Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Building an AI to predict human age from a blood sample

colekillian.com

Post reply on HN
Print view View original

11–20 of 68 posts

Re: Building an AI to predict human age from a blood sample

#11

Post by chewxy »

deepnotderp wrote: post #10
chewxy wrote: post #5

>Back to the computer science: 470,000+ features sounds nice at first, but is a recipe for overfitting when we only have 700 samples at our disposal. Proceeds to use (1024^2 * 2 + 1024) parameters in the neural network.

I know this is a joke, but the theory of generalization in NNs is rapidly advancing and it's not quite that simplistic: https://arxiv.org/abs/2003.02139

Ya. And the choice of optimizer (in this case adam) also imposes upon it some regularization scheme.

I just thought I'd highlight a bit of funniness.

Re: Building an AI to predict human age from a blood sample

#15

Post by cjbprime »

echelon wrote: post #8

Far-reaching prediction: they're going to do facial prediction from blood samples as well. Law enforcement really wants to generate sketches from unknown DNA found at crime scenes. That is, of course, in addition to the all-encompassing family trees we're providing them with 23andme.

In what sense in this a prediction? The challenge of estimating faces from DNA is one that researchers have already been competing and publishing papers on for years: https://www.pnas.org/content/early/2017/08/29/1711125114

Re: Building an AI to predict human age from a blood sample

#17

Post by cletus »

So I'm not ML guru or anything but what I learned was that if you have m features on n samples you want n > m to prevent over-fitting, no?

Also, with so few samples, how do you do your hyperparameter tuning and validation?

I mean you could eliminate certain features in isolation but that doesn't capture dependent features. And how would you do dimensionality reduction?

Re: Building an AI to predict human age from a blood sample

#19

Post by jonathankoren »

unwoundmouse wrote: post #13

Great work! I love how well presented all of the information is. I'd be really interested to see how well a baseline linear model using those features would perform - it seems like it could do pretty well.

A linear model should always be compared to these DNNs.

There was a paper last year or so that compared correctly tuned linear models to various deep belief net papers and found that the performance "gains" suddenly evaporated or were not nearly as great as originally published.

If I can track down that paper, I'll post it.

Post reply on HN