Titulo Estágio
Open-World Active Learning in Self-Driving Cars
Áreas de especialidade
Sistemas Inteligentes
Local do Estágio
DEI-FCTUC
Enquadramento
Closed World Artificial Intelligence [4, 5] assumes that every class is knownwhen a model is tested. However, this assumption is unrealistic, since situati-ons never seen before will most likely occur in a real world scenario. In these realworld domains, the environment in which the AI agents act are dynamic, fullof uncertainty, and even high-risky, demanding a continuous update of sensoryinformation and knowledge bases of the agents. Self-driving cars are a para-digmatic example of these agents. The environment is continuously populatedwith novel classes of objects, novel patterns of behaviour, that were not acquiredduring pre-training and simulation. This is a hard problem for the close-worldassumption of AI, being better faced with and open-world AI approach.As such, many techniques in the field of Open World Artificial Intelligence(AI) or Open Set Recognition [4] have been developed recently. Unlike the clas-sical approach to AI, also called closed set AI [4, 5], open set recognition assumesthe possibility of unseen or unknown classes during testing. This is a more rea-listic approach to problem solving, since in real world scenarios machines mustbe ready to deal with situations they have not yet encountered.Many advances in recent years have been made in order to advance this field,as various algorithms mainly for image and text document classification havebeen proposed. For instance, Openmax [1] is an algorithm for calculating theprobability of unknown classes relying on the already established Softmax algo-rithm for probability calculation on closed sets [6, 15, 3]. The DOC algorithm[13] for text documents was proposed in 2017 and represents a great advance inthe area, as this algorithm became a very well established baseline to comparewith new algorithms [17, 9, 14, 7]. Other relevant advances in the area is theability to integrate new classes into the set of known classes at run time [5, 16],without the need to retrain the algorithm, since this is a very time consumingprocess.MND [2] andInformed Democracy[10] are another two examples of recentadvances within Open World AI. The former, MND, is based in a SegregationNetwork and its main difference is that it verifies at each step if all the imageshave the right labels. The latter,Informed Democracy, is based on a democracysystem, constituted byleadersandcouncils. The main new feature in this1
algorithm is the possibility of the initial decision be revoked by the council anda different one be issued as final.Typical active leaning algorithms [12, 11] are based on the closed worldAI assumption, assuming that the set of classes are know at beginning of thelearning task and remain the same along the learning task. Modifying thisparadigm is mandatory so that active learners can be more adapted to realworld situations [8]. A possible solution may be adapting current open worldAI algorithms to the active learning setting.
Objetivo
The focus of this work is open world AI in the context of Active Learning,in the sense that: (i) the algorithm will try to learn new classes over timeand integrate them in the training set, just like classic active learning; (ii)the algorithm considers the possibility of finding new unknown classes when selecting unlabelled instances to be queried to and then annotated by the oracle.The general goal is bringing together Open World AI and Active Learningalgorithms in the context of self-driving cars.In concrete, we aim at reaching the following goals:
- Analysis of Open AI and Active Learning algorithms
- Analysis of data sets and learning algorithms in self-driving cars
- Analysis and specification of Open AI algorithms for Active Learning
- Implementation and testing of the proposed Open World, Active LearningAlgorithm in the domain of self-driving cars.
Well known public data sets from popular projects of autonomous cars such as those available in https://waymo.com/open/ (Waymo's data set), https://analyticsindiamag.com/top-10-popular-datasets-for-autonomous-driving-projects/ or https://datasetsearch.research.google.com/search?query=self%20driving%20car&docid=L2cvMTFsajBxbDdqdw%3D%3D will be analysed and considered for the implementation and testing phases. Most of them comprise not only simulation but also large amount of real data, which have been used by real autonomous cars that populate a few cities in the world.
Plano de Trabalhos - Semestre 1
1- State of the Art [Sept – Oct]
2- Analysis of public data sets, and Open AI and Active Learning algorithms for self-driving cars [Nov]
3- Design, development and First Implementation of an Open AI and Active Learning algorithm for self-driving cars [Dec – Jan]
4- Thesis Proposal Writing [Dec – Jan]
Plano de Trabalhos - Semestre 2
5- Improvement of the Open AI and Active Learning algorithm for self-driving cars [Feb – Apr]
6- Experimental Tests [Apr – May]
7- Paper Writing [May – Jun]
8- Thesis Writing [Jan – Jul]
Condições
The eligible student will have at disposal all the necessary computational plat-forms, tools and devices.
