🧑‍🔬 Peter Battaglia

Peter Battaglia's papers here concern relational reasoning in neural networks. They ask how deep learning systems can represent entities and the relations among them, and so generalize by recombining familiar parts into new inferences and actions. Both works ask which structural assumptions should be built into an architecture rather than left to be learned from scratch, and both belong to the topic of Reasoning.

The more cited of the two, Relational inductive biases, deep learning, and graph networks (2018), is a position paper and literature review with no new experiments. It argues that hand-built structure and end-to-end learning need not be treated as rivals. It then gathers a range of earlier approaches into a single graph network framework grounded in relational inductive biases. The earlier A simple neural network module for relational reasoning (2017) introduced the Relation Network, a small add-on component. It applies one shared function to every pair of input objects, adds up the outcomes, and feeds that total into a second network. The objects can be cells of a convolutional feature map, encoded sentences, or rows of a state description.

Taken in order, the two papers move from a single pairwise module toward a general framework for computation over graphs. Battaglia's current affiliation is Google DeepMind in London. Earlier bylines include the Massachusetts Institute of Technology, Google in Mountain View and the University of Minnesota. Adam Santoro, David Raposo and Mateusz Malinowski appear as co-authors on both papers here. Other collaborators include Timothy Lillicrap, Caglar Gulcehre and Francis Song.

Peter Battaglia has 2 papers in our library, in 2017–2018, most of them on Reasoning; often with Adam Santoro, David Raposo, Mateusz Malinowski.