Multi-Context Multi-Domain Spatio-Temporal Disambiguation from Meta-Data, Audio and Video Stream inputs as Time-Series - ON-233
Preferred Disciplines: Computer science, Computer engineering, Electrical engineering (PhD student or Post-Doc)
Company: Caliber Communications Inc
Project Length: 4-6 months
Desired start date: As soon as possible
Location: Hamilton, ON
No. of Positions: 1
About the Company:
Caliber Communications Inc., Canada’s leading tech-innovator that has engineered and developed a unique communications platform which operates on the cellular backbone. With its core technology, Caliber aims to change the way people view security utilizing their patented communication technology offering the world’s first modular security system. This platform is specifically utilized for the application of live remote security monitoring of sites across a variety of industries.
Caliber’s Insurance approved technologies are able to perform 24/7 – answering the sharp increase in private security demand, while dramatically lowering the costs. With highly engineered, accurate monitoring technologies, they provide the coverage, reliability and versatility that security guards just cannot.
Since its inception in 2014, Caliber has embarked on an ambitious set of projects to allow automated assistance and monitoring capabilities in a single platform, SyncroReports. As part of this system, alerts including emails and SMS messages are sent exactly when their computer vision systems match events per specified parameters. In the past year, Caliber commenced a comprehensive computer vision research project, which leveraged machine learning techniques to detect presence of specific events within live video streams, so as to assist monitoring staff using prompts.
In 2019, Caliber aims to further refine this computer vision technology through a research project aimed at improving multi-context multi-domain disambiguation after performing spatio-temporal segmentation of events and objects within video frames, as well as interactions between objects – akin to multi-attention systems. Using this approach, Caliber aims to test the improved multi-context multi-domain model against a live person’s decision-making, so as to eventually make this a reinforcement learning problem for the investigation in future research projects.
- Extend the current research (via both: Caliber’s internal proprietary knowledge base, as well as publicly available literature) by creating a multi-mode meta-model and associated policy hierarchies around identified domains (e.g. audio and video streams, and meta-data) to generate actionable recommendations for a user (similar to prescriptive analytics).
- Create a framework that is extensible for testing the improved multi-mode meta-model and policy hierarchies against a live person’s decision-making, so as to set the ground-work for future research towards reinforcement learning projects.
To be discussed
Expertise and Skills Needed:
- Masters or Ph.D. in Computer Science (preferably) or Electrical/Computer/Software Engineering
- Extensive Experience in Machine Learning, Deep Learning, especially multi-attention scenarios
- Working experience in (or sound knowledge of) DevOps, containerization and orchestration
- Knowledge of applied machine learning is a plus, when it comes to computer vision applications, including OpenCV, TensorFlow Object Detection API, YOLO, R-CNNs and SSDs, etc.
- A strong quantitative background (ideally in the field of Applied Mathematics) or experience with applied machine learning for computer vision is a definite plus.
For more info or to apply to this applied research position, please