Teaching a Fridge to See

http://slides.com/dpeskov/deck/live

About Me

  • Work for PwC in D.C.

     ( Analytics Innovation Accelerator)

  • Computer Scientist
  • Have used refrigerator

The Project

Build a fridge system that is:

  • Real-time

  • Global

  • "Smart"

Agenda

  • Machine Vision Overview
  • Implementation
  • Leveraging the Cloud

Why Should You Care?

  • “Computers now better than humans at recognising and sorting images” -The Guardian.  May 2015
  • [Artificial Intelligence is our] “greatest existential threat” - Elon Musk
  •  

Deep Learning for Machine Vision

How to classify this:

Failure:

Images are UNSTRUCTURED DATA

  • Computationally expensive
  • Relevant and irrelevant data combined 

 

What is Deep Learning?

The Stack

The Tools

Implementation

IoT, Data Generation, Model Building

Output

Set-Up

Data Generation

Factors to consider: 

  1. Sample size
  2. Duplication issues
  3. Backgrounds
  4. Number of classes
  5. Lighting
  6. Angles

Model Building

Two Tasks:

Model Approach:

vs

1)

2)

vs

Region Identification

Selective Search

rCNN

Remove
Overlap

Classify Object

One Step!

Leveraging the Cloud

Building an App

  • Flask
  • Creating a python environment
  • Version control with Git
  • Documentation

Cloud GPU

  • 6x faster
  • Cheap
  • Reliable

The Interface

Good

Fast

Cheap

Easy - (er)

Painful

Back up your data!

Work together

Overfitting

Lessons Learned

Thank you!

denis.peskov@pwc.com

Made with Slides.com