Deep Learning could be Good

You’ve probably followed the developments in Deep Learning in a similar way as I have. Exciting. Naively picking up – win at Go, or more recently, protein folding. That is clever.

You may not have had so many years of cynicism about the AI effort. God Damn [1].

I read a lot of fiction books about robots, way before you were born. Actually studied psychology for a year. Because my girlfriend at the time wanted to do it, but there weren’t enough candidates to make the class viable (at Stocksbridge College), until I signed up.
Best book ever was “Introduction to Psychology”, an A level and/first year text. I suspect now a lot of it was wrong, wrong, wrong. But it did place models in there, aside from of some discredited things, lots about using statistics.

Slipping my point.

Ok, if you are in the middle of your PhD now, close your ears.
Intuitively, it’s really easy. Make a model, play with it, input vs. output.
My maths is pitiful, I still don’t get DFT, but these things just work as black boxes.

A while ago I found out that earthquakes sometimes have audio frequency radio wave precursors. Simultaneously, Deep Learning kicked off. For mental health reasons I’ve had to put that on pause, but it’s there. I think very easy to do, conceptually. Getting the bloody machines to do it, I’m not a great coder, hard work.

A friend has ordered an ultrasound scanner for herself. She is by profession a gynecologist, terrifyingly altruistic person, now doing Covid tests. Worked the other year in Rwanda, now up and down the mountains here testing people.

As a gyno, can see the attraction of an ultrasound scanner (she plans to take it to Africa for normal purposes). But in the immediate, she’s getting it for lung scans on Covid patients ($1800 out of her own pocket).
I google “pulmonary covid scans deep learning” – yeah. Yup, first hit was a convincing paper suggesting it’s better that hman. That. Isn’t peer-reviewed yet, but a paper on how this gives better results than humans.

I don’t know, as a screening technique, this could be useful.

This is useful?!

I don’t see anything like GAI happening until people get out of the record-play kind of approach. Next step will be real-time feedback loops.

[1] Steppenwolf, before your time


Leave a Reply

Your email address will not be published. Required fields are marked *

Post comment