Here is a brief summary of topics discussed, thus far, in classes. Slides and coding scripts would be distributed after each lecture via the Google groups mailing-list. Refer to the home page on how to subscribe.
06/10/2020 -- Introduction to Advanced Machine Learning. The first class introduces Machine Learning, its applications and main research areas, with particular reference to Computer Vision, as well as Deep Learning.
13/10/2020 -- Basic concepts in Computer Vision and Digital Image Filtering. Topics include: basic terminology of Computer Vision, including image formation, cues for 3D reconstruction, object detection and recognition; linear filtering, including convolutions, smoothing and sharpening.
16/10/2020 -- Smoothing as inference and multi-scale image representations. Topics include: filtering for reducing noise, additive i.i.d. noise, box and Gaussian average filters, separable filters, template matching, multi-scale Gaussian pyramid, edge detection, first and second derivative filters.
20/10/2020 -- Edge detection and Hough Transform. Topics include: image gradient, magnitude and orientation; thinning, Canny edge detector; second-order derivative for edge detection, the Laplacian operator and the Laplacian Pyramid; Hough transform, re-parameterization, generalized Hough Transform.
23/10/2020 -- Object recognition with color. Topics include: challenges in object recognition, image representation with histograms of colors, distance measures, nearest neighbor recognition, performance evaluation.
27/10/2020 -- Performance evaluation and image classification. Topics include: precision, recall, true positive rate, false positive rate, F1 measure and overall accuracy, ROC and PR curves and the area under the curves; log-loss, brier score and model calibration; image classification, L1 distance between images, k-nearest neighbor classification.
30/10/2020 -- Linear classification. Topics include: review of k-Nearest Neighbor, L1 and L2 distance; Hyperparameter tuning; Train and test sets and cross-validation; Linear classification with a parametric approach; Algebraic, visual and geometric interpretations.
03/11/2020 -- Loss functions and regularization. Topics include: multi-class SVM ("hinge") loss, softmax and cross-entropy loss, L1, L2 and elastic regularization .
06/11/2020 -- Optimization and Backprop. Topics include: gradient descent, including the numerical Vs analytical computation of the gradient; numerical gradient for "gradient check"; variants of gradient descent, including stochastic and mini-batch gradient descent; image features and end-to-end training of features; the computational graph; backpropagation via upstream gradients and local gradients for scalar variables; main patterns in gradient flow.
10/11/2020 -- Backprop continued. Topics include: backpropagation with vectors and matrices, by means of the Jacobian matrix and by means of implicit multiplications.
13/11/2020 -- Neural networks. Topics include: review of backpropagation with vectors and matrices; introduction of neural networks, including fully connected layers and activation functions; feature extraction Vs end-to-end learning; interpretation and visualization of intermediate network features; introduction to HW for deep neural networks, including CPUs, GPUs and TPUs; introduction of PyTorch.
17/11/2020 -- Pytorch. Topics include: the fundamental concepts of tensors, autograd and modules, coding examples for defining new modules, adopting nn.functional losses and using data loaders; dynamic Vs static graphs, including applications and frameworks relating to Pytorch; Introduction to CNNs, including applications and the concepts of local patterns and shared parameters and pooling.
20/11/2020 -- ConvNets. Topics include: convolutional layers and the hyperparameters of stride, pad and filter sizes; pooling layers and the hyperparameters of stride and filter sizes; the importance of depth; introduction to the course final projects.
24/11/2020 -- Semantic Segmentation. Topics include: visualization of convnets; transfer learning; things Vs stuff; (max) unpooling and transpose convolutions; intersection over union; efficient labelling for semantic segmentation.
27/11/2020 -- Object detection. In the lecture there has been discussion on object detection including details on classification and bounding box regression losses, selective search and RCNN .
1/12/2020 -- Object Detection and Instance Segmentation. Topics include: Fast and Faster R-CNN, as well as their building blocks RoI Pool, RoI Align and the Region Proposal Network; one-stage object detectors; instance segmentation with Mask R-CNN; the COCO benchmark; introductory notes on image captioning and 3D detection.
4/12/2020 -- Temporal Convolutional Networks and Recurrent Neural Networks. Topics include: introduction to sequence modelling and related computer vision applications, with specific reference to forecasting; Temporal Convolutional Networks (TCN) and the logarithmic interaction distance with convolution with dilation; Recurrent Neural Networks (RNN), including their computational graph, learning and inference, as well as interpretability of internal state cells; image captioning.
11/12/2020 -- More on RNNs. Topics include: gradient backpropagation for RNNs; image captioning with attention; brief introduction of visual question answering.
18/12/2020 -- LSTMs and Transformer for images and for forecasting. Topics include: long short-term memory models (LSTM); the role and modelling of social interaction, head pose and contact in people trajectory forecasting; transformer networks, self-attention and image transformers.
22/12/2020 -- Object, person and semantic search and meta-learning. Topics include: concluding remarks on relative self-attention and the music transformer; masking, training and fine-tuning BERT; visual object search with local, global and deep representations; triplet loss and siamese networks; re-identification with triplet loss and siamese networks; person search with OIM and QEEPS; semantic search with visual and text embeddings; introductory notes to meta-learning.