The machine's suggestion
Training the machine

Design Document for CalHacks 4.0

Group Members

Hao Le haole@berkeley.edu
Elva Chen elvachenxy@berkeley.edu
Samy Raman sraman@berkeley.edu
Claire Li shengxiao.claire.li@berkeley.edu

The Story

We have a problem. We don't know where to eat next and we want to develop a unique solution to this problem. Thus we decided on a fully image-based (as opposed to search based) recommendation web app that uses image recognition and machine learning to classify photos of food and learn what the user likes. No need for typing or searching. Just click and go. We no longer have a problem.

Description

Use machine learning to predict what restaurants the user would like to try in a new area based off of:
- Images the user selects - We use preclassified images to lower processing load on Microsoft
- From these labels form a model and import images from random restaurants
- Use Microsoft customAI image recognition to label the images
- Choose a restaurant based on label/model score with Synpatic JS ML
- Improve model with user feedback
- Store user models in MongoDB through MLAB
Front End User-Flow
- 3x3 grids of photos for preference building and model initiation
  - Users can keep picking photos
  - This data trains our neural network
  - Users can indicate diet preferences (vegetarian/vegan) (allergies optional)
  - Users can return to this at anytime to update their preferences
- Machine then gives suggestions
  - Show multiple photos
  - Show Name
  - Show location on map
  - Show Reviews from (GoogleAPI - net yet implemented)
  - Possible Responses:
  - Go there! (validates machine)
  - Maybe another time (validates machine)
  - Nah (invalidates machine)
(optional) Restaurants can be saved
- User can create custom lists. For example: "dessert/authentic/etc"
(optional) Can predict for multiple people
- User can add friends and the app will use their data as well for group decisions