Relational inductive biases, deep learning, and graph networks Artificial intelligence has made significant progress in key domains such as vision, language, control, and decision-making, driven by cheap data and compute resources. However, generalizing beyond experiences remains a challenge for AI. This paper argues that combinatorial generalization is crucial for AI to achieve human-like abilities, and that structured representations and computations are key to this goal. We advocate for an approach that combines structured and end-to-end learning, and explore how relational inductive biases can facilitate learning about entities, relations, and rules for composing them. We introduce a new building block for the AI toolkit—the graph network—which generalizes and extends various approaches for neural networks that operate on graphs, and provides a straightforward interface for manipulating structured knowledge and producing structured behaviors. We discuss how graph networks can support relational reasoning and combinatorial generalization, laying the foundation for more sophisticated, interpretable, and flexible patterns of reasoning. We also release an open-source software library for building graph networks. Graph networks are a class of models that focus on reasoning about explicitly structured data, particularly graphs. They perform computation over discrete entities and the relations between them, and can be learned without needing to specify them in advance. They carry strong relational inductive biases, in the form of specific architectural assumptions, which guide them towards learning about entities and relations. We define structure as the product of composing a set of known building blocks. "Structured representations" capture this composition, and "structured computations" operate over the elements and their composition as a whole. Relational reasoning involves manipulating structured representations of entities and relations, using rules for how they can be composed. We explore various deep learning methods through the lens of their relational inductive biases, showing that existing methods often carry relational assumptions which are not always explicit or immediately evident. We then present a general framework for entity- and relation-based reasoning—which we term graph networks—for unifying and extending existing methods which operate on graphs, and describe key design principles for building powerful architectures using graph networks as building blocks. We have also released an open-source library for building graph networks, which can be found here: github.com/deepmind/graph_nets. Graph networks support arbitrary (pairwise) relational structure and computations over graphs afford a strong relational inductive bias beyond that which convolutional and recurrent layers can provide. The main unit of computation in the GN framework is the GN block, a "graph-to-graph" module which takes a graph as input, performs computations over the structure, and returns a graph as output. The GN framework's block organization emphasizes customizability and synthesizing new architectures which express desired relational inductive biases. The key design principles are: Flexible representations; Configurable within-block structure; and Composable multi-block architectures. We introduce a motivating example to help make the GN formalism more concrete. Consider predicting the movements of a set of rubber balls in an arbitrary gravitational field, which, insteadRelational inductive biases, deep learning, and graph networks Artificial intelligence has made significant progress in key domains such as vision, language, control, and decision-making, driven by cheap data and compute resources. However, generalizing beyond experiences remains a challenge for AI. This paper argues that combinatorial generalization is crucial for AI to achieve human-like abilities, and that structured representations and computations are key to this goal. We advocate for an approach that combines structured and end-to-end learning, and explore how relational inductive biases can facilitate learning about entities, relations, and rules for composing them. We introduce a new building block for the AI toolkit—the graph network—which generalizes and extends various approaches for neural networks that operate on graphs, and provides a straightforward interface for manipulating structured knowledge and producing structured behaviors. We discuss how graph networks can support relational reasoning and combinatorial generalization, laying the foundation for more sophisticated, interpretable, and flexible patterns of reasoning. We also release an open-source software library for building graph networks. Graph networks are a class of models that focus on reasoning about explicitly structured data, particularly graphs. They perform computation over discrete entities and the relations between them, and can be learned without needing to specify them in advance. They carry strong relational inductive biases, in the form of specific architectural assumptions, which guide them towards learning about entities and relations. We define structure as the product of composing a set of known building blocks. "Structured representations" capture this composition, and "structured computations" operate over the elements and their composition as a whole. Relational reasoning involves manipulating structured representations of entities and relations, using rules for how they can be composed. We explore various deep learning methods through the lens of their relational inductive biases, showing that existing methods often carry relational assumptions which are not always explicit or immediately evident. We then present a general framework for entity- and relation-based reasoning—which we term graph networks—for unifying and extending existing methods which operate on graphs, and describe key design principles for building powerful architectures using graph networks as building blocks. We have also released an open-source library for building graph networks, which can be found here: github.com/deepmind/graph_nets. Graph networks support arbitrary (pairwise) relational structure and computations over graphs afford a strong relational inductive bias beyond that which convolutional and recurrent layers can provide. The main unit of computation in the GN framework is the GN block, a "graph-to-graph" module which takes a graph as input, performs computations over the structure, and returns a graph as output. The GN framework's block organization emphasizes customizability and synthesizing new architectures which express desired relational inductive biases. The key design principles are: Flexible representations; Configurable within-block structure; and Composable multi-block architectures. We introduce a motivating example to help make the GN formalism more concrete. Consider predicting the movements of a set of rubber balls in an arbitrary gravitational field, which, instead