Shulou(Shulou.com)06/02 Report--

Big Data Digest Column Works

Written by Christopher Dossman

Compile: fuma, Jiaxu, Yunzhou

Hello, everyone, this week's AI Scholar Weekly column meets you again!

AI Scholar Weekly is an academic column in the field of AI, dedicated to bringing you the latest, most comprehensive and in-depth overview of AI academics, and sweeping up the cutting-edge information of AI academics every week.

Monday update, do AI research, weekly from this article is enough!

Keywords of the week: autonomous driving, intelligent robots, GANs

This week's hot academic research

Robot Preschool

Robots need to learn to operate in intuitive and practical ways in order to deploy more successfully in real-world environments. To do this, the researchers trained a variant of ResNet that maps hand-eye camera images to end-effector velocities.

In their approach, human teachers demonstrate simple service-type tasks to robots via joysticks, and data collection, training, and deployment occur within an hour.

The contribution of this paper is a data-efficient and practical way to teach robots new behaviors in a very short time. All of this shows that robots can learn simple tasks quickly through practical demonstrations. Furthermore, the technique can be used to learn fast strategies for data-inefficient reinforcement learning.

Original text:

https://arxiv.org/abs/1905.09025

Driving Strategy of Auto-driving Vehicle on Freeway Based on Reinforcement Learning

For the first time, researchers have contributed to a real-time autonomous driving strategy based on reinforcement learning for vehicles traveling on highways. This strategy was implemented by deploying Double Deep Q-Network and was able to guide autonomous vehicles on highways.

Furthermore, the strategy can consider passenger comfort through a cleverly designed objective function. Ultimately, the strategy results in a collision-free trajectory where the autonomous vehicle will move forward at the desired speed while minimizing its longitudinal and lateral accelerations.

Autopilot engineers have had a great success in developing alternative path planning strategies that take into account multiple perspectives such as traffic density and safety.

This work is the first attempt by scholars to formulate reinforcement learning strategies for unrestricted highways. And provides a lot of insights into trajectory path planning problems. The generalization ability and stability of the new reinforcement learning strategy are also studied by using the established SUMO microscopic traffic simulator.

In terms of computational cost, the computational cost required to generate the action is significantly reduced and can be generalized to previously invisible driving situations. Unfortunately, at present this strategy does not guarantee collision-free trajectories.

https://arxiv.org/abs/1905.09046

Implementing Safety Awareness Software in Autonomous Vehicles

To date, there are no clear guidelines for security-aware AV (Autonomous Vehicles) computing systems and architecture design. This prompted the researchers to conduct field studies, including autonomous vehicle fleets, road conditions and traffic patterns in various regions. According to the research, traditional computing system performance indicators can not fully meet the security requirements of AV computing system design. Instead, they recommend using a "security score" as the primary metric for measuring the level of security an AV computing system is designed for. They also propose perceptual latency models to help architects estimate security scores for a given architecture and system design without physically testing it in AV.

The contribution of this paper is to propose a security scoring and perception delay model for security-aware AV computing system design. It details the detailed design and workload of AV computing systems and discusses the design implications of security-aware AV systems. Since autonomous vehicles have a long way to go before they can freely drive on the road, the work of this paper is only the starting point of the design of safety-aware audio-visual systems. This is a call to AI researchers and engineers to continue to improve its excellent work as the field advances.

https://arxiv.org/abs/1905.08453

Using domain adaptation to compensate for differences between synthetic and actual point cloud data

The researchers propose a deep learning-based adaptive approach to the field of generative adversarial networks (GANs) that acquire actual point cloud data from 3D optical radar sensors to assist in vehicle detection. The framework is based on the Cyclic Consistent Generative Adversarial Network (CycleGAN) architecture.

The framework achieves competitive results on vehicle detection tasks in real aerial view (BEV) point cloud images. When tested on real aerial point cloud images, the framework improved by more than seven percentage points over other baseline methods.

The trained model achieved an accuracy rate of 64.29%, thus demonstrating the effectiveness of this domain adaptive method on the task of vehicle detection in real aerial point cloud images. The implementation of this framework can develop state-of-the-art technology for target detection.

https://arxiv.org/abs/1905.08955

Create personalized head dialogue models from actual photos

The researchers propose a framework for adversarial generative model meta-learning capable of training highly realistic virtual conversational avatars in the form of deep generator networks. This framework creates a new model with just a few photos.

The model trained on 32 images achieves perfect realism and personalized scoring. The new algorithm is able to modify the parameters of the generator and discriminator on an individual basis, thus completing training quickly using only a small number of images.

The framework proposed in this paper can learn highly realistic and personalized conversational avatars. By performing long-term meta-learning on large-scale video datasets, the method is able to learn small samples of stranger face conversational avatar models as adversarial training problems for high-volume generators and discriminators.

The method can be used to enhance and implement human-computer interaction, natural language processing and automatic dialogue systems, and can synthesize reasonable dialogues in video sequences.

It also has practical applications in telepresence, including video conferencing and multiplayer games, as well as in the special effects industry.

Original text:

https://arxiv.org/pdf/1905.08233.pdf

Other explosive papers

New object detection and tracking system via vision, lidar and GPS / IMU positioning.

https://arxiv.org/abs/1905.08758

Simple and accurate semantic segmentation of 3D lidar point clouds using Range-Image U-Net.

https://arxiv.org/abs/1905.08748

Help students gradually implement robotics projects through research-based learning and competitions.

https://arxiv.org/ftp/arxiv/papers/1905/1905.07644.pdf

An imaging model for digitally layering overlapping brain tissue data.

https://arxiv.org/abs/1905.09231

Software that automatically generates images of highlighted brain structures.

https://arxiv.org/abs/1905.08627

AI news

New programmable chips faster than GPUs can solve the toughest optimization problems.

https://spectrum.ieee.org/tech-talk/semiconductors/processors/georgia-tech-optimization-chip-solves-huge-class-of-hard-problems

Demand for data scientists is rising.

https://www.forbes.com/sites/intelai/2019/05/22/ai-strategy-6-trends-changing-the-role-of-data-scientists/#2b5594565d5e

Companies want to leverage AI but lack the skills to assist in executing AI strategies.

https://www.artificialintelligence-news.com/2019/05/21/93-of-firms-committed-to-ai-but-skills-shortage-posing-problems/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.