Skip to content. Skip to navigation
McGill Home SOCS Home
Personal tools
You are here: Home Announcements and Events Seminars Schedule

Seminar Home
Fall 2015 Schedule
Winter 2016 Schedule
Summer 2016 Schedule

SOCS Graduate Seminar Series Seminar Schedule

Date Category Seminar Info
2013/12/02 Graduate Seminar Series Place: MC103
Time: 12 - 12:30
Speaker: Prof. Benjamin Fung
Affiliation: Associate Professor, McGill School of Information Studies
Title: Applying Data Mining to Real-life Crime Investigation

Data mining has demonstrated a lot of success in many domains, from direct marking to bioinformatics. Yet, limited research has been conducted to leverage the power of data mining in real-life crime investigation. In this presentation, I will discuss two data mining methods for crime investigation. The first method aims at identifying the true author of a given anonymous e-mail. The second method is a subject-based search engine that can help investigators to retrieve criminal information from a large collection of textual documents. If time permits, I will also provide an overivew of the ongoing and future research projects in the Data Mining and Security (DMaS) Lab.

Biography of Speaker:

Dr. Benjamin Fung is an Associate Professor of Information Studies (SIS) at McGill University and a Research Scientist in the National Cyber-Forensics and Training Alliance Canada (NCFTA Canada). He received a Ph.D. degree in computing science from Simon Fraser University in 2007. Dr. Fung collaborates closely with the national defense, law enforcement, and healthcare sectors. He has over 70 refereed publications that span across the prestigious research forums of data mining, privacy protection, cyber forensics, services computing, and building engineering. His data mining works in crime investigation and authorship analysis have been reported by media worldwide. Before pursuing his academic career, he worked at SAP Business Objects for four years. He is a licensed professional engineer in software engineering. Website:

2013/11/25 Graduate Seminar Series Place: MC103
Time: 12 - 12:30
Speaker: Rahul Garg
Affiliation: SOCS, McGill
Area: compilers
Title: Velociraptor: A compiler toolkit for compiling languages like MATLAB and Python/NumPy to target hybrid CPU+GPU systems

Multicore CPUs and accelerators such as general purpose GPUs (GPGPUs) have become mainstream. At the same time, programmers want to use high-level array-based languages such as MATLAB and Python/NumPy. Various language extensions have been proposed for these languages to allow programmers to tap into the power of hybrid CPU/GPU systems. However, language extensions for parallel and hybrid computing require building new compilers or substantially extending existing compilers, and this is a non-trivial task. My toolkit, Velociraptor, is a reusable and portable toolkit that makes it easy to build compilers targeting CPU+GPU systems. I describe the working of my toolkit and will describe how I used it to in two compilers to target multicore CPUs and GPUs. If time permits, I will also show a demo of the compiler generating code for GPUs from Python and/or McVM which is a virtual machine for the MATLAB language developed by other students in my lab.

2013/11/13 Graduate Seminar Series Place: MC103
Time: 12 - 12:30
Speaker: Ladan Mahabadi
Affiliation: McGill SOCS
Title: Music Self-similarity and complexity leveraged for composer classification and computational Turing tests

Music is an information-bearing medium, containing progressions of notes and silences (musical events) ordered in time. Its complexity derives from its information-rich structure, which enables music to be molded to the imagination of any creator, regardless of time or geography. We use the prevalence of patterns and complex structures in music to (1) investigate self-similarity features extracted from musical rhythm, and (2) to create a more secure, user-friendly computational Turing test.

(1) We use similarities hidden in different layers of musical rhythm to construct a concise structural identity (temporal fractal features). This work extends the temporal self-similarity analysis to non-Western music and presents the fractal features in rhythm as a universal feature set. These features are used for longitudinal analysis of compositions in a composer's body of work, and are applied as composer descriptors for classification in two large Western and non-Western music score repertoires. I will present both Western and non-Western composer classification results, which show that fractal features provide complimentary information about the underlying structure that can be used to improve the accuracy of existing classifiers.

(2) The ubiquity of music across all cultures, its complex structure at different layers, and the perceptual characteristics and limitations of the human auditory system are leveraged to construct more accessible, aesthetically pleasing and secure computational Turing tests. Music-based CAPTCHAs, called mCaptchas, are introduced to improve Web accessibility for individuals with visual impairments, to help address and avoid susceptibility to security flaws of existing audio CAPTCHAs, and to improve the overall user experience. I will present empirical evidence of the scheme's security for over 2000 mCaptchas, while its usability is tested by approximately 500 individuals on the Amazon Mechanical Turk (AMT) online market. These results demonstrate that humans can efficiently and accurately solve the generated music-based challenges while sophisticated computer programs fail.

