Zephyrnet Logo

DeepMind Papers @ NIPS (Part 1)

Date:

Interaction Networks for Learning about Objects, Relations and Physics

Authors: Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Rezende, Koray Kavukcuoglu

Reasoning about objects, relations, and physics is central to human intelligence, and a key goal of artificial intelligence. However many modern machine learning methods still face a trade-off between expressive structure and efficient performance. 

We introduce “interaction networks”, which can reason about how objects in complex systems interact, supporting dynamical predictions, as well as inferences about the abstract properties of the system. Interaction networks are both expressive and efficient because they combine three powerful approaches: structured models, simulation, and deep learning. They take as input graph-structured data, perform object- and relation-centric reasoning in a way that is analogous to a simulation, and are implemented using deep neural networks. They are invariant to permutations of the entities and relations, which allows them to automatically generalize to systems of different sizes and structures than they have experienced during training.

In our experiments, we used interaction networks to implement the first general-purpose learnable physics engine. After training only on single step predictions, our model was able to simulate the physical trajectories of n-body, bouncing ball, and non-rigid string systems accurately over thousands of time steps. The same architecture was also able to infer underlying physical properties, such as potential energy. 

Beyond physical reasoning, interaction networks may provide a powerful framework for AI approaches to scene understanding, social perception, hierarchical planning, and analogical reasoning.

For further details and related work, please see the paper

For applications of interaction networks to scene understanding and imagination-based decision-making, please see our submissions to ICLR 2017: Discovering objects and their relations from entangled scene representations and Metacontrol for Adaptive Imagination-Based Optimization

Check it out at NIPS:

Mon Dec 5th 06:00 – 09:30 PM @ Area 5+6+7+8 #48

Fri Dec 9th 08:00 – 6:30 PM @ Hilton Diag. Mar, Blrm. C

Source: https://deepmind.com/blog/article/deepmind-papers-nips-part-1

spot_img

Latest Intelligence

spot_img