Alan Martyn

ml engineer with background in audio engineering

[Alan Martyn @AlanMartyn]

My mum died of breast cancer when I was 16. She was 43.

I worked on this problem for 5 years at Kheiron, alongside many talented people dedicated to improving patient outcomes.

Satisfying beyond words to see the progress continue. There is still much work to do. Onwards!

[Eric Topol @EricTopol]

A review of the new trial results (below) along with all the other evidence and conclude it's time to move to AI-support for all mammography (with no cost to patients) https://t.co/hITOBNd7pg

[Eric Topol @EricTopol]

A review of the new trial results (below) along with all the other evidence and conclude it's time to move to AI-support for all mammography (with no cost to patients) https://t.co/hITOBNd7pg

[Alan Martyn @AlanMartyn]

@Austen @RCT may be a good partner. A while back I helped them build this app that has most of Da Vinci’s anatomical drawings. Translated notes marked up in html (with links I think), comments, hi-res images

A fascinating insight into an insatiably curious mind.

https://t.co/BLTEkm8OkZ

[Mathigon @MathigonOrg]

All repeated, 2-dimensional patterns have one of 17 different symmetries. These are called the “Wallpaper Groups”, and you can explore all of them at https://t.co/6eUKg1MBel

[Alan Martyn @AlanMartyn]

@irinimalliaraki Wow, looks just like a project we built once:
https://t.co/Da66YM9XyD

[Alan Martyn @AlanMartyn]

Building Helmholtz 5 years ago with @laurencegreed and @GyGalik and a too many other v. cool people: https://t.co/Da66YM9XyD

[Alan Martyn @AlanMartyn]

How do regularised models behave when two features contain the same information? A quick experiment:

https://t.co/VICpUD9RsR

[Alan Martyn @AlanMartyn]

@fchollet I’ve recently implemented U-net and Pix2Pix Gan in Keras here:
https://t.co/oGTiUAL0iX

and here:
https://t.co/Gqr410c0iG

[Alan Martyn @AlanMartyn]

Segment space odyssey.

Or, a comparison of a few U-Net variations applied to image segmentation:
https://t.co/oGTiUAL0iX

Thanks @ORonneberger

[Alan Martyn @AlanMartyn]

How residual shortcuts speed up learning... or Why do convolutional layers struggle to output their input?

My attempt at a demo, you can run the code in one click!
https://t.co/r0biLkIfb1

#100DaysOfMLCode

[Alan Martyn @AlanMartyn]

@paulg Here you go...

[Alan Martyn @AlanMartyn]

Experiments in generating a colorbar of the 5 most dominant colors in an image. K-means clustering works incredibly well for flat images.

[Alan Martyn @AlanMartyn]

The Jupyter Notebooks integration in VsCode is neat. Left pane is notebook editor, right pane is log showing cell ouputs over time. Makes the current state a bit more explicit.

[Alan Martyn @AlanMartyn]

This idea has had a pervasive impact in the field. For example it is a precursor to the Neural Ordinary Differential Equations paper which won best paper at NeurIPS 2018 and which @sirajraval describes nicely here:
https://t.co/HOhEYuHHuu

[Alan Martyn @AlanMartyn]

Here's my implementation of the CIFAR10 experiment from Kaiming He's original ResNet paper. I'll try and summarise in this thread. What did I get wrong? What did I miss?

https://t.co/rTlayms4BM

[Alan Martyn @AlanMartyn]

This simple addition allowed networks up to 110 layers deep to outperform shallower equivalents, in turn achieving state-of-the-art at the 2015 ImageNet competition.