Roadmap

Deep Learning
1

Hi everyone,

I wanted group’s opinion on the road map I should design for myself or I should not care about a road map. But before that, some context:

Context: I am some one who has been trying to make a transition from Project Management into Data Science. Education wise I am an Electrical Engineer and no / less coding experience. However, in last 4-5 years I have learnt coding and even real world modelling and deployment experience (classical Machine Learning).

I now want to get myself into into LLM space. However I have limited time given the full time job. I also have a side project where I have to build an text classifier i.e. given some text I have to identify if it is written by a fraudster for which my initial sense says BERT will be good. If I am able to apply and learn simultaneously, it is a win-win situation.

Now, 2 options -

  1. Understand LSTM, RNN, Attention, Transformer and BERT (from some online playlist or course) and then implement BERT.
  2. or Fine Tune BERT and then see how I can improve BERT and in the way whatever terms I see I study them topic by topic backwards.

Understanding Pytorch is required in both.

I am a learn by doing person and thus gained a lot by doing Part 1 of Fast AI course.
I have also done 1st 2 lectures in Andrej Karpathy’s course (NN Zero to hero course) which has a very similar approach to that of Fast AI.

Please advise how my road map should look like ?

3

Given your “learn by doing” style, focus on immediate application: Option 2 is your best bet. Start by fine-tuning a pre-trained BERT model (using Hugging Face Transformers and PyTorch) for your fraud text classifier. As you work on the project, learn the underlying concepts like tokenization, attention, and the Transformer architecture in a “backward” fashion – understanding them as they become relevant to your task or when you question why something works. This project-driven approach provides immediate wins, reinforces learning contextually, and efficiently builds your LLM knowledge while simultaneously addressing your side project.