Ernest (Khashayar) Namdar

Researcher

I do what I do because I love it.

This is pretty much what I do every day:
trying to find the global minimum.

Play with the seed, learning rate, and number of iterations.
For example, seed = 0, lr = 0.02, iters = 80 will solve this one nicely.

Randomize the initial point How big the steps should be How many steps should be taken

This process is called optimization, and the algorithm you're using here is gradient descent. Most modern AI models are trained this way.

The surface you're optimizing over is called a loss landscape. In medical AI, where most of my research lives, these landscapes are usually non-convex and non-smooth, which makes optimization a lot trickier.

What you see here is a mild, interactive version of that challenge, except that I cannot see the loss landscape visually as it has many more dimensions than 3.