>> Its vision system relied on near-infrared fluorescent (NIRF) tags placed in the intestinal tissue; a specialized NIRF camera tracked those markers while a 3D camera recorded images of the entire surgical field. Combining all this data allowed STAR to keep its focus on its target. The robot made its own plan for the suturing job, and it adjusted that plan as tissues moved during the operation.
>> The researchers trained STAR only on how to perform this particular intestinal suturing procedure. “We programmed the best surgeon’s techniques, based on consensus and physics, into the machine,” Kim said.
So it's a combination machine learning - expert system robot. Machine learning (I assume Convolutional Neural Networks) for vision, some planning algorithm for the suturing job and expert knowledge for the surgical techniques.
That's the shit I'm talking about! Let's see more of that! Don't just end-to-end train some deep net and try to have it learn to do everything from scratch. Use background knowledge! Combine techniques! Be smart, dammit!
Edt: btw, this sort of thing, systems with hard-coded expert knowledge besting human experts, that's age-old stuff. It's how it used to be done before the last winter (AI winter). Nice to see it back.
Edt 2: Apologies for the ex!cla!mation! marks! but I'm so! excited!
Ho-hum. After reading the paper, they didn't use anything like machine learning or expert knowledge. They just kludged the whole thing step-by-step into the robot. Its "planning" was to draw a line between the NIRF tags on the surface of the pig gut.
So it's not even properly autonomous- they're just misusing the term to mean that nobody was remote-controlling it.
>> Its vision system relied on near-infrared fluorescent (NIRF) tags placed in the intestinal tissue; a specialized NIRF camera tracked those markers while a 3D camera recorded images of the entire surgical field. Combining all this data allowed STAR to keep its focus on its target. The robot made its own plan for the suturing job, and it adjusted that plan as tissues moved during the operation.
>> The researchers trained STAR only on how to perform this particular intestinal suturing procedure. “We programmed the best surgeon’s techniques, based on consensus and physics, into the machine,” Kim said.
So it's a combination machine learning - expert system robot. Machine learning (I assume Convolutional Neural Networks) for vision, some planning algorithm for the suturing job and expert knowledge for the surgical techniques.
That's the shit I'm talking about! Let's see more of that! Don't just end-to-end train some deep net and try to have it learn to do everything from scratch. Use background knowledge! Combine techniques! Be smart, dammit!
Edt: btw, this sort of thing, systems with hard-coded expert knowledge besting human experts, that's age-old stuff. It's how it used to be done before the last winter (AI winter). Nice to see it back.
Edt 2: Apologies for the ex!cla!mation! marks! but I'm so! excited!