Four modules, each built from short phases with a project to ship and a checklist to track progress. Each phase has high-level resources from Andrew Ng and DeepLearning.AI for intuition and hands-on practice, plus deeper resources (lectures, papers and books) for the maths and details. Each module builds on the one before it, so follow the order.
For students and engineers who already know classical ML and want to go deep on modern AI. Realistic timeline: about 8 months at 2 hours a day (~430 hours).
Module 1 — Deep Learning Foundations
Phase 1 — Math for Neural Networks (Week 1 · ~10h)
Linear algebra (vectors, matrices, tensors), calculus (partial derivatives, chain rule, Jacobians) and probability (distributions, MLE, cross-entropy, KL divergence). Learn just enough to derive backpropagation by hand.
High-level understanding (Andrew Ng and DeepLearning.AI: intuition, big picture and hands-on practice)
- DeepLearning.AI, Mathematics for Machine Learning and Data Science Specialization
- Andrew Ng, ML Specialization, Course 1 Weeks 1-2 (derivatives, gradient intuition, vectorization)
Deeper understanding (lectures, papers and books: the maths and implementation details)
Phase 1 project
Derive the gradients of a two-layer network on paper, then verify them numerically with finite differences.
Phase 2 — Perceptron and Feed-Forward Networks (FFNN) (Weeks 2-4 · ~32h)
Neurons, activation functions (ReLU, sigmoid, tanh, softmax), forward pass, loss functions, backpropagation, and weight initialization.
High-level understanding
- Andrew Ng, Deep Learning Specialization, Course 1: Neural Networks and Deep Learning
- Andrew Ng, ML Specialization, Course 2 Weeks 1-2 (neural networks, activations, training)
- DeepLearning.AI, PyTorch for Deep Learning Professional Certificate (networks, CNNs, transfer learning, transformers, diffusion, ONNX and quantization)
Deeper understanding
- Andrej Karpathy, Neural Networks: Zero to Hero (micrograd, makemore)
- Dive into Deep Learning (free interactive book)
Phase 2 project
Build an FFNN in NumPy only and train it on MNIST. Then rebuild it in PyTorch.
Phase 3 — Optimization and Regularization (Weeks 4-5 · ~18h)
SGD, momentum, RMSprop, Adam/AdamW, learning-rate schedules, batch and layer normalization, dropout, weight decay, and vanishing/exploding gradients.
High-level understanding
Deeper understanding
Phase 3 project
Compare optimizers and regularizers on the same model and plot the loss curves.
Phase 4 — Convolutional Neural Networks (CNN) (Weeks 5-7 · ~22h)
Convolution and pooling maths, receptive fields, padding and stride. Architectures and their variations: LeNet, AlexNet, VGG, Inception, ResNet, DenseNet, MobileNet and EfficientNet.
High-level understanding
- Andrew Ng, Deep Learning Specialization, Course 4: Convolutional Neural Networks
- DeepLearning.AI, PyTorch for Deep Learning Professional Certificate
Deeper understanding
Phase 4 project
Implement ResNet from scratch in PyTorch and train it on CIFAR-10.
Phase 5 — Recurrent Neural Networks (RNN) and Variants (Weeks 7-9 · ~20h)
Vanilla RNN, backpropagation through time, LSTM, GRU, bidirectional and stacked RNNs, seq2seq with encoder-decoder, and Bahdanau/Luong attention.
High-level understanding
Deeper understanding
Phase 5 project
Train a character-level language model with an LSTM, then a GRU, and compare them.
Module 2 — Natural Language Processing and LLMs
Phase 6 — Text Processing and Embeddings (Weeks 10-11 · ~16h)
Tokenization (BPE, WordPiece), TF-IDF, Word2Vec, GloVe, FastText, and text classification baselines.
High-level understanding
Deeper understanding
Phase 6 project
Train word embeddings on a corpus and build a text classifier on top of them.
Phase 7 — Attention and the Transformer (Weeks 11-12 · ~20h)
Self-attention, multi-head attention, positional encodings, and encoder, decoder and encoder-decoder architectures.
