CS 5043: HW8: Explainable Deep Learning

Assignment notes:

Deadline:
- Expected completion: Friday, May 7th @11:59pm.
Hand-in procedure: submit a pdf to Gradescope
This work is to be done on your own. While general discussion about Python, Keras and Tensorflow is encouraged, sharing solution-specific code is inappropriate. Likewise, downloading solution-specific code is not allowed.
Do not submit MSWord documents.

Objective

Deep learning tools provide a great opportunity to create sophisticated models for making predictions across a wide range of domains. Their power comes, in part, from their large number of degrees-of-freedom (parameters). But, this property can also mean that the resulting models make incorrect decisions, or make correct decisions, but for the wrong reasons. As practitioners, it is important for us to always look carefully at how our models perform, not only in the aggregate, but also on a case-by-case basis.

Our objective for this homework is to use explainable deep learning techniques to analyze a couple of image classification models that we train using the Chesapeake Watershed data set. Our goal is to better understand how these models make their decisions and to evaluate the appropriateness of the learned models.

Data Set

We will use F0 of the Chesapeake Watershed data set. The loader code from HW 7 has been augmented with an additional function:

[new_outs, labels] = pixel_class_to_image_class(outs)

where:

outs is a pixel-level tensor (examples x r x c) that is returned from our loader
new_outs is an example-level tensor (examples x classes) that encodes binary class labels for each image. Each class is independent of one-another. The classes are as follows:
- Has water
- Has low vegetation
- Has barren
- Has impervious (other)
- Has road
Note: the following code will translate the new_outs into a count of the number of examples in each class:
```
       
np.sum(new-outs, axis=0)
```
labels is an array of strings that describe the classes

Models

Create and train two CNN-based classification models:

Single output: has road or not
Five outputs (yes, all of them)

Notes:

Use a sigmoid as the output of the last layer (don't use softmax, as that implies exclusivity across the outputs)
Use binary_crossentropy for loss
Use binary_accuracy for your primary metric
You will need more training examples than for HW 7 (at least 1000)
Expect training to be a lot faster than with HW 7
It is important that your validation performance be higher than the background accuracy (accuracy of making the best possible constant output)

Visualization

We will be using the tf-keras-vis package to visualize the behavior of your network under various conditions. The package has useful documentation, including example code.

Pick a set of 5-6 representative images from the validation data set that we will use to analyze the models. These images should have examples of each of the five classes.

For each of the models, do the following:

Use one of the saliency techniques to examine the pixel-level saliency for each of the classes and example images. If there are 6 images, then you will have a grid of 5 x 6 saliency images. Include in the title of each image the predicted probability for that class in that image and the true label for that class in that image
Do the same with one of the GradCAM methods
Use either a dense layer or convolutional layer tool to visualize what a set of 5 neurons / filters encode for the same set of images.
Note: always use 'clone=True' so you can re-use your learned models

What to Hand In

Hand in a PDF file that contains:

Code for generating and training the network. Some useful UNIX command line programs:
- enscript: translate code (e.g., py files) into postscript files
- ps2pdf: translate postscript files into pdf files
- pdfunite: merge several pdf files together
The above figures
A reflection on what you learned about the different methods and what they told you about your models.

Grades

30 pts: Model generation / visualization code. Is it correct? clean? documented?
40 pts: Visualization results
30 pts: Reflection

andrewhfagg -- gmail.com

Last modified: Mon May 3 18:10:57 2021