.

Compared with other existing methods , It also shows VOS The advantages of .

It serves as a general learning framework , It can be applied to target detection and image classification . The previous methods are mainly driven by image classification .

At present, the model has been used in GitHub The open source .

Author's brief introduction

This model is mainly composed of Du Xuefeng 、 Cai Mu and others proposed .

Du Xuefeng graduated from Xi'an Jiaotong University , Currently studying at the University of Wisconsin Madison CS Doctor .

The main research direction is trusted machine learning , Including extraterritorial object detection 、 Against robustness 、 Noise label learning .

Cai Mu , He also graduated from Xi'an Jiaotong University , At present, it is the University of Wisconsin Madison CS Sophomores .

Research interests focus on deep learning 、 Computer vision , Especially three-dimensional scene understanding （ Point cloud detection ） And self supervised learning .

The corresponding author of this paper is Sharon Yixuan Li, At present, he is an assistant professor of computer science at the University of Wisconsin Madison , I've been in Facebook AI Professor Ren .

Reference link ：

[1]https://twitter.com/martin_gorner/status/1489671903727915008
[2]https://arxiv.org/abs/2202.01197
[3]https://sites.google.com/view/mucai
[4]https://www.linkedin.com/in/xuefeng-du-094723192/details/experience/
[5]https://github.com/deeplearning-wisc/vos