MAI5063 Advances in Deep Learning

Lecture companion: a guided, interactive tour of loss, gradients, learning rate, batch size, momentum, Adam, schedules, weight decay and normalisation. Read it before the labs or return to it when training misbehaves.

Everyday examples and small demos for local windows, shared stamps, feature maps, stride, pooling, Dropout, BatchNorm and augmentation, with minimal maths. Read it before the CNN lab.

MLP, activation functions, loss, gradient descent, learning rate and overfitting on MNIST in Google Colab. Interactive demos plus the code for all nine tasks.

SGD, momentum, RMSProp and Adam, learning rate, batch size, scheduling and weight decay on Fashion-MNIST in Google Colab. Interactive demos plus the code for every task.

Convolution, pooling, Dropout, BatchNorm and data augmentation on CIFAR-10 in Google Colab. Interactive demos plus the code for every task.

The same tasks as Lab 3 on Fashion-MNIST: smaller grayscale images, a quicker download and faster training, so it runs comfortably on a CPU runtime.

VGG blocks, residual connections and a pretrained EfficientNetB0, with feature extraction and fine-tuning on Fashion-MNIST in Google Colab. Interactive demos plus the code for every task.

Use a pretrained VGG16 or VGG19 as a feature extractor, then classify with an SVM or Random Forest in Google Colab. Compare with raw pixels and with the fine-tuning in Lab 4.

Build IoU and Non-Max Suppression from scratch, explore precision and recall, then run a pretrained Faster R-CNN in PyTorch and TensorFlow in Google Colab.

Encode boxes to a YOLO grid, match anchors, then run YOLOv8 and SSD and measure speed against Faster R-CNN in Google Colab.