Structured Text and Image Generation

Structured Auto-Regressive Models with Latent Predictor Networks

(Wang Ling — DeepMind Technologies Limited)

I will describe the recent advances in auto-regressive models for sequence generation applied to the generation of natural language, images and code. Such models are highly appreciated as they learn a tractable probability density by predicting the collective object (e.g., an image or a document) from its parts (e.g., pixels and words).

In the first part of the talk, I will give an overview of the model, featuring the RNN language models used for document generation, and the PixelCNN model used for image generation. Then, I will describe the recent advances that have been accomplished in these areas of research.

The second part of the talk will focus on Latent Predictor Networks, which provides a framework for learning generation structure in auto-regressive models, while preserving the tractability of the probabilistic model. More specifically, I will show that the generation process in images and language can be augmented by learning to articulate different forms of prediction. Using this method we build image generation models that can learn to generate new pixels by adjusting existing pixels rather than always generating them from scratch, or text generation models that can learn to copy from words in the input.

Ranking Programming Languages for their Energy Efficiency

(João Paulo Fernandes — Department of Informatics Engineering, University of Coimbra)

In this talk, we present a study of the runtime, memory usage and energy consumption of twenty seven well-known software languages. We monitor the performance of such languages using ten different programming problems, expressed in each of the languages.

Our analysis of the obtained data not only establishes a global ranking for all the 27 programming languages, as it also ranks them within categories such as execution type and paradigm.

The results of such analysis show interesting findings, such as, slower/faster languages consuming less/more energy, and how memory usage influences energy consumption. We show how to use our results to provide software engineers support to decide which language to use when energy efficiency is a concern.

Text Summarization using Sequence to Sequence Modeling

(Pedro Verruma — Talkdesk)

This talk will explain how sequence to sequence modeling works and how it can be applied to many NLP tasks. We will then focus on the text summarization example with a tutorial on how to write a small application for extracting summaries from text, automatically.