ML for Scientists: L1
Carol(ina) Cuesta-Lazaro
This week in AI4Science
https://cims.nyu.edu/~tristanb/statement.pdf
This week's goal
x
Features
Labels
y
\hat{y} = f_\theta(x)
\mathcal{L}(\theta, y, \hat{y})
Multilayer perceptron
f_\theta(x)
Screen Sharing ...
"Single hidden layer can be used to approximate any continuous function to any desired precision"
Optimization
Arbitrary accurate solution exists, but can it be found?
Generalization?
Overfitting
Universal Approximators
Screen Sharing ...
Image Credit: "Visualizing the loss landscape of neural networks" Hao Li et al
Empirical Muon
https://arxiv.org/abs/2606.16899