antoine wystrach and Miaojing Shi
Thu 08 Oct 2015, 12:45 - 13:45

If you have a question about this talk, please contact: Steph Smith (ssmith32)

Speaker: Antoine Wystrach

Title: Kung-fu larva

Abstract: The larvae of Drosophila Melanogaster spontaneously crawl towards or away from stimuli such as odours, light, or temperature gradients. These taxis behaviours are described as a combination of several, distinct, behavioural modules triggering specific actions. Some actions appears determined to sample information from the environment (e.g., casting the head to the sides), while others are qualified as a 'goal-directed behaviours' (e.g., turning and crawling in a particular direction); and one can argue that the transition between active sampling and goal-directed behaviours requires ‘decision-making’ and ‘short-term memory’. Here we show that all these apparent distinctions can emerge from a simple and unique core mechanism: a continuous turning oscillator on which all sensory modalities, innate or learnt, converge. In addition to the robustness of the mechanism and the ease for its implementation, we were ourselves surprised by the number of phenomenon observed in larvae that spontaneously emerged from such a simple model, even though we had no intention of capturing them. When in a closed-loop with the environment, sampling and goal-directed behaviours become one and the same, thus sparing the need for 'decision-making’ or ‘short-term memory’. According to the 'tao', which is said to be both the source of, and the hidden force behind all apparent phenomenon; one needs to be in harmony with nature in order to achieve 'effortless action'.


Speaker: Miaojing Shi

Title: a group testing framework for similarity search

Abstract: In this talk, I will present a group testing framework for detecting large similarities between high-dimensional vectors, such as descriptors used in state-of-the-art description of multimedia documents. At the crossroad of multimedia information retrieval and signal processing, a set of group representations are produced that jointly encode several vectors into a single one, in the spirit of group testing approaches.  By comparing a query vector to several of these intermediate representations, one can screen the large values taken by the similarities between the query and all the vectors, at a fraction of the cost of exhaustive similarity calculation. Unlike concurrent indexing methods that suffer from the curse of dimensionality, proposed method exploits the properties of high-dimensional spaces. It therefore complements other strategies for approximate nearest neighbor search.