Bucky's Smart Pub

Designed and implemented an active learning siamese neural network recommender system. Applied to Bucky's Smart Pub, where Bucky tries to learn each patrons' beer taste without getting them too drunk

Comment

Please see my full readme at https://github.com/jackhwolf/BadgerHacks2021-BuckysSmartPub

Inspiration

  • siamese NN are interesting to me
  • I wanted to work on a project that involved

What it does

  • Actively selects data points for training a Siamese neural network for making pairwise recommendations
  • Active learning: instead of training on all points in data, start with a subset of the data and add points to the training data based on a scoring heuristic
  • Siamese neural network: NN that takes two inputs instead of one, used for comparing points - e.x. which point does the use like more?
  • Expose algorithm to users through API and website - BuckysSmartPub, using beers.csv from Kaggle
  • API calls are ran on the ML server in the background so site does not freeze and multiple clients can join
  • Clients can step through the learning process and get live recommendations and give feedback

How I built it

  • Pytorch for the siamese neural network
  • Dask for distributing model training
  • numpy

Challenges I ran into

  • running background code through API calls
  • finding hyperparameters for the model which gave good accuracy

Accomplishments that I'm proud of

  • designed and implemented my own ML theory
  • applied it to real world data
  • made a decent website

What I learned

  • how to run API calls in background

What's next for Bucky's Smart Pub

  • Host model on EC2 instance
  • Host site on S3
  • Send model results through SQS

Built With

Share this project:

Updates