Speaker: Dr Tim Hospedales; Reader in Image and Vision Computing, Institute of Perception, Action and Behaviour
Title: Explorations in the borderlands between vision and language
Abstract: In recent years, computational linguistics and computer vision research have developed a sizeable overlap. In this talk I will give a tour of my explorations in this area. These include learning from social media data, grounding weakly supervised text tags with individual image regions, zero-shot learning (How can we use distributional semantic models of words or paragraphs to induce visual recognisers?), learning multi-modal word embeddings and going beyond vector-valued word embeddings to embedding words as probability distributions, deep neural captioning and visual question answering. I aim to present both some mature research and some speculative ideas that I hope will generate fruitful discussion.
Speaker: Andrew Brock; Current Student, CDT in Robotics
Title: One-Shot Model Architecture Search through HyperNetworks (SMASH)
Abstract: Designing architectures for deep neural networks requires expert knowledge and substantial validation time. I propose a technique to accelerate model selection by learning an auxiliary HyperNet that generates the weights of a main model conditioned on that model’s architecture. By comparing the relative validation performance of networks with HyperNet-generated weights, one can effectively search over a wide range of architectures at the cost of a single training run. To facilitate this search, I develop a flexible mechanism based on memory read-writes that allows for the definition of a wide range of network connectivity patterns, with ResNet, DenseNet, and FractalNet blocks as special cases. I validate my method (SMASH) on CIFAR-100 and CIFAR-10, achieving competitive performance with similarly-sized hand-designed networks.