Octobre 2010 - Lisa_seminaires - mailman.iro.umontreal.ca

UdeM-McGill-MITACS machine learning seminar Fri Oct. 29@14h30, AA-3195
by Dumitru Erhan 29 Oct '10

29 Oct '10

The UdeM-McGill-MITACS machine learning seminar series is continuing its fall schedule. *Friday*'s seminar: Tale of a Neural Network: From Part-Of-Speech to Parsing. by Ronan Collobert NEC Labs Location: Pavillon André-Aisenstadt (UdeM), room AA-3195 Time: *Friday, October 29, 14:30* Abstract: We will present a single architecture which excels in performance on various Natural Language Processing (NLP) tasks. Instead of hand-crafting task-specific features, generic word representations are *learnt* from large unlabeled corpuses. We will demonstrate how our architecture naturally applies to a wide range of tasks, from simple tasks like Part-Of-Speech, to complex tasks like Parsing. Analysis and comparison with existing NLP approaches will be given. The presentation will end with the introduction of a simple standalone software, implementing all these tasks with blazing execution speed.

1 1

UdeM-McGill-MITACS machine learning seminar Tue Oct. 26@14h00, AA-3195
by Dumitru Erhan 25 Oct '10

25 Oct '10

The UdeM-McGill-MITACS machine learning seminar series<http://www.iro.umontreal.ca/article.php3?id_article=107&lang=en>is continuing its fall schedule. Next week's seminar: *Sequence Processing with Recurrent Neural Networks * by Alex Graves Technische Universität München, Germany Location: Pavillon André-Aisenstadt (UdeM), room AA-3195 Time: Tuesday, October 26, *14:00* Abstract: The need to label, predict, model or otherwise process sequential data arises in many domains. Some obvious examples are speech and handwriting recognition, inferring protein structure from amino acid sequences, and making predictions about the stock market or the weather. Although the tasks and data vary greatly, certain properties seem to be universally beneficial for sequence processing algorithms. In particular the ability to 'remember' the context surrounding each point in the data sequence, and the ability to generalise to distorted versions of sequential patterns appear to be crucial. Recurrent neural networks are a promising architecture in this regard. Furthermore, they can be generalised to treat spatial, as well as temporal, dependencies in a sequential way, expanding the range of applicability to data such as images, video, and medical scans. This talk presents a generic framework for processing spatio-temporal data using recurrent neural networks, and shows how the framework can be tailored to specific tasks.

1 1

UdeM-McGill-MITACS machine learning seminar Tue Oct. 19@15h00, AA-3195
by Dumitru Erhan 19 Oct '10

19 Oct '10

The UdeM-McGill-MITACS machine learning seminar series<http://www.iro.umontreal.ca/article.php3?id_article=107&lang=en>is continuing its fall schedule. Next week's seminar: *Large Scale Image and Music Annotation: Learning to Rank and Multi-Tasking with Joint Embeddings * by Jason Weston Google Research, NY Location: Pavillon André-Aisenstadt (UdeM), room AA-3195 Time: Tuesday, October 19, *15:00* Abstract: In the first part of the talk we will discuss large scale image annotation. Image annotation datasets are becoming larger and larger, with tens of millions of images and tens of thousands of possible annotations. We propose a well performing method that scales to such datasets by simultaneously learning to optimize precision at k of the ranked list of annotations for a given image \em and learning a low-dimensional joint embedding space for both images and annotations. Our method both outperforms several baseline methods and, in comparison to them, is faster and consumes less memory. We also demonstrate how our method learns an interpretable model, where annotations with alternate spellings or even languages are close in the embedding space. Hence, even when our model does not predict the exact annotation given by a human labeler, it often predicts similar annotations, a fact that we try to quantify by measuring the newly introduced ``sibling’’ precision metric, where our method also obtains good results. In the second (shorter) part of the talk we will discuss large scale music annotation. Music prediction tasks range from predicting the genre, style or the artist given a song or clip of audio, predicting similar artists given an artist, or predicting related songs given a song, clip, artist name or genre or style tag. That is, we are in interested in pretty much every semantic relationship between the different musical concepts in our database. In realistic databases, the number of songs is measured in the millions, and the number of artists in the tens of thousands or more, providing a considerable challenge to standard machine learning techniques. In this work, we propose a method that scales to such datasets which attempts to capture the semantic similarities between the database items by modeling audio, artist names, and genre and style tags in a single low-dimensional semantic space. This choice of space is learnt by optimizing the set of predictions tasks of interest jointly using multi-task learning. Our method both outperforms baseline methods and, in comparison to them, is faster and consumes less memory. We then demonstrate how our method learns an interpretable model, where the semantic space captures well the similarities of interest. Joint work with Samy Bengio and Nicolas Usunier.

1 1

UdeM-McGill-MITACS machine learning seminar Thu Oct. 7@14h00, AA-3195
by Dumitru Erhan 06 Oct '10

06 Oct '10

The UdeM-McGill-MITACS machine learning seminar series is back with an exciting fall schedule. Next week's seminar (see http://www.iro.umontreal.ca/article.php3?id_article=107&lang=en) Learning Spatial and Transformational Invariants for Visual Representation by Charles Cadieu Redwood Center for Theoretical Neuroscience University of California, Berkeley Location: Pavillon André-Aisenstadt (UdeM), room AA-3195 Time: Thursday, October 7, 14:00 Abstract: Learning abstract, invariant properties of the visual word is a key attribute of biological vision systems. I will describe a hierarchical, probabilistic model that learns to extract invariant spatial structure and invariant motion structure from movies of the natural environment. The first layer in the model produces a sparse decomposition of local edge and motion structure. This decomposition is achieved through a complex-valued sparse coding model in which amplitudes represent the presence of edge structure at specific positions, orientations, and spatial scales, and phases represent the precise positions of edges and how image structure changes through time. This decomposition into amplitude and phase exposes statistical dependencies that the top layer in the model captures as two types of invariances: spatial and transformational. The spatial invariants are a sparse code of the patterns in the first layer amplitude components. They code a rich set of multi-scale edges, textures, and texture defined edge boundaries. The transformational invariants are a sparse representation of patterns of change in the first layer phase components. They learn a multi-scale code of motion, spanning local and global motions, and are capable of learning complex motions such as zooming, rotation, and deformation. Besides extracting abstract, invariant properties of the visual world, I will show how the hierarchical model provides a concrete model of cortical feedback that is useful for perception under noisy or ambiguous conditions.

1 1