Project Highlights -
Fusion and Inference in Surveillance Networks

  1. Quantum feature learning: We proposed new algorithms using D-Wave's chimera network architecture to learn features from data in an unsupervised way. The work was presented at the CIFAR 2011 summer school on Neural Computation and Adaptive Perception and was recently accepted for publication at the NIPS 2011 Workshop on Deep Learning and Unsupervised Feature Learning. The results obtained thus far demonstrate that D-Wave's chip can indeed be used to carry out feature learning in images. However, there are still hardware calibration difficulties that need to be surmounted. Ultimately the goal is to prove that D-Wave's system enables us to learn features more efficiently than with classically computers.
  2. Adaptive Monte Carlo: We were able to capitalize on our earlier work on Bayesian optimization [T3] to develop new adaptive algorithms for efficient Markov chain Monte Carlo inference. One algorithm solves the hard problem of tuning the hyper-parameters of hybrid Monte Carlo and will be presented at the NIPS 2011 Workshop on Deep Learning and Unsupervised Feature Learning. Two other algorithms are tailored to discrete probabilistic graphical models with and without constraints, including the chimera model of D-Wave Systems[T4,T5]. The proposed adaptive MCMC algorithms significantly outperform existing techniques, including the popular Swendsen-Wang algorithm.
  3. Bayesian Optimization: We developed hedging strategies to automatically fuse the information from many utility models in Bayesian optimization [C44]. The approach has resulted in a very significant practical improvement on performance over existing approaches. We also made inroads in proving stronger theoretical convergence rates for Bayesian optimization with deterministic observations. This work will be presented at the NIPS 2011 Workshop on Bayesian Optimization, Experimental Design and Bandits, co-chaired by Nando de Freitas. Finally, our Bayesian optimization techniques have been applied to adapt particle filters over time so as to improve object tracking in surveillance video [C45].

  1. Distributed Particle Filtering: We proposed a new algorithm for efficient distributed particle filtering [C39]. The conference paper was runner up for Best Paper Award at the ISIF 2010 International Conference on Fusion.
  2. Multi-scale Gossip: We proposed a new algorithm for fast decentralized averaging based on peforming gossip at multiple scales in a network and analyzed the convergence rates [C38]. The conference paper won the Best Paper Award (Signal Processing and Information Theory Track) at the 2010 IEEE International Conference on Distributed Computing in Sensor Systems.
  3. Accelerated Distributed Averaging: We created an approach to accelerate distributed averaging through the use of a very small amount of local node memory [J11].
  4. Selective Gossip and Greedy Gossip with Eavesdropping (GGE):We published a paper describing and theoretically characterizing a greedy gossip algorithm that for practical network architectures has a significantly reduced convergence time compared to previous gossip techniques [J12]. We also identified a selective gossip algorithm that facilitates distributed non-linear approximation; we proved that the algorithm converged and demonstrated that it reduces communication compared to alternative methods [C37]. We participated in the development of a review article of gossiping techniques [J13].
  5. Delay-Tolerant Particle Filtering: We developed a new procedure for tracking targets when some measurements are delayed due to the wireless network [C40]. The procedure estimates the informativity of delayed measurements and only processes those that have the potential to significantly change the filtering distribution.
  6. Particle filtering and Smoothing: We have made several advances in developing novel Sequential Monte Carlo methods, including smoothing procedures [J10], analysis of diffusion processes [J9], and techniques for nonlinear target tracking [C41].

  1. Distributed Particle Filtering: We have conducted a theoretical analysis of the stability of the leader node particle filter [C22]. The conference paper was one of three nominations for Best Student Paper Award at the ISIF 2008 International Conference on Fusion. In [C35, JS1] we have generalized this analysis to address any particle filter that involves an approximation step and improved the bounds on performance.
  2. Accelerated Consensus: We have created an approach to accelerate local, linear iterative network algorithms that asymptotically achieve distributed average consensus [J7,C18]. We have provided a theoretical analysis that characterizes the improvement and examines how it scales with network size [C34, J11].
  3. Greedy Gossip with Eavesdropping (GGE): We have developed a new algorithm, GGE, for reaching consensus through local gossip-type communication in a wireless surveillance network [C20]. We have proved that GGE outperforms standard randomized gossip in its speed of convergence [C24, C33, J11].
  4. Controlling the Sensors – Stochastic Control Approaches: In [C8], we made the crucial observation that the stochastic control problem can be reinterpreted as one of trans-dimensional inference. With this new understanding, we were able to propose a novel and very efficient reversible jump Markov chain Monte Carlo (MCMC) algorithm. In [C35, N2], we were able to improve these MCMC algorithms to attack formidably complex control and active sensing problems.
  5. Camera Management: A very important issue in the management of cameras is where to look in a scene in order to infer a good model of the location of targets. We have addressed this problem in [C9, C19], developing a sequential decision-making algorithm for gaze planning.
  6. Tracking Multiple Targets: In [J4], we proposed a multiple-model implementation of the probability hypothesis density (PHD) filter. In [J3], we addressed the issue of multiple spawning targets and presented an algorithm for measurement-to-track association.
  7. Model Uncertainty: In [J5], we address tracking problems where there is only partial knowledge of the model structure and the associated parameters. We propose an EM-type algorithm that casts the problem in a joint state estimation and model parameter identification framework.
  8. Motion and Camera Models: We have focused on the development of a meaningful metric for assessing the “closeness” of two data points; this type of metric is critical for classification and object association between successive video frames (or images from multiple cameras). We identified an algorithm for learning the most appropriate metric from a set of data [C25, C26, C27].

