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

Big data Digest column works

Author: Christopher Dossman

Compilation: Jiaxu, Yunzhou

Hello, everyone, this week's AI Scholar Weekly program is with 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 AI academic overview, capturing the weekly AI academic cutting-edge information, and updating AI black mirror short stories irregularly at the end of the article.

Update on Monday, do AI research, start with this article every week!

Keywords of the week: end-to-end autopilot, fantasy text adventure games, open source data sets for instructional video analysis

AI fantasies

The field of AI is developing in two different directions, one is small micro computing, the other is giant computing.

Hot academic research this week

A simple confrontation example for end-to-end autonomous driving model

In the process of developing a multi-functional modeling framework and simulation infrastructure to study adversarial examples of end-to-end self-driving models, researchers found that some very simple and easy-to-design physical conditions, will produce a certain degree of antagonism to the self-driving model. For example, the marking lines on the road can destabilize the end-to-end driving model.

The antagonism of these examples does not show strong interference in many cases, such as driving in a straight line, but in other cases, such as turning a vehicle, it will show a strong interference.

Although it is very simple for human beings to distinguish and avoid such conflicts, for the end-to-end surveillance model, these disturbances will cause serious traffic violations. In tests and experiments, the researchers used the CARLA self-driving car simulator to prove that these physical disturbances not only exist, but also have a strong effect on even the most advanced models under specific driving conditions.

Potential application and impact

The problem that intelligent machine learning systems are extremely vulnerable to interference has caused widespread concern. These research results and their theoretical framework undoubtedly provide useful information for future research, and reveal the shortcomings of the end-to-end deep learning model, which can be improved in the future. In addition, this study also provides important insights for artificial intelligence engineers to further explore the possible attacks on a wider range of deep learning models.

Original text:

Https://arxiv.org/abs/1903.05157

Comprehensive teaching video analysis data set (COIN data set)

In order to solve the problem of lack of diversity and scale of teaching video data sets, researchers recently launched the COIN data set, which is the largest video data set used for comprehensive teaching video analysis.

The dataset is organized in a hierarchical structure, including about 11827 videos, covering more than 10 areas of daily life. All videos in the COIN dataset are professionally annotated. In addition, the researchers also proposed a simple and effective method to capture data dependencies at different description stages. This kind of method can also be combined with traditional methods to better extract the key steps in teaching video.

Potential application and impact

Researchers are very clear about the purpose of creating COIN data sets-to establish a comprehensive and comprehensive teaching video data set through rich semantic classification to provide a benchmark for teaching video analysis. Through the establishment of COIN, researchers also hope to promote the in-depth study of teaching video analysis in the AI community in the future.

COIN dataset:

Https://coin-dataset.github.io/

Original text:

Https://arxiv.org/abs/1903.02874v1

Quantification without sacrificing accuracy

In general, quantization will reduce the accuracy of the model. Nevertheless, researchers are still committed to reducing the decline in accuracy in the quantization process by improving quantization techniques. In a recent paper, they proposed a focused quantization technique that can effectively quantify the weight of the model.

The quantization method proposed in this paper uses Gaussian mixture representation to locate the high probability regions in the weight distribution of the model and quantifies them at a fine level. In addition, this quantization technique only uses the power of-2 to provide an effective computing model.

Through a series of integrated pruning and coding, they have managed to demonstrate high-end compression ratios on a variety of CNN. For example, they achieve a compression ratio of 18.08 × in ResNet-50, and the loss of model accuracy is only 0.24% negligible, which exceeds the current compression technology.

Potential application and impact

The quantized parameters integrated with the quantitative calculation of the deep neural network have a very high ability to achieve a large number of computational gains and optimize performance. The focus quantization proposed in this paper can reduce the model size and computational cost, transform it into high compression ratio, and improve the efficiency of current and future CNN.

Original text:

Https://arxiv.org/abs/1903.03046v1

LIGHT: learn to speak and act in fantasy text adventure games

Imagine if there were a game in which multiple computers and humans could play game roles. What would such a game look like? Researchers recently launched a fantasy text adventure game in which players can interact with knowledge, action and expression while establishing conversations with multiple players.

