วีดีโอ

In this video we introduce Long-Short-Term-Memory Networks (LSTMs), which were developed to address some of the difficulties in training vanilla RNNs. You can find the next video of the series here: th-cam.com/video/QaqICaFZ_bg/w-d-xo.html

In this video we introduce the backpropagation algorithm which is used for training RNNs. You can find the next video of the series here: th-cam.com/video/0B5Bmagjxx8/w-d-xo.html

In this video we discuss why training vanilla RNNs can be difficult and unstable. You can find the next video of the series here: th-cam.com/video/-Fy6CDkqgPg/w-d-xo.html

In this video, we discuss common design patterns of RNNs. You can find the next video of the series here: th-cam.com/video/HVXUQv2WLrI/w-d-xo.html

In this video we introduce RNNs and discuss their differences with feedforward neural networks. You can find the next video of the series here: th-cam.com/video/AV64R_GTSdY/w-d-xo.html

This section covers the role of pooling layers in CNNs, which reduce the spatial dimensions of feature maps, making computations more efficient while retaining key information. This is the last video of this series. You can find the first video of the next series, "RNNs", here: th-cam.com/video/A3o06Gi6KWs/w-d-xo.html

In this lecture, we explore advanced convolutional techniques such as 3D convolutions, transposed convolutions, and dilated convolutions. We briefly discuss their transformative role in enhancing the capabilities of convolutional neural networks for diverse and complex tasks.

Explore the benefits of convolutions, including their ability to leverage sparse interactions, share parameters across the network, and maintain spatial consistency through equivariant representations.

A detailed explanation of the convolution operation at the heart of CNNs, showing how kernels are used to extract critical features from input data by sliding across the image.

Discover the neuroscience-inspired origins of CNNs, tracing their development from Hubel and Wiesel’s research on visual perception to pioneering models like Fukushima’s Neocognitron and LeNet-5.

This lecture provides an overview of CNNs, explaining their layered structure and why they are particularly effective at processing image data and other grid-like structures.

In this video we discuss Dropout, a method for randomly dropping units during training that prevents co-adaptation of neurons, thereby increasing the robustness of deep neural networks. You can find the next video of the series here: th-cam.com/video/v4A9h2sXWaE/w-d-xo.html

In this video, we discuss regularisation cocktails, a method that combines multiple regularisation techniques, such as dropout and data augmentation, that can lead to stronger model performance, especially in tabular data. This is the last video of this series. You can find the first video of the next series, "CNNs", here: th-cam.com/video/qFEf-qS4D7E/w-d-xo.html

In this video, we'll explore ensemble techniques that combine multiple models to improve accuracy and minimise variance, especially in situations with data shifts. We'll also look at snapshot ensembles, a computationally efficient method that takes snapshots of model weights at different training stages before each restart, allowing us to achieve ensemble-level performance without significantly...