The University of Waterloo ECE Machine Learning Lab (UWECEML) carries out work on a variety of topics within Artificial Intelligence and Machine Learning with a focus on using real world problems to discover computationally hard questios for modelling uncertainty, learning predictive models and enabling decision making. Below are some of the topics, domains and concepts of ongoing and prior work in the lab.
Convolutional Neural Networks (CNNs) are a type of multi-stage, deep neural network which use alternating layers of convolutions, to correlate nearby points, and pooling, to reduce the data size to encourage abstraction. We are exploring the ability of CNNs to perform more powerful reasoning in spatiotemporal domains like forest fire management and brain imaging.
Anomaly detection is the problem identifying within a given dataset a collect of events or datapoints that lie outside the “normal” behaviour for the system. This is similar to outlier detection in regression analysis and can be thought of as a two class, unsupervised clustering problem where one class contains normal datapoints and the other contains the anomalies. We may know something about the anomalies ahead of time or, more often, we will need to learn to identify them by their rarity or their dissimilarity with the vast majority of normal datapoints. In the driving domain the most important anomalies are unsafe driving situations.
In computational sustainability problems in ecology, climate science, migration modelling and forest management an important class of anomalous situations are catastrophic changes to the system. These could be very large forest fires which reach a destructively high temperature; or they could be a widespread increase in mortality or fertility in plants or fish; or it could be an average yearly temperature beyond an historical threshold. In all these cases the catastrophic event could be recognized as an anomaly when looking back on the data, but looking forward we would need to predict the imminent change in class or state that results in the anomaly. This problem of predictive clustering and catastrophe detection/prediction is one of the longer terms goals of my research in several areas.
Optimization and Decision Making Under Uncertainty
This includes work on Reinforcement Learning using direct policy search as well as exact solution to Markov Decision Processes with confidence bounds particularly in complex domains especially when there are vary large state and action spaces. A wide range of optimization algorithms have been developed that are each appropriate in different types of problems. Equilibrium Policy Gradients utilize a parametrized policy defined as the equilibrium distribution of a Markov chain built from local causal policies. We have used this algorithm to deal with the complexity caused by spatial structure common in computational sustainability domains. There is also work on MDP planning algorithms with provable confidence intervals upon completion and other Reinforcement Learning approaches.
Some of out work looks at modelling, inference and learning of interpretable probabilistic graphical models. Of particular interest are cases where the underlying data is relational, such as social networks, or there is a need for both causal and correlational structure, cyclic structure in other words, within the same model. This arises naturally when trying to represent spatial policies using graphical models but cyclic structure can arise in many relational domains and existing fully directed or fully undirected models have trouble dealing with it.
The UWECEML lab carries out fundamental AI/ML research into the above topics but most projects are centred around a concrete, applied domain or dataset that drives the needs to algorithmic and theoretical advances. These projects range widely across automotive, medical, physics, environmental and IT systems domains.
We have worked with car companies such as DENSO and Magna International on various projects relating to analysis of data collected in modern vehicles up to an including how to control cars in complex environments.
Driver Behaviour Learning - We are working with Magna International to collect high resoltion, multi-modal data from actual drivers on fixed scenarios. The resulting data will be used to learn predictive models of driver behaviour and anomalous situations. These models can then be used to enable the next generation of smarter Adaptive Driver Assistance Systems (ADAS) such as intelligent cruise-control and eventually autonomous driving. The research questions are how to build more robust, safety aware predictive models and how to model human driver style to make ADAS more user friendly.
Related Projects: DBL Project, CapsuleNetsForLIDAR
People: Dan Chung, Sahil Perriera
Multi-Agent Collaborative Control - There are various algorithms ongoing in the lab focussed on learning collaborative multi-agent policies for driving cars autonomously in simulated environments. We focus on software simulators and physical treadmills with remote controlled cars for these experiments. We have are developing Multi Agent Reinforcement Learning (MARL) algorithms which learn inter-agent communication protocols automatically and which then optimize policies to act in environments where communication is noise or sometimes impossible.
People: Sushrut Bhalla, Sriram Subramanian
Bounding Box reconstruction from LIDAR data Using Capsule Networks - LIDAR sensors have a very high accuracy within their limited range but for general use they need to be better fused with other sensor modalities and contextual information. The research goal here is to utilize the new ideas from capsule based deep neural networks to improve the state of the art in object detection in LIDAR and LIDAR+visual data from the automotive domain.
