Mechanistic interpretability for interspecies communication
- commitment
- 10 hrs/week
- flexibility
- No mentee proposal
mentored by
Siddharth PuttaBoston University
The project
There is a working project on using mechanistic interpretability tools on bioacoustic encoder models to understand encoded structure. This will be a continuation + branched extension of the project.
- Continuing interpretability pipeline on bioacoustic encoder models
- Training SAEs with acoustic-based feature vocabulary
- Reproducing and quantifying linear probe methodology
- Extension project on interpretability pipeline for Variational Autoencoder + Vision transformer model batch
- Still idea stage, but plans to use interpretability techniques on models that compress and reconstruct vocalizations - cross-comparing between species.
The mentee's role
- ML Engineering
- Research analysis(interpretation + bioacoustic domain analysis)
- Writing, discussing, and project-scoping.
It will be a very co-work based project.
Who I'm looking for
Must-haves:
- Actual interest in the project
- Good work ethic
Nice-to-haves:
- Comfortable with machine learning, python, and coding agents
- Background/knowledge in linear algebra and interpretability
- Bioacoustics domain knowledge
Questions for applicants
Answer EITHER of these:
- Tell me about a time where you worked on something interdisciplinary, what was challenging?
- Tell me about a time where you had to learn/accumulate knowledge about a topic very fast - how did you do it?
Support offered
- I will be very hands-on and collaborative with this project, so my support will mainly be related to the research at hand and technical work.
- I can also offer connections to other mentors/advisors that will help with this project.
- I will both help with shaping direction and drafting, as well as contributing to the direct output.
- format
- Research & writing · Project scoping
- topic
- Other