I am a mathematician; my doctoral studies involve exploring the areas of topology, geometry, and algebraic combinatorics( thesis).
Machine learning? Oh, I've dived headfirst into that fascinating pool! My journey began with the hat of a data engineer,
constructing models and crafting end-to-end pipelines for the AI industry. As I plunged deeper into the swirling currents of computer vision
and natural language processing algorithms, I realized I was absolutely smitten with the state-of-the-art work in these realms. And guess what?
My prior adventures in Topology and Geometry, especially the thrilling expeditions in Topological data analysis,
perfectly equipped me to blaze trails in the domains of Topological deep neural networks and Combinatorial complexes. Dive in with me!
github
linkedin
Papers
My research primarily focuses on exploring the intersection of Topological Data Analysis (TDA) with two advanced fields: Natural Language Processing (NLP), particularly in transformer architectures, and Generative Artificial Intelligence (AI), specifically in Generative Adversarial Networks (GANs).
In our recent study, we investigated the use of TDA for detecting out-of-distribution (OOD) data in natural language contexts. We applied TDA to attention maps in transformer-based language models, specifically evaluating our approach on BERT. Our method was compared against traditional OOD detection techniques based on BERT’s CLS embeddings. We discovered that our TDA-based approach excels in distinguishing in-distribution data (such as politics and entertainment news articles from Huff- Post) from significantly different out-of-domain samples (like IMDB reviews). However, its effectiveness diminishes with near out-of-domain (CNN/DailyMail) or same-domain (business news articles from HuffPost) datasets. These findings are detailed in our paper, you can read more in my blog post, where I also cite the paper