ML/DL Engineer Learning Path
This 16-week study plan takes you from mathematical foundations to deployment-ready deep learning engineer. Each week has specific readings from the Archon knowledge base, hands-on exercises, and a milestone. Budget 15-20 hours per week. This path cross-references the 69 EDA pages, 25 data pipeline pages, and AI engineering pages for complete coverage.
Who This Is For
- Computer science graduates starting ML/DL careers
- Software engineers pivoting into ML research engineering
- Data scientists who want deeper understanding of neural network internals
- Anyone who wants to implement architectures from scratch, not just call APIs
Prerequisites
- Python programming (comfortable with classes, list comprehensions, decorators)
- Basic linear algebra (vectors, matrices, dot products)
- Basic calculus (derivatives, chain rule)
- Basic probability (Bayes' theorem, distributions)
Total estimated time: ~250 hours across 16 weeks (15-20 hrs/week)
Overview
Phase 1: Foundations (Weeks 1-4)
Week 1: Mathematical Foundations and Python for ML
Read:
- Deep Learning Overview -- understand the landscape
- Math Foundations -- linear algebra, calculus, probability for ML
- Python ML Ecosystem -- NumPy, pandas, scikit-learn overview
- NumPy -- array operations, broadcasting, vectorization
EDA cross-reference (for data intuition):
- Data Types Deep Dive -- understand data before modeling
- Understanding Distributions -- probability distributions in practice
Exercise:
- Implement matrix multiplication from scratch in Python
- Implement gradient descent for linear regression using only NumPy
Milestone: You can explain the chain rule, compute matrix products, and write vectorized NumPy code.
Week 2: Neural Network Basics
Read:
- Neural Network Basics -- perceptrons, activations, backprop, optimizers
- Architecture Selection Guide -- overview of all architectures
ML foundations cross-reference:
- Linear Regression -- the simplest model
- Logistic Regression -- classification foundations
- Evaluation Metrics -- how to measure model quality
Exercise:
- Implement a 2-layer MLP on MNIST from scratch with NumPy
- Achieve >97% accuracy by tuning learning rate and architecture
Milestone: You can derive backpropagation by hand and implement a working MLP from scratch.
Week 3: PyTorch Fundamentals
Read:
- PyTorch Fundamentals -- tensors, autograd, nn.Module, DataLoader
Exercise:
- Rewrite your NumPy MLP in PyTorch using nn.Module
- Train CIFAR-10 classifier
Milestone: You can write a complete PyTorch training loop from memory.
Week 4: Training Techniques
Read:
- Training Techniques -- BatchNorm, dropout, weight init, LR scheduling, mixed precision
- DL Checklist -- 40-item project checklist
- Cross-Validation -- proper evaluation methodology
- Hyperparameter Tuning -- grid, random, Bayesian
Exercise:
- Systematically add BatchNorm, dropout, data augmentation, cosine LR schedule to CIFAR-10
- Measure improvement from each technique
Milestone: You can apply every standard training technique and understand when each helps.
Phase 2: Architectures (Weeks 5-8)
Week 5: Convolutional Neural Networks
Read:
- CNN -- convolution math, LeNet to ResNet to EfficientNet
- Image Classification -- augmentation, ViT, transfer learning
EDA cross-reference:
- Image & Audio EDA -- understanding image data
Exercise:
- Implement ResNet-18 from scratch
- Fine-tune pretrained EfficientNet on a custom dataset
Milestone: You can implement ResNet from scratch and apply transfer learning.
Week 6: Sequences: RNNs and LSTMs
Read:
- RNN and LSTM -- RNN equations, vanishing gradients, LSTM gates, GRU
- NLP Fundamentals -- tokenization, Word2Vec, embeddings
EDA cross-reference:
- Text Features -- text data exploration
- Text Preprocessing -- cleaning text for NLP
Exercise:
- Implement LSTM from scratch in PyTorch
- Train sentiment classifier on IMDB -- target >87%
Milestone: You can implement LSTM from scratch and explain why it solves vanishing gradients.
Week 7: Transformers
Read:
- Transformers -- self-attention, multi-head attention, positional encoding
Exercise:
- Implement a transformer from scratch (attention, multi-head, encoder, decoder)
- Visualize attention patterns
Milestone: You can implement a transformer from scratch and explain every component.
Week 8: NLP with Transformers
Read:
- Language Models -- n-gram to GPT, pre-training, scaling laws
- BERT Family -- BERT, RoBERTa, DeBERTa, sentence-transformers
- Text Generation -- decoding strategies, RLHF, DPO
Exercise:
- Fine-tune BERT on CoLA and measure Matthews correlation
- Build a semantic search engine with sentence-transformers
Milestone: You can fine-tune any HuggingFace model for classification, NER, or generation.
