CLOUD COMPUTING SEMINAR SERIES
TITLE: Small is Beautiful: A Knowledge Centric Approach for Small Language Models
SPEAKER: Manoj Agarwal, GiKA.AI
DATE/TIME: Monday, Nov 4, 430-530PM
VENUE: CDS 102 Seminar Room
ABSTRACT: LLMs, although remarkable in generating seemingly intelligent answers, have no real understanding of the data they are trained on, i.e., the semantic understanding of the concepts and meanings behind the words (though they “simulate” that understanding). Hence, these models present a few fundamental challenges in their adoption for critical use cases such as:
- Hallucination: The hallucination is a foundational problem with these models, even though their world knowledge is deep and growing.
- Data Leakage: Giant language models have to be called via their APIs and this leaks the private and sensitive data to these centralised large models.
- Limited Context: For many use cases, the amount of data to be shared is larger than the context window, to make a meaningful inference. At the same time, a large context window, even if possible, may throw the model off track.
Besides, some of the other factors impacting the adoption of these models are lack of up-to-date information, latency and cost. Some recent approaches, such as Retrieval Augmented Generation (RAG), are proposed to handle a few of these limitations. In this talk, we present a novel approach which is knowledge centric instead of document centric, to address the challenges outlined above. Primarily, our approach comprises:
1. Understanding the data context: Given a corpus of documents, it remains a hard question to answer “What is this data about?”. Our first step is to semantically understand the user data. The data is parsed, processed, and cleaned and is used to improve the contextual semantic understanding of the small language model using a knowledge centric approach.
2. Knowledge Layer: The knowledge layer is the central component of our system. It captures the data taxonomy and the relationships between the entities and represents the data as knowledge graph. We propose a novel KG-RAG approach, that is highly effective and efficient to handle the problem of hallucination by grounding the small language model in the factual knowledge while also improving its reasoning capabilities.
3. Domain-Aware Semantic Query Engine: Semantic query engine fetches the factually accurate results by grounding the response in the data thus facilitating more meaningful, personalised and contextual search.
BIO: Dr. Manoj Agarwal is a co-founder of a deep tech company, GiKA.AI that aims to build next generation data intelligence systems for better semantic search and analytics. Before this, he was Senior Staff Engineer in Discovery intelligence team at Uber AI. In Uber, Manoj introduced the semantic search for Uber Eats. Besides, he worked on automatic enrichment of taxonomy using merchant data. Prior to joining Uber, Manoj worked as Principal Applied Scientist at Microsoft – AI and Research and as a senior researcher in IBM Research. Manoj was the chief architect for building a web scale product knowledge graph for Microsoft – Shopping, comprising a few hundred million products. Manoj also worked as adjunct faculty in IIT-Gandhinagar. Dr. Manoj Agarwal completed his PhD from IIT-Bombay where his thesis was awarded ACM India Doctoral Dissertation Award (Honorary mention). His research interests are in the areas of web mining, graph mining, pattern recognition, data mining, knowledge graphs, language models and information retrieval with more than 30 patents and over 25 research paper.
Host: Yogesh Simmhan
About: The IBM-IISc Hybrid Cloud Lab (IIHCL) hosted at IISc is curating the Cloud Computing Seminar series with guest speakers from Industry and Academia speaking about the latest technologies and research on Cloud and edge computing, distributed computing systems, and AI/ML/Big Data platforms. More details at: http://iihcl.iisc.ac.in .
ALL ARE WELCOME
======================================================================