High-level understanding
- DeepLearning.AI, NLP Specialization, Course 4: Attention Models (transformers, translation, QA)
- DeepLearning.AI, PyTorch for Deep Learning Professional Certificate
Deeper understanding
- Jay Alammar, The Illustrated Transformer
- Karpathy, Let's build GPT from scratch
- Paper: Attention Is All You Need
Phase 7 project
Build a mini-GPT from scratch and train it on a small text dataset.
Phase 8 — Pretrained Language Models (Weeks 13-14 · ~20h)
BERT, RoBERTa, GPT, T5 and BART. Fine-tuning with Hugging Face Transformers and the Trainer API.
High-level understanding
Deeper understanding
Phase 8 project
Fine-tune a BERT-family model (for Arabic, try AraBERT) for classification or NER and report proper metrics.
Phase 9 — Machine Translation and Question Answering (Weeks 14-16 · ~22h)
Translation with transformers (MarianMT, NLLB, mBART) and metrics (BLEU, chrF, COMET). Extractive QA on SQuAD, generative QA, and open-domain QA.
High-level understanding
Deeper understanding
Phase 9 project
Build an English to Arabic translator plus an extractive QA model, and evaluate both.
Phase 10 — Large Language Models (Weeks 16-18 · ~30h)
Pretraining and scaling laws, instruction tuning, RLHF and DPO, parameter-efficient fine-tuning (LoRA, QLoRA), quantization, prompting techniques, and LLM evaluation.
High-level understanding
- DeepLearning.AI, Generative AI with LLMs
- DeepLearning.AI, Finetuning Large Language Models
- DeepLearning.AI, Quantization Fundamentals with Hugging Face
Deeper understanding
Phase 10 project
Fine-tune an open model with QLoRA on a custom dataset and compare it with the base model.
Phase 11 — RAG Systems (Weeks 18-19 · ~20h)
Embeddings and vector databases, chunking strategies, hybrid search, rerankers, query rewriting, RAG evaluation (faithfulness, context recall), and an intro to agents and tool use.
High-level understanding
- DeepLearning.AI, Retrieval Augmented Generation (RAG)
- DeepLearning.AI, Advanced Retrieval for AI with Chroma
Deeper understanding
- Paper: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
- Ragas documentation (RAG evaluation)
Phase 11 project
Build an end-to-end RAG app over a real document set and publish an evaluation report.
Module 3 — Computer Vision
Phase 12 — Image Fundamentals and Classical CV (Week 20 · ~10h)
Pixels, color spaces, filtering, edges, histograms, feature detection (SIFT, ORB), and OpenCV.
High-level understanding
Deeper understanding
Phase 12 project
Build a document scanner with edge detection and perspective transform.
Phase 13 — Image Classification and Transfer Learning (Week 21 · ~10h)
Data augmentation, fine-tuning pretrained backbones, class imbalance, and Grad-CAM for explainability.
High-level understanding
- Andrew Ng, Deep Learning Specialization, Course 4: Convolutional Neural Networks
- DeepLearning.AI, PyTorch for Deep Learning Professional Certificate
Deeper understanding
Phase 13 project
Fine-tune a pretrained classifier on a custom dataset and add Grad-CAM visualizations.
Phase 14 — Object Detection (Weeks 22-23 · ~22h)
IoU, NMS, mAP. Two-stage detectors (R-CNN family), one-stage detectors (SSD, YOLO), and DETR.
High-level understanding
- Andrew Ng, Deep Learning Specialization, Course 4: Convolutional Neural Networks
- DeepLearning.AI, Advanced Computer Vision with TensorFlow (detection, segmentation, interpretability)
Deeper understanding
- Stanford CS231n, Deep Learning for Computer Vision
- Ultralytics YOLO documentation (detection, tracking, export)
- Paper: You Only Look Once (YOLO)
Phase 14 project
Train a YOLO detector on a custom dataset and report mAP.
