Lab 3: Classification
Due Wed. April 3rd in class

Neural network classifier

Introduction

We will again be using Google colab to experiment with neural networks in a Python notebook. Follow this link to start the lab. Before you begin working, save a copy of this notebook to your own drive.

You may work with a partner on this lab, if you'd like.

Analysis

After you have experimented with the provided notebook, write up your results in a document. Be sure to include the following in your results:

1. Begin by discussing the CMU faces dataset.
   • At the bottom of page 135 in Atlas of AI, Crawford argues that every dataset used to train a machine learning system contains a world view. Describe the world view of this dataset.
   • Do you think it was collected with the participants' consent and did they likely know it would be made publicly available?
   • How diverse is the dataset?
   • What do you notice about how these images were posed?
   • How well do you expect classifiers trained on this dataset would generalize to novel images that were not posed in the same way?
2. Next discuss the results of the sunglasses classification task.
   • For one particular run of the network, describe which hidden layer neurons are active when the network is presented with faces with and without sunglasses.
   • Include screen shots of the visualizations of the weights to each hidden neuron.
   • Try to explain what each hidden neuron is likely computing based on both the hidden layer activations and weights.
3. Then discuss the results of the direction classification task.
   • For one particular run of the network, describe which hidden layer neurons are active when the network is presented with people facing straight ahead, left, and right.
   • Include screen shots of the visualizations of the weights to each hidden neuron.
   • Try to explain what each hidden neuron is likely computing based on both the hidden layer activations and weights.

Bring in a printed copy your analysis to class next week.