AI Research Scientist Roadmap
From math and deep learning theory to paper reproduction, experimentation rigor, and novel model development — the structured, free path to becoming an AI research scientist.
Mathematical & Theoretical Foundations
6–8 weeksThe math that separates a research scientist from someone who calls `.fit()` — you need to derive it, not just use it.
What you'll learn
- Linear algebra and vector calculus for deep learning
- Probability, information theory, and entropy
- Optimization theory: convexity, convergence, saddle points
- Statistical learning theory basics (generalization, VC dimension)
Deep Learning Architectures & Theory
8–10 weeksDerive backpropagation yourself, then understand why every major architecture decision exists.
What you'll learn
- Backpropagation and gradient computation from first principles
- CNNs, RNNs, and transformer architectures
- Normalization, residual connections, and why they stabilize training
- Vanishing and exploding gradients — causes and fixes
Courses
Optimization & Training Dynamics
4–6 weeksWhy training works, why it fails, and how to diagnose the difference — a core research-scientist skill.
What you'll learn
- Gradient descent variants and stochastic optimization
- Variance reduction techniques in SGD
- Hyperparameter sensitivity and regularization theory
- Diagnosing a training run: loss curves, gradient norms, learning rate schedules
Reading, Reproducing & Critiquing Papers
4–6 weeksThe unglamorous skill that determines whether you can actually build on prior work instead of reinventing it badly.
What you'll learn
- Reading a paper efficiently: abstract → method → results → limitations
- Reproducing a published result from scratch
- Spotting weak baselines, cherry-picked results, and benchmark contamination
- Maintaining a running literature map for your subfield
Free Tutorials
Experimentation & Evaluation Rigor
4–6 weeksThe discipline that separates a real research result from a lucky seed.
What you'll learn
- Designing controlled ablations that isolate one variable at a time
- Statistical significance and variance across random seeds
- Avoiding evaluation leakage and overfitting to a benchmark
- Reproducible experiment tracking and versioning
Novel Model Development & Adaptation
6–8 weeksTurn theory into a contribution: fine-tune, adapt, and test new ideas against strong baselines.
What you'll learn
- LoRA, QLoRA, and parameter-efficient fine-tuning (PEFT)
- Designing and running architecture ablations
- Agentic and RAG systems as research substrates
- Prompt-based evaluation of new methods
Research Communication & Writing
3–4 weeksA result nobody can read or trust does not advance the field — or your career.
What you'll learn
- Writing a clear, reproducible methods section
- Building result tables and figures that hold up to scrutiny
- Responding to reviewer critique without getting defensive
- Presenting a research result to non-research stakeholders
Interview Preparation
OngoingPrepare for a loop that tests theory depth, implementation skill, and research judgment — not just LeetCode.
What you'll learn
- Presenting and defending your own past research
- Whiteboard research ideation and open-ended critique
- Implementing a method from a paper under time pressure
- Explaining a failed experiment and what you learned from it
Ready to start?
All courses are free and backed by real open source products you can use immediately.