TY - CONF
T1 - Automatically composing representation transformations as a means for generalization
AU - Chang, Michael B.
AU - Levine, Sergey
AU - Gupta, Abhishek
AU - Griffiths, Thomas L.
N1 - Funding Information:
The authors would like to thank the anonymous ICLR reviewers and commenters, Alyosha Efros, Dinesh Jayaraman, Pulkit Agrawal, Jason Peng, Erin Grant, Rachit Dubey, Thanard Kurutach, Parsa Mahmoudieh, Aravind Srinivas, Fred Callaway, Justin Fu, Ashvin Nair, Marvin Zhang, Shubham Tulsiani, Peter Battaglia, Jessica Hamrick, Rishabh Singh, Feras Saad, Michael Janner, Samuel Tenka, Kai-I Shan, David Chang, Mei-ling Hsu, Tony Chang and others in the Berkeley Artificial Intelligence Research Lab for helpful feedback, discussions, and support. The authors are grateful for computing support from Amazon, NVIDIA, and Google. This work was supported in part by the Berkeley EECS Department Fellowship for first-year Ph.D. students, travel funding from Bloomsbury AI, contract number FA8650-18-2-7832 from the Defence Advanced Research Projects Agency (DARPA) under the Lifelong Learning Machines program, contract number FA9550-18-1-0077 from the Air Force Office of Scientific Research (AFOSR), and the National Science Foundation (NSF) Graduate Research Fellowship Program. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of DARPA, AFOSR, or the NSF.
Publisher Copyright:
© 7th International Conference on Learning Representations, ICLR 2019. All Rights Reserved.
PY - 2019
Y1 - 2019
N2 - A generally intelligent learner should generalize to more complex tasks than it has previously encountered, but the two common paradigms in machine learning - either training a separate learner per task or training a single learner for all tasks - both have difficulty with such generalization because they do not leverage the compositional structure of the task distribution. This paper introduces the compositional problem graph as a broadly applicable formalism to relate tasks of different complexity in terms of problems with shared subproblems. We propose the compositional generalization problem for measuring how readily old knowledge can be reused and hence built upon. As a first step for tackling compositional generalization, we introduce the compositional recursive learner, a domain-general framework for learning algorithmic procedures for composing representation transformations, producing a learner that reasons about what computation to execute by making analogies to previously seen problems. We show on a symbolic and a high-dimensional domain that our compositional approach can generalize to more complex problems than the learner has previously encountered, whereas baselines that are not explicitly compositional do not.
AB - A generally intelligent learner should generalize to more complex tasks than it has previously encountered, but the two common paradigms in machine learning - either training a separate learner per task or training a single learner for all tasks - both have difficulty with such generalization because they do not leverage the compositional structure of the task distribution. This paper introduces the compositional problem graph as a broadly applicable formalism to relate tasks of different complexity in terms of problems with shared subproblems. We propose the compositional generalization problem for measuring how readily old knowledge can be reused and hence built upon. As a first step for tackling compositional generalization, we introduce the compositional recursive learner, a domain-general framework for learning algorithmic procedures for composing representation transformations, producing a learner that reasons about what computation to execute by making analogies to previously seen problems. We show on a symbolic and a high-dimensional domain that our compositional approach can generalize to more complex problems than the learner has previously encountered, whereas baselines that are not explicitly compositional do not.
UR - http://www.scopus.com/inward/record.url?scp=85083951403&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85083951403&partnerID=8YFLogxK
M3 - Paper
AN - SCOPUS:85083951403
T2 - 7th International Conference on Learning Representations, ICLR 2019
Y2 - 6 May 2019 through 9 May 2019
ER -