Build Strong Foundations Through Our Deep Learning Course in Abu Dhabi

Module 1: Deep Learning Foundations & Neural Network Basics

3 hrs

• AI vs. Machine Learning vs. Deep Learning — precise definitions and relationships
• Why deep learning now — the convergence of data, compute, and algorithms
• Landmark breakthroughs: AlexNet (2012), AlphaGo, GPT, and Stable Diffusion
• Hardware for deep learning — CPUs, GPUs, TPUs, and cloud acceleration
• Perceptron and MLP — weights, bias, activation, and forward pass
• Backpropagation and gradient descent — batch, stochastic (SGD), and mini-batch variants
• Activation functions — Sigmoid, tanh, ReLU, Leaky ReLU, PReLU, ELU, GELU, Softmax
• PyTorch — dynamic computation graphs, tensors, and autograd
• TensorFlow/Keras — the production workhorse with eager execution

• Building and training an MLP from scratch in PyTorch
• Understanding gradient descent and backpropagation mathematically
• Choosing between PyTorch and TensorFlow for research vs. production use cases
• Activity: Build and train an MLP from scratch in PyTorch on the MNIST digit recognition dataset

Module 2: Training Deep Networks — Optimization, Regularization & Tuning

3 hrs

• Momentum, RMSProp, Adam and AdamW — the default optimizers for most deep learning tasks
• Learning rate schedulers — step decay, cosine annealing, and warm restarts
• L1 and L2 weight regularization — penalizing large weights to prevent overfitting
• Dropout — randomly deactivating neurons during training as a form of ensemble
• Batch Normalization, Layer Normalization, and Group Normalization
• Data augmentation — creating artificial training variety to improve generalization
• Xavier/Glorot and He/Kaiming initialization — breaking symmetry and controlling gradient scale
• Diagnosing underfitting vs. overfitting using learning curves
• Hyperparameter tuning — grid search, random search, and Bayesian optimization

• Applying optimizers, regularization, and tuning strategies to achieve peak performance
• Diagnosing and resolving gradient vanishing, exploding, and underfitting issues
• Reading and interpreting training curves to guide model improvement decisions
• Activity: Train a deep network with and without regularization — measure and visualize the difference

Module 3: Convolutional Neural Networks (CNNs) for Computer Vision

4 hrs

• Convolution — filters, kernels, feature maps, local receptive fields, stride, and padding
• Pooling layers — max pooling, average pooling, and global average pooling
• Hierarchical feature learning process — edges to textures to objects
• CNN architectures — LeNet-5, AlexNet, VGGNet, Inception/GoogLeNet, ResNet, EfficientNet
• Object detection — YOLO, Faster R-CNN, and SSD architectures
• Semantic segmentation — FCN, U-Net, and DeepLab for pixel-wise labeling
• Instance segmentation — Mask R-CNN overview
• Vision Transformers (ViT) — applying transformer architecture to image patches
• Image preprocessing pipelines — resizing, normalization, and data augmentation
• Visualizing CNN features — Grad-CAM and activation maps for interpretability

• Building image classifiers, object detectors, and segmentation models with CNNs
• Applying transfer learning from landmark architectures to custom tasks
• Visualizing and interpreting CNN feature maps for explainability
• Activity: Train a ResNet model to classify medical X-ray images — achieve clinical-grade accuracy

Module 4: Sequence Models — RNNs, LSTMs, GRUs & Time-Series

3 hrs

• RNN architecture — hidden state, weight sharing, unrolling through time, and BPTT
• Vanishing and exploding gradients in RNNs — why they struggle with long sequences
• LSTM cell — forget gate, input gate, output gate, and cell state explained in detail
• Bidirectional LSTMs — reading sequences in both directions simultaneously
• Stacked LSTMs — deeper representations for complex sequential patterns
• GRU architecture — reset gate and update gate; GRU vs. LSTM comparison
• Preparing time-series data — windowing, stationarity, and feature engineering
• Sequence-to-sequence (Seq2Seq) architecture — encoder-decoder for forecasting
• Temporal Convolutional Networks (TCN) — CNN-based alternatives for sequences

• Designing LSTM and Transformer architectures for time-series and NLP tasks
• Applying Seq2Seq models for multi-step forecasting scenarios
• Comparing RNN, LSTM, GRU, and TCN to select the right architecture for each problem
• Activity: Build an LSTM network for multi-step stock price forecasting and energy demand prediction

Module 5: The Transformer Architecture — Attention and Beyond

4 hrs

• Limitations of RNNs — why sequential processing bottlenecks long-range learning
• Scaled dot-product attention — queries, keys, and values explained mathematically
• Self-attention — every token attending to every other token in the sequence
• Multi-head attention — learning diverse representation subspaces in parallel
• Positional encoding — injecting sequence order into attention without recurrence
• Encoder block (BERT), decoder block (GPT), and encoder-decoder (T5) variants
• Vision Transformer (ViT) — patch embeddings and global attention for images
• Swin Transformer — hierarchical attention with shifted windows
• FlashAttention, Rotary Positional Embeddings (RoPE), and efficient transformer variants
• Implementing multi-head attention in PyTorch — every line explained

• Building a Transformer from scratch in PyTorch for sequence-to-sequence tasks
• Understanding the attention mechanism and its application across vision and NLP
• Applying efficient transformer variants for scalable deep learning systems
• Activity: Build a Transformer from scratch in PyTorch and train it on a machine translation task