Phase 3: Advanced Topics (Weeks 9-12)
Week 9: Generative Models
Read:
- Autoencoders -- vanilla AE, VAE, ELBO, reparameterization trick
- GANs -- minimax, mode collapse, WGAN-GP, conditional GAN
- Diffusion Models -- DDPM, Stable Diffusion, LoRA
Exercise:
- Implement VAE from scratch on MNIST
- Fine-tune Stable Diffusion with LoRA
Milestone: You can implement VAE and GAN from scratch and explain diffusion math.
Week 10: Advanced Computer Vision
Read:
- Object Detection -- R-CNN family, YOLO, DETR, mAP
- Image Segmentation -- U-Net, DeepLab, Mask R-CNN, SAM
- Transfer Learning -- feature extraction, fine-tuning, CLIP, few-shot
Exercise:
- Train YOLOv8 on a custom dataset
- Implement U-Net from scratch for medical imaging
Milestone: You can train object detectors and segmentation models on custom data.
Week 11: Large Language Models and Alignment
Read:
- Language Models (revisit scaling laws)
- Text Generation (revisit RLHF and DPO)
- Papers Reading List -- read Transformer through DPO papers
AI engineering cross-reference:
- Fine-Tuning -- practical fine-tuning guide
- AI Guardrails -- safety and alignment in practice
Exercise:
- Fine-tune an open LLM with LoRA on a custom instruction dataset
- Implement DPO on a small preference dataset
Milestone: You can train a small LM from scratch and fine-tune open LLMs with LoRA/DPO.
Week 12: Multimodal Models and Reinforcement Learning
Read:
- Multimodal Models -- CLIP, VQA, image captioning
- Reinforcement Learning -- MDP, Q-learning, DQN, PPO
- Graph Neural Networks -- message passing, GCN, GAT
Exercise:
- Build an image search system with CLIP
- Train DQN on CartPole
Milestone: You can build multimodal systems and train RL agents.
Phase 4: Production (Weeks 13-16)
Week 13: Model Optimization
Read:
- Model Optimization -- pruning, quantization, distillation, ONNX, TensorRT
Data pipeline cross-reference:
- Preprocessing Pipeline -- production data preprocessing
- Data Contracts -- schema contracts for ML pipelines
Exercise:
- Quantize BERT to INT8 and measure speedup vs accuracy
- Export a model to ONNX and benchmark inference speed
Milestone: You can optimize a model for production (2-4x speedup with minimal quality loss).
Week 14: Deployment and MLOps
Read:
- DL Checklist (revisit deployment sections)
- Model Serving -- TorchServe, Triton, FastAPI
- GPU Kubernetes -- GPU scheduling
- ML Pipelines -- end-to-end ML pipeline design
- AI Testing -- testing ML systems
EDA cross-reference:
- Data Drift -- detecting distribution shifts in production
- Data Quality Validation -- automated quality checks
Data pipeline cross-reference:
- Pipeline Monitoring -- monitoring data pipelines
- Great Expectations -- data validation framework
Exercise:
- Wrap a model in FastAPI, containerize with Docker, deploy
- Set up model versioning with MLflow
Milestone: You can deploy a model as an API endpoint with monitoring.
Week 15: Capstone Project
Choose one end-to-end project:
Option A: Image Classification Pipeline
- Collect custom dataset
- Train CNN or ViT with full training recipe
- Optimize (quantize + prune)
- Deploy as web API with monitoring
Option B: NLP Pipeline
- Fine-tune BERT for real-world classification
- Build semantic search with sentence-transformers
- Quantize for production
- Deploy with FastAPI + monitoring
Option C: Generative AI
- Fine-tune LLM with LoRA
- Add safety filters
- Deploy as chat API
- Evaluate with human preferences
Week 16: Portfolio and Interview Prep
Read:
- Papers Reading List -- review 30 must-read papers
- Architecture Selection Guide -- justify architecture choices
- Algorithm Selection Guide -- classical ML decisions
- ML Checklist -- production ML checklist
- ML Interpretability -- explaining model decisions
Activities:
- Write up capstone project
- Create portfolio with 3-5 projects
- Practice explaining projects in 2 minutes
What You Will Be Able to Do After This Path
- Implement any neural network architecture from scratch in PyTorch
- Train and fine-tune models for vision, NLP, and multimodal tasks
- Apply model optimization (quantization, pruning, distillation) for production
- Deploy models with monitoring, versioning, and automated retraining
- Read and implement research papers
- Build end-to-end ML pipelines from data to deployment
Cross-References to Related Paths
- AI/ML Engineer Path -- LLM integration, LangChain, RAG, agents
- Data Scientist Path -- Math foundations, EDA (69 pages), statistics
- Data Engineer Path -- Data pipelines (25 pages), orchestration, quality
- Backend Engineer Path -- APIs and infrastructure for ML systems
- All Deep Learning pages: Deep Learning Overview -- index of all 25 topics
- All Machine Learning pages: Machine Learning Overview -- index of all 30 topics
- All EDA pages: EDA Overview -- index of all 69 topics
- All Data Pipeline pages: Data Pipeline Overview -- index of all 25 topics