Observações
[1] Abhijit Bendale and Terrance E. Boult. Towards open set deep networks.CoRR, abs/1511.06233, 2015
[2] Supritam Bhattacharjee, Devraj Mandal, and Soma Biswas. Multi-classnovelty detection using mix-up technique. InIEEE Winter Conference onApplications of Computer Vision, WACV 2020, Snowmass Village, CO,USA, March 1-5, 2020, pages 1389–1398. IEEE, 2020.
[3] Ken Chatfield, Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman.Return of the devil in the details: Delving deep into convolutional nets. InBritish Machine Vision Conference, 2014.
[4] Geli Fei and Bing Liu. Breaking the closed world assumption in text clas-sification. InProceedings of the 2016 Conference of the North AmericanChapter of the Association for Computational Linguistics: Human Lan-guage Technologies, pages 506–514, San Diego, California, June 2016. As-sociation for Computational Linguistics.
[5] Geli Fei, Shuai Wang, and Bing Liu. Learning cumulatively to becomemore knowledgeable. InProceedings of the 22nd ACM SIGKDD Interna-tional Conference on Knowledge Discovery and Data Mining, KDD ’16,page 1565–1574, New York, NY, USA, 2016. Association for ComputingMachinery
[6] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classi-fication with deep convolutional neural networks. In F. Pereira, C. J. C.Burges, L. Bottou, and K. Q. Weinberger, editors,Advances in NeuralInformation Processing Systems, volume 25, pages 1097–1105. Curran As-sociates, Inc., 2012.3
[7] Xue Lei, Yi Cai, Jingyun Xu, Da Ren, Qing Li, and Ho fung Leung. Incor-porating task-oriented representation in text classification. In Guoliang Li,Jun Yang, Joao Gama, Juggapong Natwichai, and Yongxin Tong, editors,Database Systems for Advanced Applications - 24th International Confe-rence, DASFAA 2019, Proceedings, Lecture Notes in Computer Science(including subseries Lecture Notes in Artificial Intelligence and Lecture No-tes in Bioinformatics), pages 401–415. Springer-Verlag, January 2019. 24thInternational Conference on Database Systems for Advanced Applications,DASFAA 2019 ; Conference date: 22-04-2019 Through 25-04-2019.
[8] Martin Mundt, Yong Won Hong, Iuliia Pliushch, and Visvanathan Ra-mesh. A wholistic view of continual learning with deep neural networks:Forgotten lessons and the bridge to active and open world learning.CoRR,abs/2009.01797, 2020.
[9] Qi Qin, Wenpeng Hu, and B. Liu. Text classification with novelty detection.ArXiv, abs/2009.11119, 2020.
[10] Alina Roitberg, Ziad Al-Halah, and Rainer Stiefelhagen. Informed demo-cracy: Voting-based novelty detection for action recognition. InBritishMachine Vision Conference 2018, BMVC 2018, Newcastle, UK, September3-6, 2018, page 52. BMVA Press, 2018.
[11] Burr Settles. From theories to queries. In Isabelle Guyon, Gavin C. Ca-wley, Gideon Dror, Vincent Lemaire, and Alexander R. Statnikov, editors,Active Learning and Experimental Design workshop, In conjunction withAISTATS 2010, Sardinia, Italy, May 16, 2010, volume 16 ofJMLR Proce-edings, pages 1–18. JMLR.org, 2011.
[12] Burr Settles.Active Learning. Synthesis Lectures on Artificial Intelligenceand Machine Learning. Morgan & Claypool Publishers, 2012
[13] Lei Shu, Hu Xu, and Bing Liu. DOC: Deep open classification of textdocuments. InProceedings of the 2017 Conference on Empirical Methodsin Natural Language Processing, pages 2911–2916, Copenhagen, Denmark,September 2017. Association for Computational Linguistics.
[14] Lei Shu, Hu Xu, and Bing Liu. Unseen class discovery in open-world clas-sification.CoRR, abs/1801.05609, 2018.
[15] C. Szegedy, Wei Liu, Yangqing Jia, P. Sermanet, S. Reed, D. Anguelov,D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with con-volutions. In2015 IEEE Conference on Computer Vision and PatternRecognition (CVPR), pages 1–9, June 2015.
[16] Hu Xu, Bing Liu, Lei Shu, and Philip S. Yu. Open-world learning andapplication to product classification.The World Wide Web Conference,2019.
[17] Jingqing Zhang, Piyawat Lertvittayakumjorn, and Yike Guo. Integratingsemantic knowledge to tackle zero-shot text classification. InProceedingsof the 2019 Conference of the North American Chapter of the Associationfor Computational Linguistics: Human Language Technologies, Volume 14 (Long and Short Papers), pages 1031–1040, Minneapolis, Minnesota, June2019. Association for Computational Linguistics
Orientador
Luís Miguel Machado Lopes Macedo
macedo@dei.uc.pt 📩