The term “Internet of Things” can be used to describe a huge range of different technologies, from sensors to gateways to back-end systems that organize data and keep machine-to-machine networks secure. Lots of attention is rightly paid to the way IoT systems gather data and how it moves from place to place. However, for some parts of the IoT, the issue of how to keep sensors powered may be just as important.

Particularly in the case of IoT systems that feature small sensors and sensors that might be far away from each other or from the rest of the system, energy usage is a critical concern, because traditional wired power may simply not be an option.

Agriculture, utilities, and transportation are among verticals where widely spaced, low-power deployment is important. Scientists studying a volcano might not be able to run a power cable all the way from the closest part of the grid to their vibration sensors. Soil moisture testers in a farmer’s field could face the same problem, and so on.

There are, however, other options, and choosing the best solution has everything to do with understanding what the desired business outcome and how to attain it with peak efficiency, according to Gartner vice president and analyst Al Velosa.

 “The fundamental question is ‘hat’s it cost to deploy the infrastructure?’” he said. “If you’re managing a couple thousand miles’ worth of assets…the bigger cost is sending a truck to that asset than anything else.”

That’s particularly important for the first and probably the most common option for powering small, remote IoT assets – battery cells. No battery lasts forever, so, eventually, they have to be swapped out. The current state of the art focuses on silver oxide cells akin to watch batteries and those used in hearing aids, according to Forrester vice president and principal analyst Frank Gillett,

“One of the problems you run into is that some battery chemistry won’t make it 10 years,” he said.

Silver oxide cells remain popular because their charge-to-weight ratio is comparatively very high. Even a small battery of this type can power a simple sensor outfitted with a low-power, infrequently used radio for years, potentially. They are not, however, so powerful as to free device makers from the responsibility of designing for maximum efficiency.

At its most basic level, an IoT sensor needs to be able to collect information, express that information in a digital format, and transmit it up the chain, whether that’s to a nearby edge device for collation and processing or directly to the back end. Each part of that process has an energy cost, and while advancing technology has dramatically increased power efficiency in both processing and transmission, energy is still one of the primary limiting factors in IoT device design.

“It’s figuring out how to maximize what a low-power device can do, and a lot of that is making the radios more efficient,” said Gillett. “The flipside is making the compute part of the IoT endpoint very power-efficient as well. Ideally you have them both fairly integrated.”

Battery technology, he added, advances comparatively slowly, compared to processors, chips and sensors. That’s part of why some companies are looking elsewhere to power their IoT devices.

One option is solar power. Increasingly efficient solar cells mean that it’s easy enough to add appropriately sized panels to small devices, and the cost of those panels has dropped of late, also.

That’s great in theory, according to Velosa, but in practice, many deployments using solar energy aren’t going to be that much more efficient than those using battery power. Solar is still dependent on the panels getting enough exposure to the sun, and what’s more, they’re far from maintenance-free. Dust and dirt can degrade their ability to generate power.

“We are still looking at normal deployment of an asset being 1-5 years, with some service call in the middle of that,” he said.

The post IoT power: battery, wired or wireless? appeared first on Artificial Intelligence.

To help engineers develop AI systems, MathWorks has added deep learning capabilities to its latest update of MATLAB and Simulink. The update, called R2020A, includes a “Deep Network Designer” app, which is said to help engineers train neural networks. Designers can also manage several deep learning experiments at a time in another app, Experiment Manager.

Users will have more network options when creating deep learning code. 

What else is included in this update and how will it benefit engineers in various fields?

Core Updates for AI Development

The latest update from MathWorks includes a range of updates for AI systems, including code conversions, deployable web apps, and managers. Many of these are included in an updated Deep Learning Toolbox.

As mentioned, the two major updates of R2020A is the Deep Network Designer app, which allows users to train neural networks, and the Experiment Manager, which allows users to organize multiple experiments. The Experiment Manager also allows users to track experimental data, like core variables and results, providing a comprehensive view of a deep learning project.

Deep learning tools have also dramatically improved over a range of applications to help engineers with common AI tasks. AI involved with video and image data now includes features such as a signal labeler, pixel label datastore, audio datastore, and image datastore. In the network realm, MathWorks now allows users to build advanced network architectures, such as GANs, Siamese networks, attention networks, and variational auto-encoders.

Deep Learning Frameworks

MathWorks Deep Learning now supports a new range of other deep learning frameworks such as MobileNet-v2, ResNet-101, Inception-v3, SqueezeNet, NASNet-Large, and Xception. When it comes to the deployment of AI systems MathWorks now supports Arm Mali GPUs and includes a GPU Coder. The update is said to enable automatic deployment to Jetson AGX Xavier and Jetson Nano Platforms.

MathWorks can also generate code for networks such as YOLO V2 object detector, DeepLab-v3+, MobileNet-v2, Xception, DenseNet-201, and recurrent networks. AI algorithms for reinforcement learning are now included for systems for walking and driving. Other generic robotics use algorithms including DQN, DDPG, A2C, and PPO.

Classes and Interface Upgrades

The R2020A update does not just affect deep learning applications; it also includes features that can help designers interface their designs with other systems and improve organization. For example, MathWorks now offers C++ classes from MATLAB classes. It also offers new messaging protocols.

The R2020A update allows MathWorks users to work with UTF-8 characters and includes new graphic options for graphs. Users can also access a live editor tool, which allows them to view tasks currently being executed. 

Three new products included in the R2020A update include the Motor Control Blockset, Simulink Compiler, and MatLab Web App Server.

To help designers generate compact code, the Motor Control Blockset offers a library of motor control algorithms. It also provides “out-of-the-box support” for a number of motor control hardware kits.

With the Simulink Compiler, designers can use Simulink models to create software, web apps, and standalone applications that will run simulations without actually installing Simulink. 

Finally, the last major product—MATLAB Web App server—gives engineers organizational control over MATLAB web apps across one’s organization from a web browser.

A Deeper Understanding of Designs

MathWorks says the tools included with R2020A will not only help engineers continue to develop AI systems but also help them to better understand their designs. For example, because users can now label signals in video applications, they can visualize stimuli; likewise, datastores allows them to manage audio and visual data.

By targeting more AI frameworks, MathWorks extends its reach and therefore widens the deployment possibilities of a trained system. The new debugging features will allow designers to better understand why their AI systems produce the results they do (i.e. explain their decisions), while the inclusion of reinforcement algorithms allows for engineers to explore new methods for training systems.

A Useful Asset for Automotive and Wireless Designers

MathWorks asserts that the automotive tools included with R2020A will dramatically improve the automotive sector, with tools like the HD map to road data converter. These features are said to decrease the time needed to set up a traffic simulation environment and define road lengths and intersections. 

The update allows automotive designers to read high-definition road maps to create road data for use in traffic simulations. Optimized shift schedules can then be used with fuel economy and emissions analysis.

The update may also give designers in wireless communication a leg up with better support for wireless and 5G signal analysis, including waveform generation and cell detection.

A “Comprehensive” AI Platform

David Rich, MATLAB’s marketing director, says that this R2020A update gives designers a comprehensive platform for building an AI-focused system. He explains, “We’ve taken three decades of product, consulting and support experiences and applied it to an AI workflow that empowers engineers and scientists to clean data, build models and deploy them in production IT or embedded systems.”

Have you seen deep learning become a more central part of your workplace? If so, how? Share your experiences in the comment below.

The post MATLAB and Simulink Get a Major Deep Learning Facelift appeared first on Artificial Intelligence.