Jordan shapes for deep learning - ON-147
Preferred Disciplines: Deep Learning, Machine Learning, AI, Computer Science (Masters; PhD or Post-Doc)
Company: ShapeVision Inc.
Project Length: 4-6 months (1 unit)
Desired start date: As soon as possible; ideally no later than September 1, 2018
Location: Ottawa, ON
No. of Positions: 1
Preferences: University of Toronto, including Vector Institute, and University of Montreal, including MILA.
About the Company:
ShapeVision Inc, an Ottawa R&D startup, applies Jordan shape analysis to medical images, computer vision, and photography. Founder & CTO Dr. Martin Brooks describes ShapeVision as an opportunity to leverage diverse image representations. ShapeVision will provide the project with a proprietary web service to transform pixels to shapes.
- List & prioritize all the methods by which neural net training & inference might leverage simultaneous use of two image representations, pixels and Jordan shapes.
- Implement & evaluate at least three of the most promising methods.
- A Jordan shape is represented as a circular sequence of points, in the continuous image plane, defining the shape’s outer perimeter, including also a circular point sequence for each of the shape’s holes.
- ShapeVision software will be used to transform pixel images into hierarchies of Jordan shapes.
- Shape data may be used in conjunction with pixel data to segment and identify objects in the scene.
- Shape data may be used in one or both training and inference.
- In some experiments, shape data may be used in the loss function, or as input, or for explanation.
- Research results will provide an overview and in-depth analysis of how shape data may impact deep learning training, inference and transparency.
- Create and disseminate knowledge of how deep learning training & inference can leverage multiple, diverse data representations.
- In particular, determine if & how Jordan shape representation of 2-dimensional images can best be used with deep learning for instance segmentation and scene understanding.
ShapeVision will manage the project in four phases of full-time intern work. The intern will work closely with Dr. Martin Brooks, who will be available at least 1.5 full days each week. Collaboration will include site visits and videoconference.
- Planning phase: Choose: shape integration strategies for subsequent experiments, datasets to be used in the experiments, and compute platforms for each experiment; and specify data collection, evaluation, and reproducibility requirements for each experiment.
- Experiment phase: For each chosen shape experiment, use three part protocol: (1) hack & train; (2) proof of reproducibility; (3) formal test and data collection.
- Evaluation phase: Summarize data, with discussion of possible conclusions; and, self-assessment of project strengths and weaknesses.
- Dissemination phase: Reproducibility resources made publicly available; intern-written code made available as open source.
Post-project: Academic publication by the intern.
Expertise and Skills Needed:
- Knowledge of deep learning principles, tools, and datasets for computer vision.
- Experience training deep learning on multi-GPU platforms, including pre-trained models.
- Ability to code at all levels of the deep learning stack, including extensions to deep learning models (e.g. new loss function), new network structures (e.g. injecting shape data into a middle layer), and integration with external processes (e.g. request shape image; or, human in the loop).
For more info or to apply to this applied research position, please
- Check your eligibility and find more information about open projects
- Interested students need to get the approval from their supervisor and send their CV along with a link to their supervisor’s university webpage by applying through the webform or directly to Nadia Dubé at ndube(at)mitacs.ca or to Iman Yahyaie at iyahyaie(at)mitacs.ca.