Welcome back to Beyond the Pattern.

So it has been a while since the release of the last video and we still had a wonderful

and wonderful talk that I don't want to keep from you.

So we had a guest virtually from South Korea in our lab that was Jung-Hoon Choi and he

received his bachelor's and master's degrees in electrical engineering and computer science

from Seoul National University in Seoul, South Korea in 2003 and 2008.

And he received a PhD degree from the University of Maryland, College Park in 2015 under the

supervision of Professor Larry S. Davis.

He is currently an associate professor of Yonsei University in Seoul, South Korea and

During his PhD has worked as a research intern in US Army Research Labs, Adobe Research Labs,

Disney Research Pittsburgh and Microsoft Research Redmond.

He also was a senior researcher at Comcast Applied AI Research Washington, DC and a research

scientist at the Allen Institute for Artificial Intelligence Seattle, Washington.

He is currently focusing his research on visual recognition and weekly supervised data for

semantic understanding of images and videos and visual understanding for edge devices

and household robots.

We have had the great pleasure to have had an introduction by him on continual learning

and in particular continual learning in practical scenarios.

I think Hyeonjin, he really did a wonderful job in summarizing the state of the art and

give directions into the future of research.

And I think this talk is definitely a must see if you're interested in continual learning.

So without further ado, Jeonghyun, the stage is yours.

Thank you.

Thank you very much for a very warm introduction of me, Andreas.

Hello, my name is Jeonghyun and I'm studying computer vision and machine learning at Yonsei

University, South Korea.

Today I'm going to talk about, as Andreas introduced, continual learning in practical

scenarios.

Okay.

So I'm very glad to see you in all online, not in person yet, but I'd like to visit you

at some point to discuss about the research and something else.

Okay.

So most of AI, you probably know that we need large scale label data set and we have some

deep neural network architecture to do some task.

For example, in this case, we are talking about image classification.

But one thing that everyone is missing is that we are only considering that this kind

of training is happening once.

We call this kind of one batch of training using the notion of data set.

But as a human, we are learning something continuously from the environment and some

information that we are facing every day.

So as a human, we get to know something new and also we forget something that we have

learned in the old days.

And we are just continuously updating our knowledge base throughout the time, in the

course of time.

But in the machine learning, we're not having that kind of notion that we can continuously

update the model.

We are not continuously updating knowledge that we have once learned in the past.

So I think it is more futuristic notion of learning for the machines and also for the

humans.

We may have this kind of loop in the future that human can provide more data and this