Deep Learning
Neural networks end to end: backprop, PyTorch, CNNs, RNNs, LSTMs and transformers — from perceptron to attention.
What Is Deep Learning?
Why deep learning took over AI: representation learning, scale, and the hardware that made it possible.
The Perceptron & Linear Units
Start with the neuron: linear combination, threshold, and the step up to learnable weights.
Activation Functions
Nonlinearity is what makes deep networks expressive — sigmoid, tanh, ReLU and friends.
Forward Propagation
Push data through the network layer by layer and understand the shapes at every step.
Backpropagation
The chain rule in action: propagate errors backward to compute gradients for every weight.
Loss Functions
Pick the right objective: MSE for regression, cross-entropy for classification, and why it matters.
Optimizers: SGD, Momentum & Adam
From plain SGD to Adam — how optimizers navigate loss landscapes and converge faster.
PyTorch Tensors & Autograd
The PyTorch fundamentals: tensors, dtypes, devices, and the autograd graph.
Building an MLP in PyTorch
Assemble a multi-layer perceptron with nn.Module, layers and activations — the skeleton of every modern net.
The Training Loop
Write the loop that every training run shares: batches, forward, loss, backward, step.
Regularization for Deep Learning
Dropout, weight decay, early stopping and data augmentation — fight overfitting without shrinking capacity.
Convolutional Networks
Why convolutions beat dense layers for images: local structure, weight sharing, and translation invariance.
Classic CNN Architectures
LeNet, VGG, ResNet — the ideas (depth, skip connections) that shaped modern vision models.
Transfer Learning
Stand on the shoulders of pretrained models: freeze features, swap the head, and fine-tune.
Recurrent Neural Networks
Process sequences with a hidden state that carries context through time.
LSTM & GRU
Gated memory cells that actually remember — the fixes that made sequence modeling practical.
Transformers
The architecture behind modern AI: self-attention, positional encoding, and parallelized sequence processing.
Attention Mechanisms
Query-key-value attention, multi-head splits, and how attention replaced recurrence.
Training at Scale
GPUs, mixed precision, data loaders and distributed training — the practical side of big models.
Evaluating Deep Learning Models
Accuracy is not enough: confusion matrices, calibration, and failure-case analysis for neural nets.
Deep Learning Roadmap
Synthesize the course: choose a specialization (vision, NLP, generative), and plan real training projects.