Hacker News new | ask | show | jobs
by musingsole 1439 days ago
But don't we now have thousands of hours of real-use? Surely in the age of Dall-e, a system to use that as input to ape if not fully simulate the surgical environment is within reach.
3 comments

While DALL-E might seem like something out of the world, what it outputs at the end of the day is pretty basic: a 2D image.

To run a high-fidelity simulator of human body that is useful for surgeons, you need a LOT more. And I doubt it can be data-driven. Data-driven simulators for things like autonomous cars are just coming up, and these are way simple as the agent (the autonomous car) doesn't directly change the environment when it's driving around (as in, you don't have to simulate the car colliding into a traffic pole and then breaking it, etc.)

To simulate a human body, you need to be able to capture the material properties of different layers, some of which are fluids, and then also the interaction between them and how different organs react to a surgical operation. It's a very hard problem.

Why do all that when you have animal corpses readily available? There's no need to create a problem that doesn't exist.

Content note: surgery gore

a) thousands of hours is far far far too little data. This is a total ballpark estimate, with only knowledge of the ML and basically zero knowledge of the surgery side, but I would expect three to six orders of magnitude greater, i.e. millions to billions of hours necessary to train a machine learning model to do something like that, to a standard.

b) the big problem is not the procedures themselves, but edge cases and things going a little bit wrong.

c) you are fine with tiny flukes in your generated images, but you won't be very happy with an ML model whose tiny flukes lead to internal bleeding.

d) unfortunately (with contemporary models) we can't easily explore latent space, which might possibly contain regions corresponding to catastrophic robot arm movements. the chance isn't that high, but it's not a chance most'd be willing to take at this point.

I think this overestimates what Dall-E is actually doing, and underestimates the complexity of the task.