This project is related to the Magna Driver Behaviour Learning project but works as a stand-alone contribution.
Related Projects: DBL Project
People: Dan Chung
Alzheimer’s Classification - learning predictive classification models for diffusion MRI data to provide decision support for degenerative brain diseases using Deep Neural Network methods currently only used for 2D image classification. This domain is challenging due to the 3D structure of the data as well as the non-visual properties which do not necessarily carry over from other domains.
People : Laura McCrackin
Physics and Chemistry
Combustion Modelling - Our lab is looking at a few approaches for improving modelling of chemical interactions and building more scalable models of combustion.
People: Sushrut Bhalla
Collaborator: Jean-Pierre Hickey
One major guiding inspiration for some of the theoretical directions in the lab are problems from the growing field of Computational Sustainability which applies AI/ML methods to data analysis and optimization problems in sustainability related domains. These domains include ecology, agriculture, water systems, energy systems and climate change.
Forest Management - Prof. Crowley’s early research and the UWECEML lab’s ongoing research cover a range of problems in decision making and prediction for Forest Management. This includes algorithms for :
- optimization of sustainable harvest policies under spatial constraints using Reinforcement Learning and Policy Gradient Search.
- learning dynamics of forest fire spread from simulations and satellite data using Deep Reinforcement Learning
- generation of future fire spread scenarios from image series using Long-Term Recurrent Convolutional Neural Networks
- modelling of decision making about treatment against forest wildfires using Markov decision processes
People: Pardis Zohouri, Sriram Subramanian
Invasive Species - treatment of plant species along river banks to inhibit spread of invasive plants using Markov decision processes. The goal was to provide exact solutions in an inherently exponential action space. To balance these needs we developed a novel PAC MDP algorithm called DDV which reduced the sample complexity needed for confidence interval guarantees.
People: Mark Crowley
Fundamental AI/ML Problems in Large Scale and Streaming Data
Anomaly Detection in Embedded Systems - anomaly detection in large-scale streaming datasets from automotive systems and IT system traces. This problem is challenging for a number of reasons. It is an unsupervised problem, so anomalies are defined by the infrequency or different behaviour relative to other data. Also, as the data logs arrive so quickly they cannot be saved for later processing. Thus, an anomaly detection (AD) algorithm needs to make new predictions while simultaneously learning normal and abnormal patterns from incoming data.
We are developing algorithms for continuous learning of models for anomaly detection in the presence of large scale, streaming data using modifications of existing, diverse ideas from ensemble methods such as Random Forests, Isolation Forests and others.
People: Maria Samad
Data Reduction - Despite the reducing costs of data storage, given the increasing amounts of data being generated it is inevitable that often only subsets of available data can be used for offline learning or deployed AI/ML systems. This is obviously true in mobile devices, remote autonomous systems such as space probes but also many other domains. Our lab is developing a number of algorithms looking at the fundamental issues of data reduction in terms of dimensionality and numerosity. While many classic algorithms have addressed this, the dawn of Deep Learning do not solve all of the problems of expanding data automatically, so further theoretical and applied advances of faster methods will be highly impactful.
People: Benyamin Ghojogh
Machine Learning on Encrypted Communications - This project focussed on Network Communication Security Breach Detection. The goal was analyzing network meta-data to classify anomalous patterns of usage which may indicate security breaches in a communication network.
Prof. Crowley and PhD student Jennifer Fernick presented the results of this project at RSA 2017 in San Francisco.
Some of the related concepts of interest to the lab's research.
Spatially Spreading Processes
Learning, modelling, predicting spatially spreading processes in multiple diverse domains such as Forest Wildfire management, Invasive Species management and Brain Imaging.
The field of Computational Sustainability is at the intersection between computational sciences (such as artificial intelligence, computational modelling, optimization and planning research) with applied research in environmental/ecological domains (such as land use management, invasive species spread, sustainable ecology management, smart grids and species tracking)
In any system with rich states which can change and interact over time there is the possibility that small change in dynamics, initial state or action policy could lead to large changes in the type or value of the outcome. This property is the essence of a complex system and there is a wide variety of research ongoing in different fields in such as social science, engineering, ecology, physics, economics and computer science to understand complex systems and develop ways to bring some order to chaos. Some members of our lab are affiliated with the Waterloo Institute for Complexity and Innovation (WICI) which helps to bring researchers together on these topics.