AI & Deep Learning From Scratch In Python

Acquire fundamental python and numpy skills for deep learning, including data types, control structures, and array operations, to implement a cnn from scratch with convolution, pooling, and fully connected layers.

Explore object oriented programming in Python by building a conv layer class that encapsulates filters and biases, uses self for attributes, and implements forward propagation with convolution and ReLU.

Explore the derivative of a function, including product, quotient, and power rules, and the chain rule. Apply these tools to deep learning, velocity interpretation, partial derivatives, and gradients.

Explore derivatives by the chain rule with composite functions and a neural network example, deriving gradients and backpropagating to update weights in a fully connected layer.

Master backpropagation in fully connected networks, tracing the forward pass, loss, and cost to compute gradients for weights, biases, and inputs with ReLU and mini-batch gradient descent.

Learn how the softmax activation transforms the last-layer outputs into a normalized probability vector for classification, including the max-subtraction trick to prevent numerical overflow.

Leverage transfer learning to reuse pre-trained models for new tasks, freezing learned features and training only the new layers to save data, time, and resources, especially with limited data.

Initialize weights and biases to ensure stable gradients. Compare Xavier and He initializations, explain their use with sigmoid/tanh and ReLU in CNNs, including fan-in, fan-out, and zero biases.

Explore the Adam optimizer, combining momentum and rmsprop to provide per-parameter adaptive learning rates, with moment estimates and bias corrections across CNNs and fully connected layers.

This lecture demonstrates the YOLO v1 algorithm, detailing the cost function, gradient computation, and a CNN architecture for multi-object detection with IoU and mean squared error loss.