OpenAI yesterday unveiled its Open AI Microscope, which provides visualizations of every significant layer and neuron in eight of today’s most popular computer vision (CV) models. Interactions between neurons indicate the abilities of neural networks, and with machine learning trending toward increasingly complicated neural networks it is important for researchers to be able to quickly and easily conduct a closer inspection of these thousands of interactions. This is where AI Microscope comes in.
Just as biologists gain insights into organisms by putting model specimens under their microscopes, AI Microscope was designed to help researchers analyze the features that form inside leading CV models.
OpenAI explains that its Microscope models are composed of a graph of nodes — neural network layers connected via edges. “Each op contains hundreds of ‘units’, which are roughly analogous to neurons. Most of the techniques we use are useful only at a specific resolution. For instance, feature visualization can only be pointed at a ‘unit’, not its parent ‘node’”.
Feature visualization allows users to see how a neural network builds up its understanding of images over many layers. When pointed at a given unit, feature visualization indicates an artificial, optimized image that maximizes activations of the unit.
Other AI Microscope visualizations include the 2015 image-hallucination software DeepDream, dataset examples, and synthetic tuning curves.
Using dataset examples, users can identify pieces of images from the training dataset that result in the largest activations from a given unit. Neuroscience-inspired synthetic tuning meanwhile plots test patterns and shows how strongly a unit reacts to specific artificial inputs.
OpenAI says Microscope can reduce the neuron exploration feedback loop from minutes to seconds, enable prompt scrutiny of neuron research results, and help keep interpretability accessible.
The company believes the AI Microscope’s primary value will be in facilitating the long-term comparative study of models, and hopes the project could also benefit neuroscience researchers by helping to reveal the internal workings and connections of neurons.
The AI Microscope concept stimulated the imagination of Google Brain Research Scientist David Ha, who tweeted “This fun project makes me think that in the distant future, we can go to the computer history museum and look back at this thing called Deep Learning that had been so popular back then.”
The Open AI Microscope currently examines eight CV models in detail. OpenAI is planning to expand the scope in the future. Users who find interesting things under Microscope are invited to share these via a dedicated Distill Slack.
Journalist: Fangyu Cai | Editor: Michael Sarazen