Deep Neural Networks
This course builds deep neural networks from the ground up: the smallest mathematical units (neurons), how they relate to linear and logistic regression, then layers, depth, training with gradients, practical patterns for dense architectures, and a NumPy from-scratch capstone tied to the lexai-fellowship notebook.
Prerequisites
Foundations of Regression
Linear regression, gradient descent, sigmoid, log-likelihood
Lessons
From linear regression to the perceptron
A neuron is linear regression with a switch
Layers in a deep neural network
Stack neurons, get intelligence
Activation functions
Non-linearity is the key to learning anything
Loss functions
The loss defines what "better" means
Forward pass — how networks predict
Trace a number from input to output
Backward pass — how networks learn
Gradients flow backwards, weights update
Trainable parameters & hyperparameters
Know which knobs the model sets vs which you set
Overfitting in neural networks
The model memorises instead of generalising
Neural network architecture
Architecture is where engineering meets design
Neural network from scratch (NumPy)
If you can build it, you truly understand it
Unlocks
Deep Sequence Modelling — RNN
Apply these building blocks to time series and language
Deep Computer Vision — CNN
Specialise these layers for image recognition
Attention Is All You Need
Progress toward the transformer, built on the same foundations