Beyond the mainstream: NLP solutions in the age of LLMs

Language-related technological solutions have the potential to solve a plethora of problems in our everyday lives. At this event, we will dive into the challenges surrounding NLP in the age of Large Language Models (LLMs).

We'll look at:

• The role of human input in different languages;

• Acquiring and deploying LLMs at scale;

• Paving a path towards an equitable and inclusive future.

Through invited talks, we will go beyond mainstream, everyday solutions to explore novel opportunities and dazzling possibilities that state-of-the-art NLP tools can offer. Join this event to engage with complex questions and draw valuable insights from experts in the industry at the forefront of these technologies.

Target audience: This meetup will be relevant to ML engineers, researchers, data scientists, and anyone interacting with AI. Content is geared toward an audience with intermediate experience in AI.

Agenda

* The time is indicated in Bangalore time zone (GMT+5:30)

Does one size fit all or is there a need for domain-specific LLMs?

Massive-sized generic models such as GPTs, Bard, PaLM and others trained on a large generic corpus have shown exceptional promise and behavior beyond what they were originally designed for. GPT-4 is able to tackle multimodal information and can also perform complex reasoning using chain-of-thought. So what happens when these generic models meet highly specific industry vertical requirements? Can they perform the same way they do for other tasks?

We see numerous experimentations being performed with LLMs trained on domain-specific corpuses such as finance, medicine, legal, education, etc. These models are smaller in size (of the order of 50 B parameters) but trained on a much larger number of tokens than the generic models (compared to their parameter sizes). They have been consistently showing better performance than the generic models (which are 5-10x their sizes) on these industry-specific tasks. So should every industry vertical have a domain-specific model or are generic models good enough? What are the cost-performance tradeoffs? I will also share some thoughts on how the education industry, where we work, can possibly benefit from the best of both these worlds.

Scaling challenges of LLMs

LLMs have seemingly super-human capability that is magical and scary at the same time. Does that mean we have solved the problem of language understanding? Besides several theoretical and philosophical problems of LLMs, these models also face fundamental scaling challenges. In this talk, I will discuss three such challenges – deployment at scale, the related responsible AI issues, and serving the next 1000 languages. These are solvable, yet long-term problems which will require us to rethink the design strategies for LLM-based applications.

Low-resource, high-impact: Modern NLP solutions for under-resourced languages

None of the languages are equal until all of the languages are. In this session, we'll consider low-resourced languages and their significance in NLP, outlining the unique challenges that low-resourced NLP faces. We will examine a current state of a field; discuss the effectiveness of LLMs and the risks they entail, juxtaposing this in the context of languages other than English. We'll propose innovative approaches to tackle modern low-resourced NLP challenges based on harnessing the power of crowdsourcing.