Convolutional Neural Networks for Visual Recognition

This lecture introduces the historical challenges and foundational concepts of computer vision, from Camera Obscura to early feature extraction methods like SIFT and HOG, setting the stage for deep learning approaches.

This lecture formalizes the problem of image classification and introduces data-driven methodologies. It explores K-Nearest Neighbors and Linear Classifiers, establishing foundational concepts of hyperparameters and cross-validation for model evaluation.

This lecture introduces the formalization of loss functions, specifically multiclass SVM loss and multinomial logistic regression loss, for quantifying model error in image classification. It details regularization techniques like weight decay to combat overfitting and introduces stochastic gradient descent as a primary optimization algorithm.

This lecture transitions from linear classifiers to fully-connected neural networks, detailing the backpropagation algorithm for gradient computation. It briefly explores connections between artificial and biological neural networks.

This lecture introduces the foundational concepts of Convolutional Neural Networks (CNNs), tracing their historical evolution from early models like the Perceptron and Neocognitron to influential architectures such as LeNet and AlexNet. Key components including convolution, pooling, and fully-connected layers are formally defined.

This lecture details practical considerations for optimizing deep neural networks, examining various activation functions, data preprocessing strategies, weight initialization schemes, and the mechanism of batch normalization for stabilizing learning.

This lecture delves into practical aspects of deep neural network training, covering advanced optimization methods such as AdaGrad, RMSProp, and Adam, along with regularization strategies including dropout and data augmentation. It concludes with an introduction to transfer learning and fine-tuning methodologies.

In Lecture 8 we discuss the use of different software packages for deep learning, focusing on TensorFlow and PyTorch. We also discuss some differences between CPUs and GPUs.

This lecture surveys prominent convolutional neural network architectures. It details design principles and contributions of models like AlexNet, VGGNet, GoogLeNet, and ResNet, which have achieved success in the ImageNet challenges.

This session explores recurrent neural networks (RNNs) for sequence modeling, covering architectures like LSTM and GRU. It demonstrates their application in language modeling, image captioning, and the integration of soft spatial attention mechanisms.

This lecture extends convolutional networks to advanced computer vision tasks, focusing on semantic and instance segmentation using FCNs. It also examines object detection methods, including the R-CNN family, SSD, and YOLO.

This session delves into techniques for interpreting internal mechanisms of CNNs, such as saliency maps and feature inversion. It also covers generative applications like DeepDream and neural style transfer, showcasing image synthesis.

This lecture introduces generative modeling as a form of unsupervised learning. It elaborates on autoregressive models like PixelRNN/PixelCNN, traditional and variational autoencoders (VAEs), and generative adversarial networks (GANs).

This session formalizes reinforcement learning using Markov Decision Processes, policies, and value functions. It discusses key algorithms such as Q-Learning, policy gradients, and Actor-Critic methods, illustrating applications in games like Atari and AlphaGo.

Guest lecturer Song Han presents algorithms for accelerating deep learning inference and training, including pruning, quantization, and parallelization. The lecture also covers specialized hardware like GPUs, FPGAs, ASICs, and TPUs.

Guest lecturer Ian Goodfellow explains the phenomenon of adversarial examples in deep learning and their exploitation in attacking machine learning systems. The lecture explores defense strategies and the utility of adversarial training to enhance model robustness.