In automatic casting applications, the aim is to accurately recognise facial regions that correspond to a same actor appearing in a movie to produce described video. In particular, this project will focus on challenging tasks of capturing and modeling the facial trajectory for each person appearing in a movie in order to predict when/where the principal actors appear. This is a challenging task because recent movies are typically high quality and faces are often occluded and their appearance varies significantly according to pose, illumination, blur, etc.
In this project, we will apply machine learning to perform image style classification. We will build a system that uses image style classification to increase user engagement in an eCommerce platform setting. We will study the effects of user preferences for particular image styles on their engagement with the platform.
Image style classification is the task of categorizing an image based on attributes such as composition style (e.g., minimal, geometric, etc.), atmosphere (hazy, sunny), or colour (pastel, bright).
The primary goal of this project is to explore a variety of new and existing Natural Language Processing (NLP) techniques to improve the performance, and further the automation of, Knoteâs text analysis software â specifically with entity recognition. Entity recognition is the process of identifying all groupings of words in a collection of documents that fall within that entityâs purview, such as proper names or chemical compounds.
This project proposes to explore and implement a method of storing and retrieving data relating to genetic variation across a population of individuals. Due to the large amount of genetic information each person possesses, such a database requires special attention to minimize the amount of data stored and to create efficient methods of accessing the data. This work will research and test different strategies to build a compact data store that will return results quickly. This data store will be incorporated into the PhenoTips software provided by Gene42 Inc.
Cellphones get notifications from different companies every day, but we do not know whether these notifications have a significant impact on customersâ behaviour. Knowing the impact of these notifications would provide useful insights to marketing strategists. Since user behaviour will determine the efficacy of push notifications, this project initially aims to build a behavioural model, which will group customers based on their web site navigation behaviour.
UBC and Microsoft intend to collaborate on an applied research project where 3D models of the brain will be used to create an interactive Holographic lecture using Microsoftâs new augmented reality device, the HoloLens. The will form the basis for a lesson or âHoloLecture,â and will feature new interactions to take advantage of the HoloLensâs technology. The ability to manipulate the 3D objects and dynamically adapt them to a live lecture format will form the basis of a HoloLecture prototype that can be applied across disciplines.
Kobo is an online e-book retailer that provides recommendations for future purchases to its user base. One difficulty that recommendation systems face is what is known as the âcold-userâ problem. In this scenario, when we know so little of a userâs preferences (for example, if they are new to the platform), we do not have any basis for recommendations. The goal of this project is to develop an interactive application that can elicit such preferences from users about whom we have little information, and that can help improve recommendations for power users.
Information published by financial news agencies is used as one of the inputs to make investment decisions. News articles from multiple sources can be used to gauge market sentiment towards an industry or a specific company. Deep learning techniques have been successful in producing state of the art results on various benchmark datasets (Dai & Le, 2015; Miyato et al., 2016). Most of the popular algorithms extract features from words, sentences or paragraphs and represent them as fixed-length vectors (Mikolov et al., 2013; Le & Mikolov, 2014).
This project will develop and apply machine learning techniques to predict the valuation of the properties in Nova Scotia. The techniques will help Property Valuation Services Corporation (PVSC) assessors with more efficiently and accurately valuing properties. The ultimate goal is to help PVSC reduce the number of annual appeals â which is a costly undertaking. It will also reduce the need to send assessors directly to the property locations, instead they will use machine learning techniques to more accurately predict property values.
Representations are fundamental to Artificial Intelligence. Typically, the performance of a learning system depends on its data representation. These data representations are usually hand-engineered based on some prior domain knowledge regarding the task. More recently, the trend is to learn these representations through deep neural networks as these can produce significant performance improvements over hand-engineered data representations. Learning representations reduces the human labour involved in any system design, and this allows in scaling of a learning system for difficult problems.