Deep Learning Researcher for Video Applications
Braincreators / Amsterdam (NL)Apply on site
The BrainMatter platform currently works on large data sets of images, predicting bounding boxes and segmentation masks for industrial asset inspection applications. Video data has a big potential for our clients and their use cases. However, the step from images to video has ramifications on all aspects of our platform.
- First, there is the representation learning itself. The temporal aspect of video and the recognition of actions and dynamic object states require different DL techniques. There is the issue of annotations and supervision . Since annotation for video will typically be more expensive than for images, we are looking at state-of-the-art self-supervised representation learning techniques.
- There are infra-structural questions. How much can be done on a certain edge client, and how much pre-training needs to be done on heavy duty machines in the cloud? Privacy concerns may require us to do more on the edge than in the cloud. There is space for optimization here that will need to be exploited if we want fast and reliable applications.
- Finally, there are issues of hashing, indexing, and retrieving very large amounts of video samples via their vector representations. Database optimization, the ability to distribute data over a cluster of servers, and handling many streams of videos to be embedded and compared all make this extremely challenging.
Within BrainCreators, the Science Team is tasked with selecting, reproducing, and prototyping relevant academic work in order to prepare it for integration into our platform and productization by our other teams. The current work on video representation learning at BrainCreators is in the hands of the Science Team, due to the experimental nature of the work and the necessary literature research needed to make the right choices in this early stage.
- Selecting relevant academic research papers, evaluating them in terms of code quality and ease of integration
- Setting up and implementing experiments to evaluate the effectiveness of the methods on both academic and industrial data sets
- Combining and integrating methods that achieve better results in practice than using each method separately
- Reporting end results in written form, documented code, and presentations in our internal weekly workshop
- Helping the rest of the Science Team to find and understand new ideas that might be relevant for both your and all our other projects (e.g, our work on 3D Point Cloud, Learning under Distributional Shift, and on Neural-Symbolic Hybrid Systems)
Experience and skills
- You are well versed in ML/DL technology, know how to implement complex pipelines, and know how to ask the right questions to get projects to the next level.
- Preferably, you have at least a completed Master degree in Artificial intelligence, or similar field, or can prove to us you have gained equivalent experience from previous projects.
- You are a capable PyTorch programmer, familiar with Git technology, and can develop non-trivial ML experiments.
- Your control of the English language is good enough to read, understand and explain scientific papers from the field, as well as present them to the team and to our clients.
- You are available early in 2021.