Lab 3. Deep Learning Lab (25 pts)

OIDD 2550, Tambe

Introduction

This lab proceeds in two parts. It first asks you to make use of Google’s Teachable Machine to investigate the capabilities and tradeoffs of deep learning algorithms, and to further understand how data fits into the machine learning pipeline. It is recommended that you use the Chrome Browser for these exercises, as things seem to work a little better there. The second part of the lab asks you to use a Python Notebook to further explore some of the choices behind deep learning and explainable models.

Deliverables

For this assignment, you are encouraged to discuss with others, but please complete the assignment and submit on your own. Submit

Assignments are to be submitted through Canvas. See Canvas for the due date.

Part 1. Identifying pneumonia in chest x-ray data (6 pts)

Question 1A

Use this data source along with Teachable Machine to train a deep learning engine that can identify pneumonia in x-ray data. The data source contains training images for normal and diseased lungs as well as test images for normal and diseased lungs. Note that the data source is a large one, and may take a while to download. We have already cut the available x-ray data down but you may want to cut it down further depending on the capabilities of your computer. Try to work with the largest set of images that your laptop will allow for the model training. For this assignment, submit a screenshot of classifier output for a diseased x-ray and another screenshot for a normal x-ray. (3 pts)

Question 1B

Briefly describe (in no more than a few sentences) what characteristics of this pneumonia x-ray task make it well suited for using a deep learning based approach. (3 pts)

Part 2. Prototype a deep-learning based “product” (9 pts)

This question is unrelated to Part 1.

Turning back to the Teachable Machine, build a prototype for something that you feel would be a marketable, deep-learning based product. This is an open ended question and you may receive extra credit for ideas that are especially creative! Then, publish it to an “endpoint URL” using Teachable Machine. Submit the endpoint URL and provide a short description of what it does and how it should be used. You receive three points for providing an URL that leads to a prototype we can view and test. Then, answer the following six questions about your product:

Question 2A

What would you name it? (1 pt for any reasonable answer)

Question 2B

Who is the market for this product? Be specific about the target market, including specifying how large this market is. (1 pt)

Question 2C

How would you reach this market to sell your product? What are the distribution channels and the unit costs of marketing and distribution? (1 pt)

Question 2D

What would you propose as the pricing model for this product? (1 pt)

Question 2E

What could provide a lasting advantage? In other words, what would prevent competitors from copying your product? (1 pt)

Question 2F

Do you anticipate that your product would introduce social/ethical challenges? If so, what might they be and how would you address them? (1 pt)

Part 3. Deep learning and explainability (10 pts)

Download this iPython notebook [this will be made available in class on Feb 22nd] and upload it into a Google Colab environment. Then, upload the x-ray data from above into the Colab environment.

Question 3A

Walk through and execute the notebook to understand the different parts of the deep learning process. Upload a screenshot of a classified image from the test data to verify this (it does not matter if it is correct or incorrect). (2 pts)

Question 3B

Looking at the tf-explain grad CAM output, how would you describe in simple terms which parts of these images are most important for this image classification task? This does not have to be a very specific or precise answer, but should describe what the grad CAM output is telling you in broad strokes. (2 pts)

Question 3C

If you were a product manager for a healthcare AI company, how would you integrate this information into a doctor-AI collaborative workflow in order to maximize effectiveness? At what point in the medical workflow would you introduce it? How would you present it and to whom? (2 pts)

Question 3D

In your view, what skills does a next-generation radiologist need in this context and how is it different from the current job description for a radiologist? (2 pts)

Question 3E

If you were the Director of Training for a large radiology program, how might you adjust your training program for medical students (i.e., future doctors)? (2 pts)