Street View

Seeing neural nets, piece by piece

In a new video, Jane Street ML researcher Ricson Cheng takes ideas from neural net verification and turns them into something you can see: an animation that shows how ReLU networks carve the plane into piecewise linear regions, and how those regions evolve as you move through weight space.

Built in Python and rendered with Matplotlib, the visualization encodes output as brightness and gradients as hue, doubling as both an artistic experiment and a research tool. It's the kind of curiosity-driven work we're excited by at Jane Street – something that started as a personal interest and was expanded and refined with help from colleagues. For more background on the ideas behind the visualization, you can read Ricson's original blog post on piecewise-linear neural nets.

Curious about ML at Jane Street? Learn more at janestreet.com/join-jane-street/machine-learning/.