Importance of Relationships

May want to model relationships between objects in an image
Like-wise in language, we also want to model more than just sequential structure
- Model inter-relationships like parse-trees, grammatical structures, etc

To represent structured relationships, we need special networks that can do that. This comes from their architectural biases.

In graphs, connected nodes should have more similar embeddings compared to unconnected nodes.

Space of Architectures

Previously discussed structures (FC-NN, CNN) cannot model these relationships effectively
Want to learn connectivity patterns of structured representations given some data and the structure in it
- Architectural biases like RNN, Attention-Based Networks and Graph-Based Networks can model these

Graph Embedding

Embedding: Learned map from entities to vectors of numbers that encodes similarity
- Word Embeddings: word → vector
- Graph Embeddings: node → vector
With graph embedding, connected nodes should have similar embeddings compared to unconnected nodes
- This can be optimized via gradient descent

To propagate information, we use a vector that represents not only the individual items, but also the relationships between items (edges in the graph)

By querying using the attention mechanism, we are able to combine information from other vector sets based on the current node.

This allows us to represent arbitrary sets of information

<aside> 📌 SUMMARY: Relationships in data are important as they can tell us which part is useful to look at. Instead of local dependencies like CNNs, we can have the network capture these dependencies instead like non-local neural networks.

</aside>