5 Short Deep Learning Lessons
Details
We will be holding an intensive meetup with 5 lectures of 20-30 minutes each given by our researchers. The following topics will be covered:
-
Nir Levine will tackle the paper: "Rainbow: Combining Improvements in Deep Reinforcement Learning" by DeepMind. Nir will review the different improvements over the vanilla DQN version, showing each improvement contribution, together leading to a very significant gap from the original version
-
Shai Bagon will review loss functions. Starting from the classical cross-entropy loss and why unbalanced data "breaks" this loss and how to handle such situations using focal-loss and dice-loss.
3.Netanell Avisidris will review fast inference methods for CNNs. Inferencing CNNs can take a long time, specifically most of the time is consumed in the convolutional layers. Netanell will review the standard implementation (im2col) and show approximate and exact methods for speed-up convolutional layers.
-
Yossi Bitton will tackle the paper: "Deformable Convolutional Networks". Yossi will go over the intuition and implementation details of two specific modules: deformable convolution and deformable roi-pooling.
-
Assaf Mushinsky will review recent advancements in human body pose estimation using CNNs. The talk will discuss both the iterative method of pose estimation from "Convolutional Pose Machines" and the method for multi-person pose estimation introduced in "Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields".
This meetup is intended for advanced practitioners with domain familiarity.
About Brodmann17:
Brodmann17's patented deep learning technology enables everyday devices to understand their environments in real-time. The use cases are tremendous in scope and enable new areas and platforms, such as augmented reality, robotics, home security, smart cities, and autonomous vehicles, for the first time to deploy real-time high-performance, hardware-agnostic, video analysis at 5% of the computing cost or in turn to produce same detection abilities with low cost, low energy consumption processors. Its algorithms have already been integrated into several commercial products by leading global players.