Module 6: Transfer Learning & Fine-Tuning Pre-Trained Models

3 hrs

• Feature extraction vs. fine-tuning — two fundamentally different strategies
• Domain adaptation — transferring knowledge across different data distributions
• Few-shot and zero-shot learning — generalizing from minimal examples
• Freezing and unfreezing layers — which layers to train and when
• Fine-tuning ImageNet models (ResNet, EfficientNet) for custom image tasks
• Learning rate differentiation — lower rates for early layers, higher for the head
• Hugging Face Transformers — loading, configuring, and fine-tuning BERT, RoBERTa, ViT
• Task-specific heads — adding classification, regression, and token labeling layers
• LoRA and QLoRA — fine-tuning billion-parameter models with minimal GPU memory
• Adapter layers and prefix tuning — adding task-specific parameters without touching weights

• Applying transfer learning and LoRA to adapt billion-parameter models to custom tasks
• Fine-tuning ViT and BERT models for image and text classification tasks
• Avoiding catastrophic forgetting with gradual unfreezing and discriminative fine-tuning
• Activity: Fine-tune a ViT model for satellite image classification and a BERT model for legal document tagging

Module 7: Generative Deep Learning — GANs, VAEs & Diffusion Models

4 hrs

• Variational Autoencoders (VAEs) — encoder, bottleneck, decoder, ELBO loss, and latent space interpolation
• GAN framework — generator and discriminator in adversarial training
• GAN training dynamics — Nash equilibrium, mode collapse, and training instability
• DCGAN, Progressive GAN, StyleGAN — state-of-the-art photorealistic face generation
• Conditional GAN (cGAN) — controlling what the generator produces
• CycleGAN — unpaired image-to-image translation
• Diffusion process — DDPM and the mathematical foundation
• U-Net denoising architecture — skip connections and cross-attention for conditioning
• DDIM — faster sampling with deterministic denoising
• Stable Diffusion — latent diffusion models and the role of CLIP conditioning
• Synthetic data generation and super-resolution with generative models

• Training GANs and diffusion models to generate photorealistic images and synthetic data
• Building text-conditioned image generation pipelines
• Applying generative models for data augmentation and super-resolution
• Activity: Train a GAN on a custom dataset and build a text-conditioned image generation pipeline

Module 8: Deep Learning for NLP & Large Language Models

3 hrs

• Tokenization strategies — BPE, WordPiece, SentencePiece, and tiktoken
• Word embeddings — Word2Vec, GloVe, FastText, and contextual embeddings
• Masked Language Modeling (MLM) and Next Sentence Prediction — BERT pre-training objectives
• BERT, RoBERTa, DeBERTa — evolution of encoder-only models
• Sentence-BERT (SBERT) — dense retrieval and semantic similarity at scale
• Fine-tuning BERT for classification, NER, and question answering
• Causal language modeling — autoregressive next-token prediction (GPT family)
• Instruction tuning and RLHF — how raw LLMs become conversational assistants
• Distributed training for LLMs — data parallelism, model parallelism, and pipeline parallelism
• LoRA, QLoRA, and adapters for parameter-efficient fine-tuning of custom LLMs

• Fine-tuning a GPT-2 model on a domain-specific corpus
• Building a text generation API and deploying an NLP model to production
• Applying distributed training and PEFT techniques for large language model customization
• Activity: Fine-tune a GPT-2 model on a domain-specific corpus and build a text generation API

Module 9: Deployment, Optimization & Production Deep Learning

2 hrs

• Quantization — INT8 and FP16 inference for 2–4x speedup with minimal accuracy loss
• Pruning — structured and unstructured pruning to reduce model size
• Knowledge distillation — training a small student model to mimic a large teacher
• ONNX — framework-agnostic model export for cross-platform deployment
• TorchServe, TensorFlow Serving, and Triton Inference Server — multi-framework GPU inference
• FastAPI wrappers — building lightweight REST APIs around any DL model
• TensorFlow Lite and PyTorch Mobile — on-device inference for Android, iOS, and microcontrollers
• ONNX Runtime — CPU and GPU inference across platforms

• Optimizing, quantizing, and serving models as APIs using TorchServe, Triton, and FastAPI
• Benchmarking inference speedup from quantization and pruning
• Deploying deep learning models to mobile and edge devices
• Activity: Quantize a trained model, benchmark the speedup, and deploy it as a REST API with FastAPI

Module 10: Capstone Project & Career Readiness

1 hr

• Option A — Computer Vision: train a custom object detection model and deploy as a real-time API
• Option B — Generative AI: build a text-to-image pipeline with a fine-tuned diffusion model
• Option C — NLP and LLMs: fine-tune a language model on domain data and build a RAG-powered chatbot
• Option D — Time-Series: build and deploy a multi-step forecasting system with LSTM and Transformer
• Deliverables: GitHub repo, training notebook, deployed FastAPI or Streamlit endpoint, project write-up
• Building a deep learning portfolio on GitHub, Hugging Face, and Papers With Code
• Writing a DL-focused CV — what to highlight and what hiring managers look for
• Acing deep learning interviews — theory questions, coding tests, and system design

• Completing a portfolio-ready capstone project showcasing end-to-end deep learning engineering
• Building a standout DL portfolio on GitHub and Hugging Face Hub
• Approaching deep learning interviews and research scientist roles with confidence
• Activity: Graduates receive a Deep Learning Specialist certificate and optional 1:1 portfolio review