Additional Details and FAQ

Who Should Attend:

This hackathon is for anyone in the machine learning world!

  • Students:

    • Masters, PhD, or Undergrad

    • Computer Science, Machine Learning, Robotics, Bioinformatics, Data Science, anywhere that PyTorch is used

  • Industry Professionals:

    • Startups that want to to get that MVP good enough for market

    • Software Companies aiming to deliver better solutions to their customers

    • Companies running models on the cloud hoping to save money on GPU inference

Judgement Criteria:

Projects will be judged based on how much optimization they are able to achieve from their initial training pipeline. The following is the exact criteria in order of weight:

  1. The model that is improved

    • Higher scores for published papers by number of citations, lower scores for simple custom models

  2. The Improvement

    • Accuracy increases and compression are both important factors.  Both will be scored in terms of percentage improvements in comparison to the percentage improvements other teams are able to achieve.

  3. Amount of customization required

    • Lower scores for applying working examples to new datasets

    • Higher scores for working with new models which require no customization for optimization.

    • Highest scores for working with new models which do require creation of customized tensor processing modules for optimization.

  4. Report and Presentation

    1. Delivering code for your projects is optional. Instead you will submit a write-up with content similar to our first testimonial and, if selected, give a presentation in March.

What to Bring:

Unlike a traditional hackathon, for this hackathon you can select your project in advance. Come to the first day with your PyTorch pipeline built that is already learning properly on your dataset. This hackathon is about seeing how much you can improve your initial model with Perforated Backpropagation™.

Notes:

  • This can be any project you want - Academic research, commercial product, or a hobby project. Feel free to even pick another Kaggle hackathon to compete in and use this compeition to top the scoreboard.

  • Optimization is easier if you have quick access to the entire training pipeline - i.e. not using a trainer framework like PyTorch Lightning or Huggingface.

  • The one downside of the optimizer is training time. You should choose a system that takes less than 48 hours to train from start to finish.

FAQ

  • How’s IP work?

    • Anything you build during this hackathon is yours to keep. Take that MVP and go get customers, publish a paper with the new state-of-the-art results you achieve, or put your new model into production. We just want to tell your story on our testimonials page.

  • Is my proprietary data and code safe?

    • Yes. Perforated Backpropagation™ is installed with pip just like PyTorch, NumPy, or any other package. It runs locally on your own computer and does not require internet access at all to ensure no data leaks for our users. You may find additional details in our privacy policy.

  • The algorithm sounds interesting, do you have a more detailed technical document?

    • Yes, here is a link to a preprint of the paper we have written.

  • Can I still participate if those days don’t work for me?

    • Yes, the hacking day will be recorded and can be viewed at any time. If you can’t make the presentation day it may impact the scoring, but you will be allowed to submit a video rather than a live presentation.

  • Do I get a hoodie for any submission?

    • To receive a hoodie a submission must be a successful experiment. This is defined as a 5% reduction in loss metric or 10% compression without negative impact on loss.

What we'll Provide

  • Python package to use Perforated Backpropagation

  • Hands on Implimentaiton instruction

  • Cash Prizes

  • Perforated AI Winter Beanie for all participants

  • Food and Drinks for both days

  • Optional Authorship on the Hackathon Paper

Hackathon Partners