Speaker:  Manik Varma
Title: Extreme Classification: A New Paradigm for Ranking & Recommendation

The objective in extreme multi-label classification is to learn a classifier that can automatically tag a
data point with the most relevant subset of labels from a large label set. Extreme multi-label classification
 is an important research problem since not only does it enable the tackling of applications with many
 labels but it also allows the reformulation of ranking and recommendation problems with certain 
advantages over existing formulations.

Our objective, in this talk, is to develop an extreme multi-label classifier that is faster to train and more 
accurate at prediction than the state-of-the-art Multi-label Random Forest (MLRF) algorithm
 [Agrawal et al. WWW 13] and the Label Partitioning for Sub-linear Ranking (LPSR) algorithm 
[Weston et al. ICML 13]. MLRF and LPSR learn a hierarchy to deal with the large number of labels but 
optimize task independent measures, such as the Gini index or clustering error, in order to learn the 
hierarchy. Our proposed FastXML algorithm achieves significantly higher accuracies by directly 
optimizing an nDCG based ranking loss function. We also develop an alternating minimization 
algorithm for efficiently optimizing the proposed formulation. Experiments reveal that FastXML can be 
trained on problems with more than a million labels on a standard desktop in eight hours using a single 
core and in an hour using multiple cores.

Speaker:  Jason Weston
Title: Hashtags, Clicks and Likes: Supervision for Content-based Posts

We study the problem of understanding the content of short textual posts and images popular in social networks. There are an abundance of weakly supervised signals for such a goal, including hashtags provided by the users themselves, as well as various kinds of click, comment and like
interactions. We show how employing word embedding and image embedding
convolutional neural network models can effectively utilize this supervision. This is joint work with Keith Adams, Lubomir Bourdev, Sumit Chopra, Misha Denil, Emily Denton, Rob Fergus, Manohar Paluri, Marc-Aurelio Ranzato, Ledell Wu and Ming Yang.


Speaker:  John Langford
Title: The Elusive Theory of Efficient Classification

What is known about the theory of efficient classification?  How far does the theory take us?  And why doesn't it take us all the way?  I'll discuss the key results and open questions I see for extreme multiclass classification.


Speaker:  Jia Deng
Title: Knowledge-Driven Recognition of Objects and Actions

A fundamental challenge in visual recognition is generalizing to new concepts. As we expand the label space from basic object categories to fine-grained classes and to compositions of visual entities (i.e. actions), most of the visual concepts will be in the long tail, i.e., most will have very few or even zero training examples. It is thus critical to leverage semantic knowledge, facts about how concepts are related, to enable generalization from common concepts to rare ones. In this talk I will
discuss some recent work in this direction. I will first present a new visual classification model that incorporates pre-defined semantic relations between labels, namely hierarchy and exclusion. Then I will show how to jointly learn the visual classification model and the semantic
relations in the context of action recognition.