Paper Talks - Emergent Abilities of Large Language Models

Hello,

We are happy to invite you for an event organise by our sponsor - GetInData | Part of Xebia on February 2nd, 2023!

What is Paper Talks?

Paper Talks is an initiative created by their Advanced Analytics team. It was born from the need to follow the newest achievements in the machine learning world and stay up to date with state-of-the art developments. The field of interest is around different hot topics in data science, especially machine learning and deep learning.

For whom?

Everyone interested in data science and machine learning is welcome to join the discussion, however, we used to be pretty curious about the details and did many deep dives, so the atmosphere is definitely more technical and academic than business oriented. If you are a data scientist or a machine learning engineer and you work either in industry or academia, surely you will be able not only to get something interesting from it for you, but also to contribute to the discussion by sharing your ideas and experience.

Formula nad Paper

It is a one-hour Zoom meeting where we discuss different hot topics in data science, especially machine learning and deep learning. We encourage everyone to contribute to the meeting by adding comments and thoughts to the paper (you can find link here)

Session theme:

Emergent Abilities of Large Language Models

Scaling up language models has been shown to significantly improve performance and sample efficiency on a wide range of downstream tasks, which is quite a predictable outcome. So instead, this paper is going to discuss an unpredictable phenomenon that we refer to as the 'emergent abilities' of large language models. Such emergent abilities can be defined as abilities that occur in models above a certain size and do not occur in smaller models. The cause of emergence has not yet been conclusively determined – if at all there is a single cause.

However, recent research by Google on emergent abilities has shed some light on this phenomenon. It turns out that there are more than 100 examples of emergent abilities that have already been empirically discovered by scaling language models such as GPT-3, Chinchilla and PaLM. The results of this research raise many new questions. Are such leaps the exception? Or are we just scratching the surface of the depths of large language models? And finally, can such undiscovered capabilities also be found in other AI models?

Read more about the Paper Talks here

Regards,

Klaudia from WDTT