You're finishing your second year of a data science degree with a solid grasp of the fundamentals, and you feel stuck. That's not a setback, it's a signal that you're ready for the next layer of learning. The confusion you're experiencing isn't a lack of direction; it's a healthy response to an overwhelming amount of choice. The real challenge isn't finding good material, it's knowing which material to trust and in what order to engage with it.
Let's start with the practical fork in the road: PyTorch versus Keras. If your goal is to understand how neural networks actually work under the hood, PyTorch gives you more control and is the dominant framework in research and industry. Keras, now part of TensorFlow, is more abstracted and easier to pick up, but it can obscure the mechanics. Given that you already have a comfortable command of Python and sklearn, you're ready for the steeper learning curve of PyTorch. Start there. Pair it with fast.ai's practical, project-driven approach, Jeremy Howard's course is designed to get you building real models quickly without bogging you down in theory first. Andrew Ng's deep learning specialization is excellent for theory, but save it for after you've built something tangible. Theory sticks better when you've already seen a model fail and need to understand why.
Kaggle competitions are not a learning path, they are a testing ground. Use them after you've completed a structured course and built a few projects from scratch. Jumping into Kaggle without that foundation will leave you reading other people's code without understanding the decisions behind it. Instead, build your own projects. Start with a classification problem that matters to you, predicting something from a public dataset, like housing prices or customer churn. Apply logistic regression, then a decision tree, then a simple neural network. Compare them. That comparison is where real learning lives.
Your SQL skills are a quiet advantage. Most data science courses underemphasize querying, but in practice, you'll spend more time pulling and cleaning data than training models. Keep sharpening that skill alongside your Python work. The combination of solid SQL, a strong foundation in classical models, and a growing comfort with neural networks will set you apart from peers who chase every new framework without mastering the basics.
Here's the concrete plan: finish fast.ai's Practical Deep Learning for Coders. Then build two projects from scratch using PyTorch, one for image classification, one for tabular data. Then take Andrew Ng's deep learning course to formalize your understanding. Finally, enter one Kaggle competition with the goal of learning, not winning. That order respects your existing knowledge, builds confidence, and turns confusion into momentum.