Data Machina #230 – Data Machina

Graph Deep Learning and GNNs. I’ve just read that DeepMind has developed a new kind of Graph NN for discovering millions of new materials. Nowadays, every time the AI Tech Giants publish a new research paper, the mainstream media goes wild as if everything in those papers were radically new, never seen before. Graph NNs have been around for many years. Here are few note on GNNs.

Intro to Graph Deep Learning. This is an excellent, comprehensive hands-on tutorial by Steve @TuringInstitute. The tutorial covers all the theory and fundamentals of GNNs, as well as the implementation and explainability of GNNs. Checkout the video below or the repo, slides here: Hands-on Tutorial on Graph Deep Learning.

The theory behind Graph NNs. Deep dive into the theory of GNNs. In this presentation, Peter @DeepMind, covers how to derive GNNs from first principles, the motivating for their use, and how GNNs have emerged along several related research lines.

Types and applications of Graph NNs. A great blogpost by Zhong in which he reviews the types of GNNs and their applications in a concise, clear way. At the end of the blogpost there is a nice FAQ section. Checkout: Graph Neural Networks: Extending Deep Learning to Graph-Structured Data.

Graph NNs and fraud detection. A prototypical use case of GNNs is fraud detection. In this presentation, Zhao & Qiao provide an overview of GNN-based fraud detection, applied to suspicious account detection in online marketplaces. They do a deep dive with PySpark and GraphFrames to build a graph in a scalable way and convert it to DGL (Deep Graph Library) format. They also share their experience of setting up training and inference graphs and deploying the e2e model pipeline in Airflow.

Graph NNs and clustering. Deep graph clustering, which aims to reveal the underlying graph structure and divide the nodes into different groups, is attracting intensive attention recently. This is a really awesome collection of state-of-the-art (SOTA), novel deep graph clustering methods (surveys, papers, code repos, and datasets). Checkout: ADGC: Awesome Deep Graph Clustering.

The latest on Graph NNs. If you are interested in learning about the latest on GNNs and Graph DL, you can attend the Learning on Graphs Conference 20223 (LoG.) Here is a live recording of LoG, Nov 27 2023. 6 hours streaming!

The promise of Graph NNs and Relational DBs. This is very interesting: training & inference at scale directly from relational data, and modeling the relations with GNNs. A few days ago, researchers at Stanford et. al introduced a new, e2e approach called Relational Deep Learning. The core idea is to view relational tables as a heterogeneous graph, with a node for each row in each table, and edges specified by primary-foreign key relations. Checkout the paper: Relational Deep Learning: Graph Representation Learning on Relational Databases and the accompanying slide deck: Relational Deep Learning: Where do graphs come from?

The long read for the w/e: Geometric Deep Learning. I enjoyed reading this blogpost. A long, detailed post on how we got from ancient Greek geometry to Graph NNs. Recommended, easy reading for the w/e: Towards Geometric Deep Learning.

Have a nice week.

1. Making Large Language Models Uncool Again

2. [free] Tübingen ML – The Probabilistic ML Lectures

3. Combining Bayes and Graph-based Causal Inference

4. Visualising Segment Anything Model (SAM)

5. USearch – Vector Search 10x Faster than Meta AI’s FAISS

6. Deep Dive into Vision Transformer (ViT) Paper

7. DPO vs. RL: Is RL Needed in RLHF?

8. Meta AI Seamless – Massive, Fast, Multilingual AI Translation

9. The $10mn AI Mathematical Olympiad Prize

10. LibreChat – An Enhanced ChatGPT Clone, Open-source, Self-hosting

Share Data Machina with your friends

1. Self-Operating Computer Based on Multimodal Models

2. MS TaskWeaver – An AI Agent for Seamless Data Analytics

3. unsloth – 80% Faster, 50% Less Memory Local QLoRA Finetuning

1. [free course] Local Explanations for Deep Learning Models 2023

2. Deep Learning Ultra – Open source DL Containers (DLCs)

3. Robin – Better Visual Reasoning by Merging Vision-Language Models

1. RIPOR – SOTA, Scalable, Efficient Generative Information Retrieval

2. Visual Anagrams: Optical Illusions with Diffusion Models (paper, code, demo)

3. Alibaba Animate Anyone: Img-to-Vid for Character Animation (paper, code, demo)

1. Just How Bad Is the US Cost-of-Living Squeeze?

2. Building an Interactive Temperature Weathergami

3. Visualising Survey Data with Interactive Likert Scales

1. [book] Designing ML Systems for Prod-ready Apps (pdf, 463 pages)

2. Streamlining MLOps: Using Git with LLMs

3. How to: Experiments with MLFlow & Github on GCP

1. Syrup – AI for Fashion Inventory Optimisation

2. PhysicsX – AI for Industrial Engineering

3. Kognitos – AI for Enterprise Automation with NLP

1. Stanford AIMI – A Collection of Datasets for Healthcare AI/ ML

2. Meta AI Ego-Exo4D: A Dataset for Research on Video Learning

3. ShareGPT4V – 1.2M GPT-4V Multi-modal Captions

Tips? Suggestions? Feedback? email Carlos

Curated by @ds_ldn in the middle of the night.