Bloomberg is a new entrant to the LLM game! Alibaba comes up with new Hyperbolic deep learning system!
1st April 5:11 PM IST (India Standard Time) -
BloombergGPT -
Following the ChatGPT buzz, Bloomberg has launched BloombergGPT: A Large Language Model for Financial analytics (Source: https://arxiv.org/abs/2303.17564 ). I have not tried the model but after 3 coffees and reading the entire paper, the results are interesting. it was good to see the citations to
Geoff Hinton's 2011 paper (page 56) - Generating text with recurrent neural networks. This has been my Point of View for some time. RNN & LSTM-based text generation has existed so has GAN-based image generation. To ChatGPT’s credit; it is democratizing AI at an extremely large scale.Off course, Geoff's 1st paper on RNNs & Sequential text is as old as 1986 - Learning sequential structure in simple recurrent networks (Source: https://proceedings.neurips.cc/paper/1988/file/9dcb88e0137649590b755372b040afad-Paper.pdf)
Microsoft's 2020 paper DIALOGPT: Large-Scale Generative Pre-training for Conversational Response Generation
Based on the results discussed in the paper, the BloombergGPT model looks good. This does not mean that it will replace existing financial experts in any possible way.
Hyperbolic Deep Learning
I have mentioned GAN & Poincare Glove-based chatbots and Graph neural networks to analyze network data such as Doximity in the 2nd edition of my book Healthcare Social Media Management and Analytics. But, what Alibaba has achieved in terms of Hyperbolic Deep Learning is super interesting. Ant Group (Ant Financial) - a subsidiary of Alibaba - has performed some great work on graph neural networks and come up with a graph mining platform called AliGraph (which a lot of my Linkedin followers are aware of apart from Neo4j, AWS Neptune, etc.). Aligraph currently supports product recommendations and personalized searches on Alibaba's E-Commerce platform. Initially, it was trained on a real-world dataset with 492.90 million vertices, 6.82 billion edges, and latent features. (Source: https://dl.acm.org/doi/10.14778/3352063.3352127)
Comes Feb 2023, Ant Group published a new paper on graph neural networks called Poincaré Heterogeneous Graph Neural Networks for Sequential Recommendation. The user-item representation in actual online settings may be distorted by previous recommendation algorithms' handling of such hierarchical information, which included empirical user-item sectionalization in Euclidean space. Poincaré Heterogeneous Graph Neural Networks for Sequential Recommendation (PHGR) can concurrently describe the hierarchical and sequential pattern information present in the data. (Source: https://dl.acm.org/doi/abs/10.1145/3568395 )
Try using it in your work when you are building a recommendation system! I will write about my own 2009 research paper next time and how ChatGPT changed it :)
How to unsubscribe - https://support.substack.com/hc/en-us/articles/360059788812-How-do-I-unsubscribe-from-a-free-subscription-