Give it to me straight BOT, how difficult is this shit actually to learn to the point where I can incorporate it in my own projects? I have a masters degree in mathematics, but never did a lot of courses in applied statistics and optimization beyond the basic bachelor courses. What would be the best way to teach myself ML from this starting point?
I like to think its easy, but when I see people from different fields applying it, I am not so sure. Especially biologists.
If you are a solid mathematician you might be fine.
Jax is better
look at the sklearn python library and take it from there
if you have a masters in applied maths you are already in a better position than 99% (or more) of the homosexuals that """""learn""""" ML/data science by applying a few python libraries. take a look at the mathematics for machine learning book, it's fairly terse but given your background you should be able to follow the entire thing. look into the first 5 chapters or so of PRML by bishop too, it's a classic in the field. once you're cool with that stuff you will be able to understand ML better than most people in the field
Keras which is built on top of tensorflow is fairly easy to get into. If you're going down this path, start with that maybe. Pytorch is preferred in academia though because it initially allowed models to be created withou the need to compile them, as required tensorflow 1.0. Other than that, Jax is more performant but more niche. If you are comfortable with numpy, the basics of all these frameworks are easy to grasp.
btw you dont even need to be into neural network stuff for these frameworks to be useful. If you do a lot of numerical computing involving tensors, you will rapidly see the benefit of automatic differentiation and gpu computing.
>how difficult is this shit actually to learn to the point where I can incorporate it in my own projects?
ChatGPT is good for getting the ball rolling. Read up on tensor flow, pytorch, keras, etc but by just smashing in basic prompts and examples of the data you want to use it spits out starter code that you can build on.
A lot of the industry doesn't use tensorflow. I wouldn't worry about it unless you know a company you work with uses it. DE >>>>>DS
>Give it to me straight BOT
lmfao good fricking luck
It's actually not hard, I find the hardest parts are the piping and connections.
Start with the basics of statistical learning and decision optimization (all of it is fairly straightforward, but for whatever reason everyone tries to skip or rush over it). I recommend Introduction to Statistical Learning with R, it's a free and comprehensive text that goes over most of the popular algorithms. The actual AI part of it is basically determining what to do (i.e., decision optimization) based off perception and what the stats/model is saying.
Learn state machines for the AI portion and Q-learning for a basic learning algorithm (again something fairly basic, but people skip over it).
The entire field is fairly muddy, but AI and Machine Learning encompass a large portion of applied math. Studying game theory, using and automating it is effectively AI for example. Lots of tasks like this where you use a math field then automate become part of the AI field.
But learn data analysis. That's always step one. Then learn optimization, network optimization is a good starting point. The rest is just piping data around and autistic programming tasks to make sure things don't break.
it's calc 1 shit, they work via the chain rule
Here's what I did coming from out of field
>Get project
>Search relevant ML models
>Make model in python
>Finish project
Not that hard
If you have a masters in math then no problem
> I want to build a model from scratch with zero knowldge in ML
well good luck without reading up a bit its hard because there are so many options
> I am willing to read
makes it easier to make your own model
> I am willing to reuse model/weights
makes it as easy as just importing and running train
its not hard unless you make it so don't be afraid if you don't understand how/why something works most people in the space have no clue either.
I'm learning ML now I'll give you my schizo interpretation of the area. Academicization of the field should be avoided. You don't need to know how every NN works like lstm or RNN or whatever and waste time stepping through them. Learn the basic forward and backward steps through a NN, how gradient descent works, how the chain rule works, etc in a day.
Then learn the basics of transformers. Which are ways to enhance the basic NN with better ability to handle things in their larger context. Transformers are the king of AI right now so grok how qkv can work to enhance contextual understanding. Like every word we injest into our brains has multiple definitions and meanings depending on the story, context, usage, transformers can send this context included version of words into a NN and work on everything not just words.
conv NN are next, just know they are good for features like vision finding edges, maybe look into the implementation if you want
Next we have RL, which is when the computer creates data itself in an explorer way, this means it can get stuck in certain areas etc. You'll leanr about state, actions, transition functions, new states, policy, value functions etc. Learn all of this till you grok it.
DO kot get bogged down in academics, they will teach you 5 different ways to write the policy variable. Just use chat gpt and the basic overview of each. Avoid learning 10+ implementation details of each NN full of tricks.
Academics want to corner the field so they invent endless busy work, multiple definitions, new names, and want to formalize every shit NN idea into a field.
Learn forward and backward pass.
Learn transformers qkv
Learn RL concepts avoid implementation details
Learn when to use Convnets
There you are done. Now build an application to make money with openai API because it's LLM is the best and just gluing into anything will be better use of your time than academic shit.
Everything is tinkering in the field
to be honest to get a job you don't even need to know how any of these things work, just know some libraries that can be used and how to modify the parameters of code you copy from online/chatGPT to do a particular task
Yes, ignore academic approach just go right into using them or at most get the fundamentals asap then use them. Anything academic is a meme around them
>Where to start?
pip install keras, tensorflow
It's simple, and there's no statistics involved.
I had no knowledge of python and managed to have outputs on the input with model, but I red several articles on how NN and DNN works.
Just be sure you use ada_delta and squared hinge loss.
This. Chatgpt has sped up my NN development by a ton, I just ask for code and tweak it slightly and it just works