376 Unit 6: Miscellaneous Topics

Background

Review my blog post on Mapping to Mimicry. I wrote it in one short sprint; feedback welcome!

• More resources: Ai generative art tools

Robotics

Rather than study theory, let’s look at two recent blog advances:

• An upgrade on a classic RL approach (Q-learning): Pre-training generalist agents using offline reinforcement learning – Google AI Blog
• A new kind of approach: just use Transformers for everything. PaLM-E: An embodied multimodal language model – Google AI Blog

Explainable and Human-Centered AI

ACM Selects: Trustworthy AI in Healthcare #02

What happens when AI meets people? How can we ensure that AI results are:

• Correct,
• Just, and
• Useful?

The first two are the subject of a subfield called Fairness, Accountability, and Transparency; the last is the subject of much research in human-computer interaction (HCI) and computer-supported cooperative work (CSCW). We’ll explore all three in these last two weeks of class.

Correctness and Transparency / Explainability

Read one or more of these:

• Explainable AI Guide (“Your high-level guide to the set of tools and methods that helps humans understand AI/ML models and their predictions.”)
• Interpreting Neural Nets: Skim one of these articles:
  • The Building Blocks of Interpretability
  • Activation Atlas
  • Visualizing Neural Networks with the Grand Tour
  • Multimodal Neurons in Artificial Neural Networks
• Interpretable Machine Learning: Fundamental Principles and 10 Grand Challenges | Abstract
• Techniques and limitations of post-hoc interpretability
  • Problems with Shapley-value-based explanations as feature importance measures
  • The Science Behind InterpretML: SHAP - YouTube
• Built-in interpretability: imodels: leveraging the unreasonable effectiveness of rules – The Berkeley Artificial Intelligence Research Blog
• Hundreds of AI tools have been built to catch covid. None of them helped. | MIT Technology Review

  Many unwittingly used a data set that contained chest scans of children who did not have covid as their examples of what non-covid cases looked like. But as a result, the AIs learned to identify kids, not covid.

Watch:

• Watch: Stop doing “explainable” ML

Supplemental Material

• See the “Hall of Fame” list at 5th VISxAI Workshop at IEEE VIS 2022 (also the list of examples under the Call for Participation)
• PyTorch implementation for Transformer Interpretability Beyond Attention Visualization

Justice (Fairness, Bias)

• Read one of the articles in ACM Selects on Algorithmic Fairness
• Prefer to watch something? 21 fairness definitions and their politics

Supplemental: The Effects of Regularization and Data Augmentation are Class Dependent | Abstract

Usability

Read or watch something from Human-Centered Artificial Intelligence.

Choose from one of the following notebooks, or do the Reinforcement Learning activities at the bottom of this page.

Neural Net Architecture

• Why so big? Counting parameters in sequence models (name: u13n1-count-params.ipynb; show preview, open in Colab)
• Models for Sequence Data (name: u13n2-seq-models.ipynb; show preview, open in Colab)
• Programming with Self-Attention (name: u13n3-self-attention.ipynb; show preview, open in Colab)

Reinforcement Learning

Policy, Value, and Q functions

Open up Observable RL Playground

1. Read through “Strategically Making Mistakes”.
2. What does a low epsilon do? What does a high epsilon do?
3. Try editing the maze = definition to edit the environment. What does it take to get the agent to tolerate a short-term negative reward to achieve a higher long-term reward?

Older activity that doesn’t work anymore

Go to the “Playground” at the bottom of this article.

1. Change the Algorithm to Q-Learning. We won’t look at the others at this time.
2. Try each of the “Visualization” options. What does each one show? Each one is a different function
3. Add one agent. How does completing an episode affect each of the functions that the agent is learning?

Exploration

1. Set the Explore-Exploit slider all the way to Explore. What do you notice about the agent’s behavior?
2. Set it all the way to Exploit. What do you notice now?

This environment isn’t rich enough for exploration to help much. So: go to a different playground, where we can actually edit the environment and see what the agent learns.