Course: Deep Learning for Bioscientists

Well done on completing week one of the course. In this week we have covered: key differences between machine learning and deep learning a brief look at the software tools …

PyTorch does contain some ‘off the shelf’ networks ready for use, but here we will look at how to build a simple convolutional nearal network from scratch. In the linked …

Tensors are the data type used to transmit data through deep learning networks. In the linked Colab notebook we look at: how to make tensors in PyTorch arithmetical operations with …

A quick overview of the important activation functions used in deep learning, in particular ReLU. Much of the power of deep learning networks comes from non-linearity. Loosely speaking, non-linearity is …

A recap of max pooling, what it does, and why we might need it. As we have seen in the previous videos and articles, the output of convolutional layers, especially …

As the name suggests, convolutional layers are the building blocks at the heart of convolutional neural networks. But what are they, and how do they work? Convolution If you completed …

Bounding box methods are a method that can provide information not only on the class of objects within an image, but also their location. As we describe in the video …

With multi-task learning, we can use deep learning to solve multiple tasks at once using the same network. In the video, we explain in more detail how this is done. …

What is heatmap regression? As the name suggests, it’s basically the output of a regression spread over the spatial dimensions of an image. So rather than a binary output as …

How do we use encoder-decoder networks to segment images? In this video we look more closely at how encoder-decoder networks can be trained to segment images, as well as discussing …

What are encoder architectures? Broadly speaking, these are deep learning network architectures that take detailed spatial data, such as digital images, and encode that data into a much smaller set …

A common approach to extracting spatial features from images are so-called encoder-decoder architectures. While classification networks aim to take the information in an image and compress it down to a …

How and why are network architectures chosen? Many different CNN encoder architectures have been proposed, in this video we mention just a few of the most common, and important to …

What are hyperparameters? As you may know if you’ve studied any machine learning techniques before, hyperparameters are the algorithm parameters that are set prior to training and do not change …

We’ve seen how convolutional neural networks are constructed, but how are they trained? In this video we go over the important training steps, specifically: the forward pass loss evaluation the …