2013/11/06 Graduate Seminar Series Place: MC103
Time: *1:30 - 2*
Speaker: Gayane Petrosyan
Affiliation: McGill, SOCS
Area: Software Engineering/Natural Language Processing
Title: Discovering Information Relevant to API Elements Using Text Classification

The number and the size of Application Programming Interfaces (APIs) are continuously growing. Applications become increasingly dependent on APIs. For example, Java SE 6, which is the core of all Java applications, contains 3774 classes and 203 packages. Programmers, both novice and experienced, are faced to the problem of learning about the vast number of APIs. Given the time at their disposal, it is not possible for programmers to learn all the APIs they need in depth. During this seminar I will describe a technique for discovering relevant sections of an API tutorial to help programmers to find additional related information about API elements they are interested in. The suggested technique automatically analyses API tutorials to assess the usefulness of the information contained therein using Natural Language Processing and Text Classification methods. Afterwards, I will present the results of experiments and the lessons learnt from this work.

2013/10/21 Graduate Seminar Series Place: MC103
Time: 12 - 12:30
Speaker: Gheorghe Comanici
Affiliation: McGill, SOCS, Reasoning and Learning Lab
Title: General MDP feature construction using bisimulation

We consider the process of generating feature-based representations for Markov Decision Processes (MDPs), a powerful mathematical framework for modelling stochastic sequential decision making. We look at a standard approach for learning good policies in reinforcement learning domains, which involves first the computation of a value function that associates states with expected returns. State features are then used to provide approximation schemes for value function computation. In this talk, I will present a systematic way of generating alternative representations by analyzing behavioural similarities on the set of states (i.e. using bisimulation metrics); I will present theoretical guarantees of corresponding approximation schemes; I will describe its relationship to other pragmatic methods, such as spectral clustering and Bellman Error Basis Functions; lastly, I will discuss its computational tractability.

2013/10/18 Graduate Seminar Series Place: *MC320* (not our usual location MC103)
Time: 12 - 12:30
Speaker: Ouais Alsharif
Affiliation: McGill, SOCS, Reasoning and Learning Lab
Title: End-to-End Text Recognition with Hybrid HMM Maxout Models

The problem of detecting and recognizing text in natural scenes has proved to be more challenging than its counterpart in documents, with most of the previous work focusing on a single part of the problem. In this work, we propose new solutions to the character and word recognition problems and then show how to combine these solutions in an end-to-end text-recognition system. We do so by leveraging the recently introduced Maxout networks along with hybrid HMM models that have proven useful for voice recognition. Using these elements, we build a tunable and highly accurate recognition system that beats state-of-the-art results on all the sub-problems for both the ICDAR 2003 and SVT benchmark datasets.

2013/10/09 Graduate Seminar Series Place: MC103
Time: 12 - 12:30
Speaker: Anqi Xu
Affiliation: McGill CIM
Area: Robotics
Title: Adaptive Parameter EXploration (APEX): Adaptation of Robot Autonomy from Human Participation

The problem of Adaptation from Participation (AfP) aims to improve the efficiency of a human-robot team, by adapting the robot’s autonomous behaviors based on input from the human collaborator. We propose a solution, namely the Adaptive Parameter EXploration (APEX) algorithm, that learns from the operator’s intervening commands and dynamically adjusts system parameters of the robot autonomy. We explore this approach within the context of visual navigation, where the human-robot team is tasked to cover and patrol different types of terrain boundaries, such as coastlines and roads. We present empirical evaluations of the APEX solution to AfP, deployed on both an aerial robot within a controlled environment, and a wheeled autonomous vehicle operating within a challenging university campus setting.

YouTube vid link:

2013/08/13 Graduate Seminar Series Place: MC103
Time: 12 - 12:30
Speaker: Jonathan Tremblay
Affiliation: McGill, SOCS
Area: Artificial Intelligence and Digital Games
Title: An Exploration Tool for Predicting Stealthy Behaviour

Stealthy movement is an important part of many games in the First Person Shooter (FPS) and Role Playing Games (RPG) genres. Structuring a game level to match stealth goals, however, is difficult, and can depend on subtle and fragile inter-actions between the game space, enemy motion, and other factors. In this talk we will apply a probabilistic path-finding approach to efficiently analyze a 2D space and find stealthy paths. This approach naturally accommodates variation in the level design, numbers and movements of enemies, fields of view, and player start and goal placement. Our design is integrated directly into the Unity 3D game development framework, allowing for interactive and highly dynamic exploration of how different virtual spaces and enemy configurations affect the potential for stealthly movement by players, or other NPCs.