Existing word adventure games are usually single-player, and players can't talk to humans in the game, which is why I think the game sounds interesting and adventurous.

Learning in an interactive game between people and text (LIGHT) is a multi-player fantasy text adventure game designed to help institutions study many aspects of dynamic collaborative dialogue between humans.

The text-based game is set in a rich game world, including more than 660 locations, 3460 objects and 1750 characters defined only in natural language. Researchers have collected a large number of data sets (11k sets) that deal with role-driven human-to-human interactions such as behavior, expression and dialogue, with the aim of training models to attract humans in a similar way.

The test results show that the training, generation and retrieval models under these conditions can effectively use the potential conditions or rules of the game world to adjust their predictions.

Potential application and impact

LIGHT provides a unified platform for sharing, training, and evaluating collaborative dialogue models across many different tasks, enabling people to interact with machines in interesting ways. By providing a suitable research platform, the framework allows research institutions to carry out a variety of collaboration and learning, so as to study and enhance the existing collaborative dialogue system. The researchers hope that this work will further promote and promote the research on basic language learning.

Official website:

Http://parl.ai/

Original text:

Https://arxiv.org/abs/1903.03094v1

Realization of self-adaptation of automatic Robot by Machine Learning

Researchers have proposed a method of integrated learning and quantitative planning, whose main goal is to achieve self-adaptation in highly configurable systems running in dynamic and uncertain environments such as robot systems. This technology uses configuration changes as the primary mechanism for implementing adaptation.

The innovation of this method is that it applies machine learning to discover Pareto optimal configuration without exploring all configurations, and applies the limitations of search space to the specific settings of controllable plans. In this way, it can combine learning with quantitative planning to achieve runtime adaptation.

In addition, this method is helpful to integrate information from multiple models in quantitative planning. Specifically, the researchers explored robot operations that need to consider time and energy consumption. Independent evaluation shows that this method produces a high-quality adaptation process in uncertain and dynamic environments.

Potential application and effect

As a new technology which can effectively realize robot adaptive integrated learning and quantitative planning, this method can be used in many other network physics systems. In addition, it can be extended as an online interface to the interface of run-time model updates.

Original text:

Https://arxiv.org/abs/1903.03920

Other popular style papers

Recent research shows that through demonstrations, you can now teach robots to perform two different cleaning tasks.

Original text:

Https://arxiv.org/abs/1903.05635

When allowing reliable camera attitude estimation, how to avoid divulging confidential information on the 3D scene? A recent study proposed an image-based privacy location solution.

Original text:

Https://arxiv.org/abs/1903.05572

Google has launched a new open source library that can effectively train giant neural networks.

Web page:

Https://ai.googleblog.com/2019/03/introducing-gpipe-open-source-library.html

The following is 6D object attitude estimation based on instance and category levels for augmented reality, robot control and navigation applications.

Original text:

Https://arxiv.org/abs/1903.04229

Want to know how to summarize and find meaningful concepts in biomedical texts?

Original text:

Https://arxiv.org/abs/1903.02861v1

AI News

Google has released an end-to-end speech recognizer based on neural networks to support state-of-the-art speech recognition.

For more information:

Https://ai.googleblog.com/2019/03/an-all-neural-on-device-speech.html

Good news for researchers: Deep Mind has developed an open source software library to help them deploy the TensorFlow model.

For more information:

Https://deepmind.com/blog/tf-replicator-distributed-machine-learning/

Will China eventually overtake the United States in AI research? Now China has published more papers on AI than the United States.

Read more:

Https://www.wired.com/story/china-catching-up-us-in-ai-research/amp

Introduction to the columnist

Christopher Dossman is the chief data scientist at Wonder Technologies and has lived in Beijing for five years. He is an expert in the deployment of deep learning systems and has extensive experience in developing new AI products. In addition to his excellent engineering experience, he also taught 1000 students the basics of deep learning.

LinkedIn:

Https://www.linkedin.com/in/christopherdossman/

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.