Data Scientist at Lyft
PhD in Theoretical Physics
Kaggle GrandMaster (31st out of 76000+)
You need to get results.
Deep Learning is alchemy.
Q: But I know how to iterate fast?
A: Yeah, sure.
Human Grid Search (1000+ teams)
Many DL tricks published in papers were used long ago at competitions.
Top competitors have 2 machines:
My home DevBox
Crypto mining gives $1000-2000 per month :)
Kindergarten stage (2-4 hours)
Adult stage (1-10 weeks)
split into folds
Semantic Segmantation => FCNs (UNet, SegNet, FCN8-32, PSPNet, etc)
Trained for 20 epochs => Intersection Over Union 0.992
=> Yeah, sure.
IoU = 0.992 (435 out of 735)
IoU = 0.996 =>
=> yeah, sure (294 out of 735)
IoU = 0.9970 =>
=> yeah, sure (21 out of 735)
TernausNet: U-Net with VGG11 Encoder Pre-Trained on ImageNet for Image Segmentation
IoU = 0.9972 =>
=> Finally (5 out of 735)
IoU = 0.9973 => 1st out of 735
Competitions for conferences can often be won with Kaggle baselines:
Approaches developed in competitions can be published