Rezka's master thesis defence - Tomorrow, 07/08, 13:00, F1.15

[Elia Bruni]

Hi all,

Tomorrow, 07/08, at 13:00, room F1.15, Rezka will defend his master thesis. Anyone interested is welcome to attend. I paste the abstract of the thesis below.

Best,

Elia

Learning to follow instructions

 An artificial intelligence (AI) systems will be useful to humans if it can communicate with us: understanding our intention, performing actions that we assigned to it, and learning new skills from interactions. Hence, understanding and executing instructions given by humans is an important step to develop general useful AI. A system should be able to carry out new instructions by harnessing previously acquired knowledge of other instructions. It also needs to be flexible to master a continuous flow of new instructions while at the same time maintaining a data-efficiency in understanding and executing new instructions. This thesis describes our research in exploring deep learning models to follow instructions that it has never encountered before. We explore two task consisting of a different type of instructions. The first task consists in understanding and processing descriptions of subregular languages. In this task, the instructions come in the form of grammar specifications. The second task

that we explore is the SHRDLURN task. In this task, the instructions come in the form of natural language utterances. Our results on the first task show that to some extent, a learning agent can generalize to unseen situations. We found that an LSTM is capable of generalizing if it can make use of attention mechanism to attend to the relevant parts of the input grammar. Our results on the second task demonstrate that a learning agent can also learn some useful information to adapt to instructions it has never seen before by exploiting prior knowledge which is learned automatically from another set of instructions. We found that when neural network models are exposed to similar language instructions

to the instructions they have been trained on, they can adapt efficiently to unseen words. Even if the unseen language instructions are very different, the models can still make use of the automatically learned prior knowledge