Readings

• Image “upscaling” algorithm with racial bias.
• Twitter cropping algorithm exhibits bias as well.
• (optional) Excavating AI on bias in ImageNet and other AI systems.

Group Discussion:

What are the sources of bias in AI systems?
What kinds of bias appear in these systems?

Group Discussion:

Who is responsible for harm done by an AI system?

Group Discussion:

If you were a movie villain, how would you use ostensibly harmless/helpful AI to do evil?

Bias

• AI both amplifies and disguises bias.
  • Machines are wrongly seen as unbiased.
  • In fact, they thrive on biased data.
• Theory-free AI is especially prone to bias.

Data Science

• Normally, a model is based on a theory.
• Data confirms or complicates the model.
• Theory is revised to produce a new model.

“Data science” sometimes skips a step and creates a model directly from data.

Theory-free Models

• A theory-free model doesn’t teach us anything.
• We can’t critique the theory behind it.
• It might still “work” very well.

Models and Authority

• Scientific models can be components of oppressive systems.
  • Provide “justifications” for racism, colonialism, etc.
  • Put more power in the hands of the powerful.
  • “Science” sounds authoritative.
• It’s easy to focus on whether a model or theory works, without ever asking about what effect it will have.

Regression

• Normal math: solve an equation.
• Regression: match an equation to data.

Regression

If we have $n$ data points $(x_i,y_i)$, our error is:

$E = \sum_{i=1}^{i=n} \sqrt{(y_i - f(x_i))^2}$

We can find an analytical minimum if our
function $f$ is simple, like $f(x) = mx + b$.

Gradient Descent

• What if $f(x)$ is too complicated to solve the error equation?
• We can try to find a minimum using Newton’s method as long as it’s differentiable.
• Basic idea:
  1. Take the derivative of our error function with respect to the constants in that function, $\frac{\delta E}{\delta C}$.
  2. Nudge those constants in the direction that reduces the error.
  3. Repeat until we get to zero error (or a minimum).

Overfitting

• Regression is only meaningful if the function is suitable.
• A complex enough function can fit any data.
• Only a theory can give you confidence to interpolate or extrapolate.

Videos

• Video: Deep Learning Basics including Autoencoders (14:55 to 33:40)
• Extra Video: What is a Neural Network?
• Extra Video: How Neural Networks Learn?

Core Concept

• Create a really general function.
• Fit it to whatever data you have.
• Use the fitted function.

Training

• Gradient is hard to compute for 1000s of data points.
• Just compute gradient for a few data points at a time, and repeat this process over and over again.
• Use a small learning rate to move slowly.
• Problems include overfitting as well as local minima of the error function.

A Neural Network

$$y_j = f(\sum_{i}{w_{j,i} \cdot x_i})$$

$$\left[ \begin{matrix} y_1 \\ y_2 \\ \vdots \\ y_m \end{matrix} \right] = f \left( \left[ \begin{matrix} w_{1,1} & w_{2,1} & \cdots & w_{n,1}\\ w_{1,2} & w_{2,2} & \cdots & w_{n,2}\\ \vdots & \vdots & \ddots & \vdots \\ w_{1,m} & w_{2,m} & \cdots & w_{n,m}\\ \end{matrix} \right] \cdot \left[ \begin{matrix} x_1 \\ x_2 \\ \vdots \\ x_n \end{matrix} \right] \right)$$

Network Structure

• Matrix of weights multiplied by input vector.
• Apply an activation function to each output.
• Simulates neurons in the brain: several signals are combined with different connection strengths to trigger an output.
• Weights control connection strengths; activation function determines result.

Sigmoid

$$f(x) = \frac{1}{1 + e^{-kx}}$$

Deep Networks

• A deep network has multiple layers.
• Input from one layer is fed into another layer.
• Size of output is usually smaller each time.
• Millions to billions of parameters (weights).
  • Let a computer calculate the derivative…

Example Target Function

Input ($x$)	Output ($y$)		Input ($x$)	Output ($y$)
	[1, 0]			[0, 1]

What does it learn?

Video excerpt

• Under the right conditions, it learns what humans do: how to find edges, and then patterns of edges, etc.
• The right conditions are tricky to produce.

What does it need?

• A huge amount of data
  • Minimum hundreds images; millions for high quality
• Labeled data, so that we can compute error values
  • All sources of labeled training data are biased 🙁

Autoencoding

• Set $f(x) = x$ and ask network to learn the identity function
  • Network has to learn compression, because its internal state is smaller than the original image.
• Learn the basics from unlabeled data
  • Unlabeled data is much easier to come by
  • All sources of unlabeled training data are also biased 🙁

Applications

• Image captions
• Deep fakes
• Generating art (bonus nightmare fuel)
• Face recognition
• Image super-resolution
• Crime prediction
• Etc.

Discussion

How might biases affect these applications of deep learning?

How should we feel about deep learning and AI as a scientific project?

Computational Creativity

• Can computers be creative?
• What does it mean to be “creative?”
• Could computers support human creativity?

Critical (Computational) Creativity

• Critique-by-reply.
• Building generators to deeply understand systems.
• A mix of computer science and media studies.
  • Focus on interactive media, especially video games.

Past Projects

Measuring “novelty” using an auto-encoder network.

Convolutional Networks

• Like the eye, apply the same neural structure in parallel to many pixels.
• Output is a smaller field of pattern vectors.
• Combine pattern vectors using pooling to reduce processing requirements.

Convolutional Network Explanation

Recurrent Networks

• Remember information across multiple activations using a “memory” input in addition to the usual input.
• Output a new “memory” vector in addition to usual output.
• Useful for processing or generating sequences with variable length, like text.

LSTM Explanation