Devolearn is a Python package that aims to automate the process of collecting metadata from videos/images of the C. elegans embryo with the help of deep learning models. It is specialized for the analysis of 2-D slices of C. elegans embryogenesis.
This project will focus on improving the performance of existing models and training/benchmarking new models on a broad range of data sets from different species relevant to developmental biology. Your goals will be to improve overall performance of the models in terms of accuracy and generalizability, training/adding new models and improve the library’s usability.
There would be 4 key elements in this project:
- Improving the current models
- Training and adding more useful models
- Improving usability
- (optional) Interactive online demos.
Improving/adding new models
In order to get good performance from a deep-learning model, we should use techniques like:
- k-fold cross validation
- Hyperparameter optimization
- Label smoothing
- Mixed precision inference
- Using more training data for re-training the pre-existing models for better accuracy
As the library grows, we should have a better place to host proper documentation on a static website. We also have to train devolearn’s deep-learning models on commonly available data, the more common the use case, the more helpful it is to the community.
Jupyter notebooks, however simple they may seem, are still a bit intimidating to people from non CS backgrounds, so there should be a focus on making interactive and easy to run online demos which showcase both our research and tutorials.
Proficiency in python with a good hold of tools like numpy, pandas and PyTorch will be required.
More information can be found at: Proposals-Public-Lectures/upgrading-devolearn.md at master · devoworm/Proposals-Public-Lectures · GitHub
Mentors: Bradly Alicea, Mayukh Deb
Tags: DevoWorm, Python, microscopy, deep learning, data science