**Online Edition** Selected Community Topics

This meetup will be an online edition in an interactive format:
We'll have 3 lightning talks of each 10 min presentation followed by 10 min Q/A.

*** Agenda ***

• 19:00: Introduction
• 19:10 - 19:30: Manas Deb “AI Industrialization - Current Status Review”
• 19:30 - 19:50: Dominik Seisser “Should we be afraid of Transformers?”
• 19:50 - 20:10: Valentin Kuhn & Tobias Blei “Deep Defect Detection at Accso”
• 20:10: Wrap up

*** Abstracts ***

“AI Industrialization - Current Status Review”
Adequate industrialization is key to broad adoption of all emerging technologies, and this also holds true for a very-high-potential yet quite-a-low-adoption technology as AI. The presentation will describe the key components of technology industrialization and how they apply to AI, the current maturity of these components in the context of AI and their potential impact on broad adoption of AI.

“Should we be afraid of Transformers?”
We look at the latest advances in language models, a small step forward in machine learning towards our quest for more general artificial intelligence. Powerful language models based on the "Transformer" deep learning architecture have set new benchmarks in Natural Language Processing (NLP). The release of the "BERT" model has been described as the ImageNet moment of Computer Vision for NLP - with highly successful transfer learning where a pre-trained model can be fine-tuned for a wide range of tasks. We cover a brief history of how we learn ever more useful information from data culminating in latest advances. We also cover pointers on how you can tackle your next NLP project utilizing the power of new language models.

"Deep Defect Detection at Accso"
As a general-purpose IT consulting and software engineering company, Accso started investing more in machine learning as the field matures and transitions into traditional software engineering. In the course of building more know-how, we chose to participate in a Kaggle challenge.
By organizing this challenge, steel manufacturer “Severstal” aimed at improving the quality of their produced steel plates by detecting and classifying defects on images of steel plates. Thus, the Kaggle challenge covered image segmentation and classification tasks on the provided training images and segmentation labels. We implemented and trained multiple convolutional networks to solve this challenge. Furthermore, we tuned our pre-processing pipeline to improve model performance.
We will take you on our journey from several problems in the beginning to our working solution. Besides, we will go into a few implementation details of our original TensorFlow models and re-implementation in PyTorch.

Looking forward to "seeing" many of you!