Phase 15 — Image Segmentation (Weeks 23-25 · ~16h)
Semantic segmentation (FCN, U-Net, DeepLab), instance segmentation (Mask R-CNN), and Segment Anything (SAM).
High-level understanding
- Andrew Ng, Deep Learning Specialization, Course 4: Convolutional Neural Networks
- DeepLearning.AI, Advanced Computer Vision with TensorFlow (detection, segmentation, interpretability)
Deeper understanding
Phase 15 project
Train a U-Net on a medical or satellite dataset and compare it with SAM.
Phase 16 — Vision Transformers and Self-Supervised Learning (Weeks 25-26 · ~18h)
ViT, DeiT, Swin, DINO, MAE, and contrastive learning (SimCLR).
High-level understanding
Deeper understanding
Phase 16 project
Fine-tune a ViT and compare it with your ResNet from Module 1.
Phase 17 — Video, Tracking and Deployment (Weeks 26-28 · ~18h)
Object tracking (DeepSORT, ByteTrack), video classification, pose estimation, and exporting models with ONNX and TensorRT.
High-level understanding
- DeepLearning.AI, PyTorch for Deep Learning Professional Certificate
- DeepLearning.AI, Advanced Deployment Scenarios with TensorFlow
Deeper understanding
Phase 17 project
Ship a real-time detection and tracking app with an exported model.
Module 4 — Generative AI
Phase 18 — Autoencoders (Week 29 · ~12h)
Vanilla, denoising and sparse autoencoders, then VAEs (ELBO, reparameterization trick, latent space) and VQ-VAE.
High-level understanding
Deeper understanding
- Lilian Weng, From Autoencoder to Beta-VAE
- Paper: Auto-Encoding Variational Bayes
- Stanford CS236, Deep Generative Models
Phase 18 project
Train a VAE and show latent-space interpolation on faces or MNIST.
Phase 19 — Generative Adversarial Networks (GANs) (Weeks 30-31 · ~26h)
The minimax game, DCGAN, training instability and mode collapse, WGAN-GP, conditional GANs, Pix2Pix, CycleGAN and StyleGAN.
High-level understanding
- DeepLearning.AI, Generative Adversarial Networks (GANs) Specialization
- DeepLearning.AI, Generative Deep Learning with TensorFlow
Deeper understanding
Phase 19 project
Build a DCGAN, then extend it to a conditional or image-to-image GAN.
Phase 20 — Diffusion Models (Weeks 32-33 · ~20h)
Forward and reverse processes, DDPM, DDIM, classifier-free guidance, latent diffusion (Stable Diffusion), and ControlNet and LoRA for images.
High-level understanding
- DeepLearning.AI, How Diffusion Models Work
- DeepLearning.AI, PyTorch for Deep Learning Professional Certificate
Deeper understanding
- Hugging Face Diffusion Models Course
- Lilian Weng, What are Diffusion Models?
- Paper: Denoising Diffusion Probabilistic Models (DDPM)
Phase 20 project
Train a small DDPM on MNIST, then fine-tune Stable Diffusion with a LoRA.
Phase 21 — Multimodal Systems (Weeks 34-36 · ~28h)
CLIP and contrastive image-text learning, BLIP, vision-language models (LLaVA, Qwen-VL), text-to-image pipelines, speech with Whisper, and multimodal RAG.
High-level understanding
- DeepLearning.AI, Building Multimodal Search and RAG
- DeepLearning.AI, Prompt Engineering for Vision Models
Deeper understanding
- Paper: Learning Transferable Visual Models (CLIP)
- Paper: Visual Instruction Tuning (LLaVA)
- Hugging Face Audio Course (Whisper and speech)
Phase 21 project
Build a multimodal RAG app that answers questions over PDFs containing images.
Phase 22 — Capstone Project (Weeks 36-37 · ~24h)
Combine at least two modules in one deployed application, such as a multimodal assistant or a document-understanding pipeline.
Capstone project
Publish it on GitHub with a clear README, an evaluation, and a short demo video.