Location note: We have changed location to Room 208, McCoy Building, University of Melbourne, 253-283 Elgin St, Carlton VIC 3053 (Google Maps). Call zero four one three seven zero two three three six to be let in.

It’s hard to make systems safe if you can’t understand them. So too for AI. Transparency and Interpretability is the science of figuring out how neural networks work on the inside, with the aim of preventing undesirable or harmful behavior.

During this hackathon, we’ll be exploring a particular branch of interpretability known as “Developmental Interpretability” that is being pioneered here at the University of Melbourne by Dr. Daniel Murfet and collaborators. Developmental Interpretability, or “devinterp”, studies how structure forms in neural networks over the course of training. It asks what we can learn about how ML systems work by detecting, locating, and investigating phase transitions during the learning process.

Whether you're a seasoned data scientist, a coding enthusiast, or simply curious about the inner workings of AI, this hackathon is for you! You’ll have a chance to explore a variety of systems and techniques (and suggest your own). No prior knowledge of developmental interpretability is required, though experience with Python and PyTorch is recommended. The event will include an introduction to developmental interpretability, a list of open problems that you could explore, and demo notebooks to help get you started. Groups will be formed on the day at the event, but you are welcome to work solo or form groups ahead of time. The event is primarily intended to be in-person, but there will be an online option available.

Location:
Room 208, McCoy Building, University of Melbourne, 253-283 Elgin St, Carlton VIC 3053 (Google Maps)

Schedule:

2 day event - Saturday 7th October and Sunday 8th October, 10am to 5pm each day. Lunch will be provided.

Day 1: Saturday, October 7th

10:00am - 10:30am: Opening Talk
10:30am - 12:30pm: Form groups and get started
12:30pm - 1:30pm: Lunch (provided)
1:30pm - 4:30pm: Hacking
4:30pm - 5:00pm: End-of-Day Presentations

Day 2: Sunday, October 8th

10:00am - 10:15am: Day 2 Kickoff
10:15am - 12:30pm: Hacking
12:30pm - 1:30pm: Lunch (provided)
1:30pm - 4:00pm: Hacking
4:00pm - 5:00pm: Final Presentations and Closing

Brought to you by:

- Responsible AI Development (RAID)

- Timaeus

- AI Safety Melbourne (AISM)

If you have any questions, please feel free to contact us at ai.safety.melbourne@gmail.com

To hear about future AI Safety Melbourne events, join our mailing list.

Contact the host of Hackathon - Developmental Interpretability

There were 0 errors found in the information you submitted: