LAB at NIPS 2016

We are delighted to inform that the LAB group will contribute to three different workshops and one symposium at this years Neural Information Processing Systems conference in Barcelona. The conference is one of the highlights of the machine learning year and takes place between the 5th and 10th of December, 2016.

Olof Mogren will present his work C-RNN-GAN: Continuous recurrent neural networks with adversarial training at the Constructive Machine Learning Workshop.

Mikael Kågebäck and Emilio Jorge will present their work Learning to Play Guess Who? and Inventing a Grounded Language as a Consequence at the Deep Reinforcement Learning Workshop.

Fredrik Johansson will present his, Uri Shalit and David Sontag’s work on estimating individual treatment effects at the What If? Workshop. Uri Shalit will also present their work at the Deep Learning Symposium.

Vacancy: Associate Professor in Machine Learning

There is currently a vacant position at the department of Computer Science and Engineering.

Associate Professor in Machine Learning

Major responsibilities
The position of Associate Professor is a full-time faculty position. You will teach courses at the bachelor and masters level. You will also lead research activities and supervise PhD students and master students. You are expected to be an active researcher and to prepare applications for the funding of research projects, as well as to promote the continuation of ongoing projects.

Position summary
Full-time permanent position.

We are looking for applicants from all subfields of machine learning with a PhD in computer science (or an equivalent qualification). Applicants with some of the following qualifications are especially appreciated:

* a strong research record;
* a strong pedagogical track record;
* potential to lead new research projects and teaching methods
* core competence in machine learning;
* a record of implementing scalable methods and systems;
* a record of applications in some domain;
* openness to industry collaboration.

The position of associate professor requires scientific expertise that is considerably higher than that required for a doctoral degree (corresponding to admission as docent) and at least 15 higher education credits in teaching and learning in higher education or corresponding expertise. You must be accustomed to teaching and be a skilled educator. For this reason, we place a great deal of emphasis on your pedagogical portfolio. Experience in conducting academic research and/or development in industry/the public sector is a requirement.

Deep Learning and NLP

The LAB group have seen several successes recently in deep learning and natural language processing. Olof Mogren’s work on named entity recognition using character-based LSTMs was recently accepted for publication at the Biomedical Text Mining workshop at Coling, BIOTXTM. Soon after, a paper on word sense disambiguation using LSTMs by Mikael Kågebäck and Hans Salomonsson was accepted to the workshop Cognitive Aspects of the Lexicon, CogALex. Mikael Kågebäck and Emilio Jorge’s work on inventing and learning languages by playing games will be presented at a NIPS workshop. Additionally, the work on causality by Fredrik Johansson and collaborators will be featured on the Deep Learning Symposium at NIPS 2016.

Publications in top-tier conferences

We are happy to report that our group has seen recent success through papers accepted at two top-tier conferences.

Chien-Chung Huang, Naonori Kakimura and Naoyuki Kamiyama will have their paper Exact and Approximation Algorithms for Weighted Matroid Intersection published in the very competitive SODA 2016.

A few weeks earlier, Fredrik Johansson, Ankani Chattoraj, Devdatt Dubhashi and Chiranjib Bhattacharyya had their paper Weighted Theta Functions and Embeddings with Applications to Max-Cut, Clustering and Summarization accepted for publication at NIPS 2015.


Exact and Approximation Algorithms for Weighted Matroid Intersection
Chien-Chung Huang, Naonori Kakimura and Naoyuki Kamiyama

We present exact and approximation algorithms for the weighted matroid intersection problems. Our exact algorithms are faster than previous algorithms when the largest weight is relatively small. Our approximation algorithms deliver a (1-∈)-approximate solution with running times significantly faster than known exact algorithms. The core of our algorithms is a decomposition technique: we decompose the weighted version of the problem into a set of unweighted matroid intersection problems. The computational advantage of this approach is that we can then make use of fast unweighted matroid intersection algorithms as a black box for designing algorithms. To be precise, we show that we can find an exact solution via solving W unweighted matroid intersections problems, where W is the largest given weight. Furthermore, we can find a (1-∈)-approximate solution via solving O(∈^{-1} log r) unweighted matroid intersection problems, where r is the smallest rank of the given two matroids.


Weighted Theta Functions and Embeddings with Applications to Max-Cut, Clustering and Summarization
Fredrik Johansson, Ankani Chattoraj, Devdatt Dubhashi and Chiranjib

We introduce a unifying generalization of the Lovász theta function, and the associated geometric embedding, for graphs with weights on both nodes and edges. We show how it can be computed exactly by semidefinite programming, and how to approximate it using SVM computations. We show how the theta function can be interpreted as a measure of diversity in graphs and use this idea, and the graph embedding in algorithms for Max-Cut, correlation clustering and document summarization, all of which are well represented as problems on weighted graphs.