Image by BoliviaInteligente Over the past 15 years we have seen a surge of use of Graph Neural Networks (GNNs) being used to model social networks, recommendations systems, transportation networks, and many more systems. With the ever growing use of GNNs, this has naturally led to the questions about the use of GNNs within the medical sector. More specifically, the question of using GNNs to predict the properties of molecules was brought up. Back in the early 2010s, this idea was in its infancy with few successful applications.

[Image by Justin Lim] Introduction This article will delve into diffusion models, a group of latent variable (see definitions) generative models with applications in image generation, audio synthesis, and denoising. More specifically, this article will mostly focus on the derivations and the ideas behind diffusion models, with a heavy enthuses on the ideas introduced in Ho et al. in his Denoising Diffusion Probabilisitic Models paper (DDPMs). The applications of these models will not be covered today.

Image by Cajeo Zhang Graphs have been used across many fields due to their ability to represent relationships between entities with applications including social networks, search engines, and protein-protein interaction networks. However, one growing limitation of these graphs are the amount of computational resources they require with some large-scale graphs having millions of nodes each with their own set of features and their set of edges. This has led to the creation of graph embedding methods, more specifically the deep embedding methods. These embedding methods aim to create a high-quality representation of the nodes and their edges. Rather than just incorporating the graph structural information into an embedding, these methods also include node and edges features and other hierarchical information. This results in a complicated model which are able to learn very rich representations of nodes.

Image by Justin Simmonds Generative Adversarial Networks (GANs) were first introduced in 2014 by Ian Goodfellow in his paper “Generative Adversarial Nets.” This paper presented the GAN framework, which consists of two neural networks called the generator and the discriminator. The generator takes random noise as input and outputs a generated image. The discriminator takes both a generated image and a real image as inputs and tries to determine which is real and which is generated. Their training process can be likened to a ping-pong game, with the generator trying to produce images that fool the discriminator, and the discriminator is trying to identify which images are generated.

Image by Todd Quackenbush For the past 30 years, tree-based algorithms such as Adaboost and Random Forests have been the go-to methods for solving tabular data problems. While neural networks (NNs) have been used in this context, they have historically struggled to match the performance of tree-based methods. Despite recent advancements in NN capabilities and their success in tasks from computer vision, language translation, and image generation, tree-based algorithms still outperform neural networks when it comes to tabular data. This article will introduce several reasons behind the continued dominance of tree-based methods in this domain.