We are thrilled to host the Deep Learning Meetup at our Tech Hub 1st June 2016. You can expect two great talks from Roland Vollgraf and Sebastian Arnold.
Doors open at 6 pm, talks start at 6.30 pm. Please remember to bring a valid ID for entry (requested by our security staff).
See you there!
Your Zalando Tech Events Team
Talk #1: From Items to FDNA and Back - A Journey Through the Space of Fashion Articles
Roland Vollgraf, Data Science Expert, Zalando SE
Fashion DNA (FDNA) is a general representation of fashion items in some feature space of moderate dimensionality. It has the nice property that similar items end up close to each other, which allows for interesting applications like content based recommendation or style identification.
While the mapping from items to FDNA is comparably easy to perform with the help of a forward pass of a Deep Neural Network, the reverse is not that straightforward. For a given FDNA feature vector we could look which items in Zalando's stock map closest to it. But what if there are holes in the feature space, regions none of our items map into. Is it nonsense what is represented there, or could it be that there is the trending stuff of the next century?
In order to visualize the richness of the FDNA space, we trained a generative probabilistic model for fashion article images conditioned on their FDNA vector. The method we used is Neural Autoregressive Density Estimation (NADE) with a recurrent neural network (LSTM). Our model is capable of getting a FDNA vector and sampling novel fashion articles images that would map that vector in FDNA space. We could create various images that are consistent with the looks of real Zalando products but also exhibit a significant level of novelty.
Roland obtained his PhD at the Technische Universität Berlin in Machine Learning and Statistical Signal Processing. He worked as Head of Research at GA Financial Solutions GmbH and conducted the development risk models and quantitative trading strategies.
For more than three years now, he has held the position of a Data Science Expert at Zalando, where he leads a R&D team. Roland is interested in Deep Learning and Large Scale Bayesian Inference.
Talk #2 Long Short-Term Memory Networks for Text Mining Tasks
Sebastian Arnold, Beuth University of Applied Sciences Berlin, DATEXIS Group (http://www.datexis.com)
Named entity recognition often fails in idiosyncratic domains. That causes a problem for depending tasks, such as entity linking and relation extraction. We propose a generic and robust approach for high-recall named entity recognition. Our end-to-end approach is based on deep contextual sequence learning and utilizes stacked bidirectional LSTM (Long Short-Term Memory) networks. Models are trained with only few hundred labeled sentences and does neither rely on additional syntactic features nor on domain specific lexicons. We report from our results F1 scores in the range of 84–94% on standard datasets. In the second part we demonstrate our research prototypes for interactive review analysis and extraction-as-you-type.