- What is machine learning?
- Deep Learning
- Sofwater Engineering
- AI in the world
It does whatever you train it to do
You can use ML for everything
So what's deep about deep learning?
How big are these deep networks?
Not too long ago this was totally true
People used to be tougher
There was a time for Caffee...
There was a time for Tensorflow...
Also, try on transfer learning:
If you overfit, add regularization
Try different hyper-parameters. Sometimes you hit the jackpot
Sometimes you don't hit the jackpot
If nothing else works
Finding bugs in DL is hard
You don't know what went wrong, the answer is just bad
We still don't really understand many elements of deep learning
For example, we know Batchnorm is helpful in many situations, but the explanation for why keeps changing
If you're in RL...
Let's make things simpler to understand (unlike how
f(x) = max(x, 0) became "rectified linear unit")
Bringing research into production
More on "cutting-edge" research
Implementing your own version is also tricky
Totally underappreciated: curating your own dataset.
Yepp, someone labelled those datasets
Sklearn still has its uses
Colab is actually quite awesome
I often go back and forth
I can relate to both
Some things never change
True, not just in deep learning
The difference often matters
Things look tidy until you look under the hood
Well, I do like my own implementation of everything...
What is a "data scientist"?
Precision vs recall
Okay, this is pure math, but still good
How to impress people with AI