I continue to be thoroughly impressed. Just last month, the latest debate between human / machine was pretty cool to see, with a machine making more compelling arguments against the human debate team (a quick advancement since it last learned to outperform humans at Jeopardy), or the month before, the new duplex demo. Next week, we should progress from seeing AI beat the world's best Go players to competing against professional Dota 2 players (which carries a $40m prize pool for the human players.) It's pretty high-dimensional data that these systems are making leaps and bounds toward understanding with superhuman levels of precision, although a lot of medicine is certainly image based -- there was a recent competition in China where the world's best radiologists competed to diagnose early brain cancer tumors and predict hematoma expansion... the machine outperformed the experts.l2silver wrote:Very interesting @SusanJ , I've definitely noticed a lot of talk on this form about dirty genes and Chris Masterson.
@apod in terms of data analysis and ai, on this front I'm a little skeptical. I work in this space, and I've seen a lot of disappointing results. I think the most successful applications have all involved image analysis, but we'll see what the future holds.
Just a few months ago, we saw superhuman performance at reading heart scans, outperforming cardiologists. Here's a clinical exam outperformed by machine learning just last month. There was also recent progress at outperforming dermatologists. A few weeks ago, a new algorithm (Modeling polypharmacy side effects with graph convolutional networks) was released for detecting drug interactions, outperforming the previous state of the art models. Looking at early onset Alzheimer's, this algorithm outperformed against all other previous multi-class classification methods, with the paper released just last summer.
This paper came out a few days ago: https://jamanetwork.com/journals/jamane ... le/2688539 which outperformed estimates from RCTs and SEER data, both of which are routinely used by clinicians for quantitative mortality risk predictions among patients with cancer starting chemotherapy.
I feel like I'm seeing unfathomable developments on a weekly basis in this space.
