We would like to invite you to the next Allegro Tech Talks Meetup in Kraków. This time, we are holding a Machine Learning edition. We’d like to describe ML techniques that enable some of Allegro’s intelligent algorithms running in production: accurate image assessment and fast approximate search.
REMEMBER TO REGISTER!
WITHOUT QRCODE YOU WON'T BE ABLE TO ENTER THE MEETUP
Note: Both presentations will be delivered in English.
Place: Browar Lubicz, Lubicz 17J, Kraków, Poland
18:00 - registration + beer
18:30 - A case-study of CNNs and transfer learning for image quzlity assessment at Allegro - Ireneusz Gawlik, Tomasz Bartczak
19:30 - Approximate nearest neighbors search - Karol Grzegorczyk, Marcin Cylke
20:30 - networking
1. A case-study of CNNs and transfer learning for image quality assessment at Allegro – Ireneusz Gawlik, Tomasz Bartczak
As allegro.pl marketplace is getting more professional, content quality assessment becomes more and more important. We employed deep learning techniques to ensure proper quality of images, pushing offers that fail to fulfil our guidelines further down the search results.
During our talk, we will describe how we approached image quality assessment, starting with classical computer vision based solution and moving later to CNN based transfer learning approach. We will share lots of experience gathered during our journey – regarding dataset building, model training and evaluation, hyperparameters tuning, and last but not least, production deployment.
Ireneusz and Tomasz are members of Machine Learning Research team at Allegro. Tomasz has a strong Software Engineering background and made a shift towards the field of Machine Learning in recent years. Ireneusz has several years of experience with machine learning R&D projects in academia and has been developing intelligent algorithms at Allegro for a couple of years. He is also working toward his PhD degree at AGH UST.
2. Approximate nearest neighbors search – Karol Grzegorczyk, Marcin Cylke
Many dense vector representations of different kinds of data have been proposed recently. One of their applications is similarity search, a.k.a. nearest neighbor search. Due to the high computational complexity of exact methods, it is often necessary to rely on some approximation. In the first part of this talk, we will review and compare existing approximate nearest neighbors algorithms and solutions. In the second part, we will focus on learning to hash for text documents and we will present our research results in this area.
Marcin and Karol are members of Allegro’s recommender system team. Marcin is an engineer deeply rooted in software design and architecture, recently shifting his interests towards data science. Karol is interested in machine learning (embedding methods in particular), distributed systems, and big data.