[Feature Proposal] Implementation of State of the Art Active Learning Algorithms

[Vishu26]

Lately, a lot of active learning methods have been developed for deep neural networks. Some of these state of the art methods are considered as a standard benchmark when comparing various active learning methods . However, there doesn't exist a straightforward implementation of these methods. It would be nice to implement these methods within the modAL framework that will make them easily accessible to a lot of developers.

Here are some of the state of the art active learning methods:

• Cost Effective Active Learning - CEAL
• Bayesian Active Learning by Disagreement - BALD
• Core Sets - Core Set
• DeepFool Active Learning method - DFAL
• Batch Bayesian Active Learning by Disagreement - BatchBALD
• Expected Gradient Length

I have implemented CEAL in #146 . I am also planning to implement other methods mentioned above. Let me know what you think.

[cosmic-cortex]

Hi! Yes, it would be extremely awesome if these are implemented in modAL! I have added a review for your PR #146, but I'll talk about your proposal and the future of modAL in detail.

I created this library exactly four years ago, during the holidays of 2017. Since then, I have moved away from the field, but active learning started to gain traction and the SOTA was improved several times. However, modAL was not updated, as I moved on with other projects.

This is going to change from 2022. I have plans to modernize the library, fix some old design issues, and add significant improvements. My hopes are to realize the potential I had in mind when I originally designed the library: a SOTA benchmark and a platform that catalyzes research in the active learning community.

Currently, I am working on formulating a written proposal and building an active community behind the lib. As there are going to be a few changes regarding the modAL interface, I suggest to hold on a bit with the implementations of these. However, your contributions are very much welcome! I'll keep you posted with the updates in this issue, the roadmap regarding the future of the lib is expected to be published this year.