Cohere For AI - Guest Speaker: Dr. Saquib Sarfraz, Deep Learning Lead

Abstract: The recent surge in large language and vision models necessitates powerful unsupervised tools for data comprehension and model development. Clustering and data visualization techniques play crucial roles in tasks like retrieval-augmented generation, open-world recognition, and embedding quality monitoring. This talk provides an overview of our previous work in this domain, beginning with revisiting the FINCH algorithm, a parameter-free clustering approach known for its speed and ability to reveal natural hierarchical data groupings. . Following this, the discussion encompasses prevalent dimensionality reduction techniques for visualization, such as t-SNE and UMAP, and provides a bit deeper look into a new algorithm, Hierarchical Nearest Neighbor Graph Embedding (h-NNE). Both h-NNE and FINCH excel in speed, scalability, and simplicity compared to existing methods. This translates to faster runtimes and effective exposure of inherent clustering structures within large-scale unlabeled data, making it particularly valuable for visualization and clustering tasks.
Paper: FINCH https://arxiv.org/abs/1902.11266 Github: https://github.com/ssarfraz/FINCH-Clustering
Paper: h-NNE https://arxiv.org/abs/2203.12997 Github: https://github.com/koulakis/h-nne

About the Speaker: Saquib Sarfraz is working as Lead Deep Learning at Mercedes-Benz Tech Innovation. He has more than 15 years of research and teaching experience in AI and machine learning. He works extensively in R&D support within Mercedes-Benz. Currently he also shares some of his time at Karlsruhe Institute of Technology (KIT) where he is engaged in teaching graduate courses, research and PhD students supervision.

Add event to calendar

Apple Google Office 365 Outlook Outlook.com Yahoo