relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · alvaro...
TRANSCRIPT
![Page 1: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/1.jpg)
Relational inductive biases, deep learning, and graph
networksPeter W. Battaglia, Jessica B. Hamrick, Victor Bapst,
Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski,Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner,Caglar Gulcehre, Francis Song, Andrew Ballard, Justin Gilmer,
George Dahl, Ashish Vaswani, Kelsey Allen, Charles Nash,Victoria Langston, Chris Dyer, Nicolas Heess,
Daan Wierstra, Pushmeet Kohli, Matt Botvinick,Oriol Vinyals, Yujia Li, Razvan Pascanu
Presentation: Sebastian Värv
![Page 2: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/2.jpg)
Human vs computer● Ability to make “infinite use of infinite means"● Combinatorial generalization● Representation of complex systems● Estimating if stack of objects is stable.
Battaglia et al. 0
![Page 3: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/3.jpg)
Current challenges● Complex language and scene understanding ● Reasoning about structured data● Transferring learning beyond training conditions● Learning from small amounts of experience
A.Daniels et al. 08
![Page 4: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/4.jpg)
Relational reasoning and inductive bias
![Page 5: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/5.jpg)
Relational reasoning and inductive bias● Combining relational reasoning and inductive bias● Hand engineered models do not generalize● End-to-end learning lacks control and does not have clear structure to
represent relations● Maybe combining these methods works
![Page 6: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/6.jpg)
Relational reasoning● Defining the structure - composing set of known building blocks.● Entity - element with attributes.● Relation - property between entities.● Rule - function which maps entities and relations to other entities and
relations.
![Page 7: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/7.jpg)
Inductive biases● Multiple solutions to the problem, both equally good● We use inductive bias to choose one solution
In neural nets context:
● Regularization term ○ Dropout○ L2 regularization etc.
![Page 8: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/8.jpg)
Relational inductive biases● Inductive biases which impose constraints on relationships and interactions
among entities in a learning process.
![Page 9: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/9.jpg)
Relational inductive bias in DL building blocks
![Page 10: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/10.jpg)
Graph networks
![Page 11: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/11.jpg)
Background● (Neural networks) that operate on graphs● Structure the computations accordingly● First discussed by Gori et al. in 2005.● Became popular in recent years, still alive (2016- …)● Effective at tasks thought to have rich relational structure (scene
understanding or one-shot learning)
![Page 12: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/12.jpg)
Example use case
Graph Networks as Learnable Physics Engines for Inference and Control by A.Sanchez-Gonzales et al.
https://drive.google.com/file/d/1xZme1bxvUWQeb9fWFelECR7YKUXur4XU/view
![Page 13: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/13.jpg)
Task● Predicting the state after n steps (understanding environment)
![Page 14: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/14.jpg)
Resultshttps://drive.google.com/file/d/15dEUgf5T4ddehMgZiVQ2FtXGqJi9JIDn/view
![Page 15: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/15.jpg)
Graph
![Page 16: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/16.jpg)
Graph network (GN)● Authors propose generalised GN framework which should satisfy all existing
models.● GN Block - “graph-to-graph” module
![Page 17: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/17.jpg)
Sample task ● Predict movements a set of rubber balls in
arbitrary gravitational field● Instead of bounding against each other they
have one or more springs to connect with some others.
● Task to predict the position of each rubber ball after certain time.
![Page 18: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/18.jpg)
GN Block
![Page 19: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/19.jpg)
Update sequence in GN block
![Page 20: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/20.jpg)
![Page 21: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/21.jpg)
Configurable structure
![Page 22: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/22.jpg)
Full graph network structure
![Page 23: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/23.jpg)
Obtaining graph structure● Defining how the input data will be represented as graph manually● Input may explicitly specify relational structure● This structure may be inferred or assumed.
![Page 24: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/24.jpg)
Limitations of graph networks● Limitations inherited directly from graph structure● Implementation difficulty
![Page 25: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/25.jpg)
Conclusion● Blending powerful deep learning approaches with structured representations● Moving a bit closer to combinatorial generalization● Proposed framework for representing all kinds of relational data using GN
![Page 26: Relational inductive biases, deep learning, and graph networks · 2018. 11. 4. · Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo,](https://reader035.vdocuments.net/reader035/viewer/2022071019/5fd3693ac2a22801b84b7d4b/html5/thumbnails/26.jpg)
Thank you for listening!