>>12535043/sci/ is full of snob-undergrads who think you need to understand the mathematical basis of ML in order to use it. Truth is, you can build good models having almost no real understanding of what's running under the hood, as long as you understand the basics of how to make good models. For this, I think its more important to understand statistics and experimental design than the rigorous math behind ML. I would argue that you need just the basics, and that it's more important to first get started with projects in ML and get used to coding and using different libraries (Python's sklearn, I use Pytorch for NNs, and/or R's CARET or other similar packages). What you need to know:
-Regression vs classification. Linear models. If you can take linear algebra, it will be the single most useful class for you for future work.
-What a ML model is (objective functions to minimize, understand backpropagation and how it works, and things like gradient descent/stochastic gradient descent are fucking awesome imo when it clicked for me how it works)
-Underfitting/overfitting and the value of cross-fold validation and splitting of training/test sets
-the basic ideas (and math if you can) of extremely common ML models: SVMs (and the kernel tricks), trees/CARTs, adaboost, etc)
-preprocessing data: This is where it gets artsy depending on the problem, and where its really important to know statistics/experimental setups.
-basic NNs: I can tell you, once you know how NN/deepNN/convNNs and variants such as RNNs work, you have it all down mostly. 99% of modern papers with different architecture work like this: We have a problem, we created a fuckton of different architectures, one worked really well, then we made up horseshit about how we "reasoned" our way to the new architecture in this paper. It's all trial and error.
-KAGGLE. Want to sink your teeth into big boy world? Look up notebooks and follow them. Follow notebooks posted by current competitions.
