DataTalks #16: DEEP Dive into Recommender Systems & Neural Network Quantization
Details
When Deep Learning Meets Recommender Systems
Abstract
Recommender systems are widely used by e-commerce and services companies worldwide to provide the most relevant items to their users. Recent advances in deep learning have made them even more exciting. At Fiverr, it’s our mission to grow and cultivate the world’s largest freelancer digital marketplace. We use a recommender system to provide the best match between customers’ needs with the skills of our freelancers. In this talk, I’ll present some state of the art deep learning algorithms for recommender systems, and the recent implementation of the DeepFM algorithm combined with model uncertainty estimation for dynamic personalized banner selection at Fiverr.com. I’ll share some implementation of technical details and evaluation results.
Bio
Asi Messica is Head of Data Science (Marketplace) at Fiverr.com. Before joining Fiverr, she held various product and development management positions at SAP, RSA and more. She pursues her Ph.D. in Software and Information Systems Engineering at the Ben-Gurion University under the supervision of Prof. Lior Rokach. Asi holds a B.Sc. in Physics and Computer Science, and MSc in Physics from the Tel Aviv University, as well as an MBA in Marketing from Bar-Ilan University.
Same, Same But Different - Recovering Neural Network Quantization Error Through Weight Factorization
Abstract
Deep learning and neural networks have made a big leap in the recent years and becoming a
leading method for many applications. In parallel, we are witnesses to the emerge of inference
embedded processor that can efficiently run these applications on the edge devices. One of the
main challenges is Quantization, conversion of the arithmetic used within the
net from high-precision floating-points to low-precision integers, this is an essential step for efficient deployment, however, quantization degrades network performance.
In this session, we will analyze the sources of quantization degradation. We will present a
model that tries to explain how quantization affects the dataflow in the net and how it might
degrade its performance. Based on the model we will suggest methods to modify the net to
make it more robust to quantization noise and thus reduce the quantization degradation.
Bio
Eldad Meller, Senior Machine Learning Engineer at Hailo. He was previously a researcher and a team leader in a leading technological unit of the Israeli Defense Forces. Eldad earned a B.Sc and M.Sc in electrical engineering (cum laude) from Tel Aviv University.
