← Back to Dashboard
Module 03 · Days 15–21
Deep Learning Fundamentals

Neural networks without the mysticism. You will understand exactly what forward pass computes, what backprop derives, and why PyTorch became the industry standard. Everything connects to runnable code.

D15
Neural networks from scratch in NumPy
Build a 2-layer network: forward pass, loss, backprop. Do this once; debug NNs ten times faster forever.
D16
PyTorch: tensors, autograd & the training loop
Industry-standard DL framework. Tensors, autograd, DataLoaders, and a clean training loop.
D17
CNNs: convolutional layers to ResNet
Conv layers, pooling, receptive field, skip connections. Why convolutions are the right primitive for images.
D18
Transformers: the architecture powering all modern AI
Self-attention, multi-head attention, positional encoding. The single most important architecture in AI today.
D19
Regularisation, optimisers & hyperparameter tuning
Dropout, batch norm, weight decay, Adam vs SGD, LR schedules. Systematic search with Optuna.
D20
HuggingFace ecosystem: models, tokenisers & Hub
GitHub for AI models. Load pretrained Transformers, fine-tune in minutes, push to Hub. Foundation for all NLP work.
🏗
🏗️ PROJECT: Image Classifier with Transfer Learning + Gradio Demo
Fine-tune EfficientNet on a custom dataset. Grad-CAM explainability. Deploy as a live Gradio app.