IPAB workshop - Carl Allen, Ivana Balazevic and Jack Geary

Carl Allen:

Title: Knowledge Graph Embedding with links to Word Embedding

Abstract: Knowledge Graphs (or Knowledge Bases) are large repositories of facts in the form of binary relations between words (entities). To retrieve facts for query answering or to infer unknown facts etc, representations of knowledge graphs are learned. Entities are typically represented as vectors and relations as transformations between them. The entities can be considered as points within a semantic space, much as they are in word embedding methods, e.g. word2vec and Glove. We consider a few exemplary knowledge graph representation methods and how they relate to word embeddings.

Jack Geary:

Title: Risk-and-Response-ability based Planning: Accounting for the reactive capabilities of foreign agents in Planning for Autonomous Driving.

Abstract: Effective trajectory synthesis in highly dynamic environments requires implicit coordination between agents. This is achieved by agents accounting for the reactive capabilities of other agents, and by adjusting their trajectories to account for the behaviour of other agents. This is most evident in driving, where this coordination is necessary for safe performance. However, many current planning methods fail to account for variations in other agents’ behaviour while also being computationally tractable when synthesising trajectories. As a consequence of this the viability of manoeuvres may be inaccurately evaluated during planning.

In this work we introduce the concept of Responsibility to augment the cost functions used by current planning methods. This measure accounts for the expectation that other agents in the environment will attempt to adjust their behaviour to prevent any avoidable collisions. We present preliminary results demonstrating the value of this method in a simulated unexpected-stopping scenario.

Ivana Balazevic:

Title: Hypernetwork Knowledge Graph Embeddings

Abstract: Knowledge graphs, such as Google Knowledge Graph, Wordnet, and Freebase are large graph-structured databases of facts, which typically suffer from incompleteness. Link prediction is the task of inferring missing relations (links) between entities (nodes) in a knowledge graph. We approach this task using a hypernetwork architecture to generate convolutional layer filters specific to each relation and apply those filters to the subject entity embeddings. This architecture enables a trade-off between non-linear expressiveness and the number of parameters to learn.