Deep learning is evolving fast into something some would call deep learning 2.0. Perception system with learning-based sensor and temporal fusion – data pipeline with automatic training data generation – and more.
Zenseact, Embedl and Kognic work closely together, at the forefront of Deep Learning. Join us and listen to three keynote speakers who will share insights about where deep learning is headed and how it can become much more efficient. Their key notes will be followed by a panel discussion.
Join us in an after-work event May 11th 16.00 at Zenseact, Lindholmspiren 2, Gothenburg. Food and beverages will be available.
Deep learning in current generation of passenger cars and a peek into next generation
Convolutional deep learning (DL) networks is the fundamental type of deep learning for solving computer vision problems in state-of-the-art active safety systems for understanding the traffic scene. The fundamental problem is to make safer and more comfortable cars, ultimately saving people lives.We’lll explain how deep learning networks are used in the current generation to solve problems like object detection (e.g., pedestrians and other road users) and estimation of features of the road (e.g., lanes, freespace and road edges). The concept of DL2.0 will be explained where a larger part of the active safety software stack is performed with deep learning. Finally, the need of optimization of DL interference is explained.
Keywords: Deep learning, computer vision, active safety systems for passage cars
Making Deep Learning Efficient
Deep Learning has delivered some amazing results in computer vision, enabling advanced autonomous driving systems. But this comes at a cost - the top performing models are monsters that consume huge amounts of compute and energy resources. To make the fruits of DL more widely available, especially in embedded systems, we need to make them much more efficient. Embedl uses a combination of a variety of techniques from state of the art research, to address this challenge, including pruning, quantization and neural architecture search. We will give a taster of the research behind our technology in this talk.
Keywords: deep learning, pruning, quantization, neural architecture search
The future of developing datasets
Creating a good dataset is more than just annotating data efficiently. As the world changes, datasets need to be diverse and flexible. Isak Hjortgren, Product Area Lead at Kognic, will present how Kognic leverages human-machine collaboration for creating state of the art datasets for autonomous driving.
Keywords: Human-machine collaboration, feedback tools, dataset