  1. A new calibration method for "bootstrapping" a particle filter to learn the camera observation model while tracking objects [C1], and a minimum-volume ellipsoid metric-based approach for learning local inference models [C6];
  2. A new mathematical model, based on signal detection principles, for multiple object tracking in images with low appearance information [T1];
  3. A probabilistic consensus method to enable local learning in networks, improve network robustness and decrease communication costs [C4,J12], and a characterization of the convergence rate and the variance of the error after consensus is achieved [N1];
  4. An anomaly detection approach using sparse on-line kernel density estimation that has been deployed to detect traffic jams in sequences of camera images of highways [C2];
  5. New algorithms for dynamic sensing using active exploration techniques and nonparametric statistics [C3];
  6. Mathematical models and stochastic approximation algorithms for optimal observer trajectory planning in partially observed environments with receding horizon control [J1] and theoretical results on the stability of these methods;
  7. A reduction method for mapping the infinite horizon stochastic control problem to one of inference with trans-dimensional distributions [C7,C8], and algorithms that exploit this mapping to carry out efficient control in large continuous spaces;
  8. The dynamic sensing algorithm developed in [C3] has been applied to active interfaces [C5] and recently was the winner of the SRC competition at the most prestigious computer graphics venue: ACM SIGGRAPH.

Event Highlights

  1. MITACS Seminar Series on Decentralized Processing, Sensor Networks and Particle Filtering: To date, fifteen leading researchers from industry and academia have presented seminars on a diverse range of topics.
  1. The MITACS Workshop on Fusion, Mining, and Security for Networks was held at McGill during June 16-20, 2008. The workshop was a huge success, attracting over 80 academics, industrial researchers, graduate students, and postdoctoral fellows from Australia, England, Switzerland, the United States, and Canada. The workshop opened with two days of tutorials on network analysis and security presented by international experts. The remainder of the week was dedicated to conference-style research presentations. In addition to academics, the workshop included a host of industrial and government participants including representatives from the Communications Security Establishment (Ottawa), Defence R&D Canada - Valcartier (Quebec), Lockheed Martin Canada (Montreal), Solana Networks (Ottawa), Swisscom (Switzerland), and the Telecom Applications Research Alliance (Nova Scotia). MITACS pledged $40k of support which covered the bulk of the workshop expenses, including travel funds for many participants. This award was supplemented by the Centre for Advanced Systems and Technologies in Communications (SYTACom), the Fields Institute (through their National Institute for Complex Data Structures), the Institute for Pure and Applied Mathematics (a U.S. research institute sponsored by the NSF and based at the University of California - Los Angeles), and MASCOS, the Australian Research Council's Centre of Excellence for Mathematics and Statistics of Complex Systems.
  2. Graduate Course on Sequential Monte Carlo Methods: Throughout the Fall Semester, 2008, Arnaud Doucet has been conducting a graduate course on sequential Monte Carlo methods. The graduate course is taking place at the Statistical and Applied Mathematical Sciences Institute (SAMSI) in North Carolina.
  3. MITACS Research Workshop: In March 2009, we organized a workshop at the McGill Research Facility in Barbados. The workshop was attended by four of the project investigators (Coates, de Freitas, Ferrie, Rabbat), three graduate students and a postdoctoral fellow. In addition, Prof. Richard Baraniuk (Rice University), Prof. Rui Castro (Columbia University), and Prof. Mark Crovella (Boston University) attended as external experts. The week-long workshop involved tutorial and research presentations as well as intensive discussion. The final two days were devoted to team-based problem-solving with a view to developing collaborative research activities.
  4. NATO Symposium (organized by Éloi Bossé, DRDC): Michael Rabbat attended the NATO Symposium on C3I for Crisis, Emergency and Consequence Management in Bucharest in May 2009. He presented a summary of the research performed under the umbrella of our MITACS project. Éloi Bossé, one of our primary collaborators at DRDC, organized the symposium, which was attended by representatives from NATO countries.

  1. Montreal Industrial Problem Solving Workshop (organized by Odile Marcotte, Centre de Recerche Mathematique, University of Montreal): Mark Coates led a team including Nando de Freitas and Boris Oreshkin in a week-long exercise in developing novel mathematical techniques to address a problem facing one of our industrial partners, Lockheed Martin Canada.
  2. Seminars on Reinforcement Learning and Markov Chain Monte Carlo (MCMC): Nando de Freitas (UBC) has presented a series of tutorial seminars at McGill University addressing the application of MCMC techniques in machine learning.
  3. MITACS Workshop on Data/Information Fusion in Canadian Academia (organized by our industrial partner's representative - Eliza Shahbazian, Research Director, Lockheed Martin Canada): Mark Coates and T. Kirubarajan presented seminars at the workshop, which attracted 35 attendees including 15 students.