LW - Rational Animations' intro to mechanistic interpretability by Writer

The Nonlinear Library: LessWrong

Link to original article
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Rational Animations' intro to mechanistic interpretability, published by Writer on June 15, 2024 on LessWrong. In our new video, we talk about research on interpreting InceptionV1, a convolutional neural network. Researchers have been able to understand the function of neurons and channels inside the network and uncover visual processing algorithms by looking at the weights. The work on InceptionV1 is early but landmark mechanistic interpretability research, and it functions well as an introduction to the field. We also go into the rationale and goals of the field and mention some more recent research near the end. Our main source material is the circuits thread in the Distill journal and this article on feature visualization. The author of the script is Arthur Frost. I have included the script below, although I recommend watching the video since the script has been written with accompanying moving visuals in mind. Intro In 2018, researchers trained an AI to find out if people were at risk of heart conditions based on pictures of their eyes, and somehow the AI also learned to tell people's biological sex with incredibly high accuracy. How? We're not entirely sure. The crazy thing about Deep Learning is that you can give an AI a set of inputs and outputs, and it will slowly work out for itself what the relationship between them is. We didn't teach AIs how to play chess, go, and atari games by showing them human experts - we taught them how to work it out for themselves. And the issue is, now they have worked it out for themselves, and we don't know what it is they worked out. Current state-of-the-art AIs are huge. Meta's largest LLaMA2 model uses 70 billion parameters spread across 80 layers, all doing different things. It's deep learning models like these which are being used for everything from hiring decisions to healthcare and criminal justice to what youtube videos get recommended. Many experts believe that these models might even one day pose existential risks. So as these automated processes become more widespread and significant, it will really matter that we understand how these models make choices. The good news is, we've got a bit of experience uncovering the mysteries of the universe. We know that humans are made up of trillions of cells, and by investigating those individual cells we've made huge advances in medicine and genetics. And learning the properties of the atoms which make up objects has allowed us to develop modern material science and high-precision technology like computers. If you want to understand a complex system with billions of moving parts, sometimes you have to zoom in. That's exactly what Chris Olah and his team did starting in 2015. They focused on small groups of neurons inside image models, and they were able to find distinct parts responsible for detecting everything from curves and circles to dog heads and cars. In this video we'll Briefly explain how (convolutional) neural networks work Visualise what individual neurons are doing Look at how neurons - the most basic building blocks of the neural network - combine into 'circuits' to perform tasks Explore why interpreting networks is so hard There will also be lots of pictures of dogs, like this one. Let's get going. We'll start with a brief explanation of how convolutional neural networks are built. Here's a network that's trained to label images. An input image comes in on the left, and it flows along through the layers until we get an output on the right - the model's attempt to classify the image into one of the categories. This particular model is called InceptionV1, and the images it's learned to classify are from a massive collection called ImageNet. ImageNet has 1000 different categories of image, like "sandal" and "saxophone" and "sarong" (which, if you don't know, is a k...