Skip navigation
Machine_Learning-966248982_copy.jpg

AI2Go Customizes Machine-Learning Models

Xnor’s AI2Go delivers customized machine-learning models that are optimized for performance and accuracy.

Machine learning (ML) has taken the industry by storm as hardware acceleration and applicable use cases have grown. ML can benefit from hardware, but that’s not necessarily a requirement. Its popularity among developers has opened up opportunities using platforms like a Raspberry Pi 3.

Deep neural networks (DNNs) are at the heart of many ML models. Such models can be used for a range of applications from classification to detection, and from adaptive driver-assistance systems (ADAS) to projecting failures of motors in the field.

The challenge is that all ML models are not the same. Many popular research models have made applications such as image recognition of dogs and cats easy, and a number of these models can be trained using application-specific data to create systems that are useful and accurate.

Unfortunately, though, that’s not always the case, which is why scientist and developers with this type of background are in demand to create or optimize models for specific applications. Likewise, models can be tuned for specific hardware platforms; there are optimizations for training as well as deployment. These optimizations can make the difference between requiring heftier hardware and using existing systems that may have fewer resources, resulting in more economical solutions.

A Model System

Xnor’s AI2Go (see figure) is a web-based, menu-driven system designed to deliver customized and optimized ML models for developers. It allows those with minimal understanding of ML to generate and train models while providing models that can work on edge devices like the Raspberry Pi or Toradex modules.

AI2Go_Fig_1.png

Xnor’s AI2Go starts with menu selections for hardware such as a Raspberry Pi 3, a use case (shown), and then optimizations.

The process starts with the selection of the target hardware. It can be a specific hardware platform or a more generic target like Linux x86_64 that runs on AMD and Intel processors. The next step is the use case, such as automotive or smart home with additional refinement options like car and brand classification for automotive cases.

Pretuned models are often available since Xnor has limited the options from hardware to use cases. It’s possible to do tuning with new models as well. These models are then trained so that they can be used in the field.

The service is free to try. As with many commercial offerings, there’s a price to pay to use the models in a product. Xnor’s sales group can provide the details.

The key is providing models that will work on the target hardware in a form that’s easy to utilize. Some of the models that can be generated might be comparable to open-source models. However, using and possibly tuning them can be a lengthy and costly process, especially for the uninitiated. That’s why the experts in the field are getting the big bucks. Solutions like AI2Go level the playing field at a much lower cost overall.

AI2Go doesn’t cover all of the ML bases, but there’s more than enough to hit a few home runs for the use cases and hardware platforms on its lists. Finding out whether these meet your needs is fast and relatively easy, providing developers with a way to determine whether ML solutions are applicable and affordable.

Data-Sheets

SourceESB banner with caps

TAGS: IoT
Hide comments

Comments

  • Allowed HTML tags: <em> <strong> <blockquote> <br> <p>

Plain text

  • No HTML tags allowed.
  • Web page addresses and e-mail addresses turn into links automatically.
  • Lines and paragraphs break automatically.
Publish