Скачать или смотреть NLP IL Meetup @ Big Panda #1 - Yonatan Hadar, Creative use of transfer learning for non trivial data

Deep learning has become the state-of-the-art solution for almost every machine learning problem. Finding a bird in an image, translating a sentence from Chinese or even playing poker and GTA, all of these can be done with deep neural networks.
But, deep learning has two main disadvantages that makes them impractical for many real world applications, it usually requires a large set of labeled data points and a large portion of expensive compute power. This is where transfer learning and models like BERT and GPT3 come to use. By using knowledge from a network that was trained on a large dataset we can train a very accurate network with a small dataset and limited compute power. For example, researchers in a document classification task, showed that the accuracy of a model that was trained on 30K labeled documents can be matched with transfer learning and 100 labeled documents.

In this talk I will share my experience on the effectiveness of using transfer learning on different machine learning tasks with some real life examples.
I will focus on how to creatively find or create large labeled datasets for transfer learning even on non-trivial data types or domain specific tasks.


Yonatan Hadar · Data Scientist at BigPanda
Yonatan is a data scientist with machine and deep learning experience in Natural language processing, Time series analysis, Recommendation engines, and more domains.
He has an MSc. in Industrial Engineering with a focus on NLP and currently working as a senior data scientist at BigPanda where he focuses on self-supervised and weakly supervised time series and NLP problems.


You are all invited to join the NLP IL community at:
Facebook: bit.ly/nlp-il
LinkedIn: bit.ly/nlp_il_linkedin
Talk Proposal Form: bit.ly/nlp-il-talk
YouTube Channel: bit.ly/nlp_il_youtube

Contact us at [email protected]