Deci deep-learning platform aims to ease AI application development

Deci, a deep-learning software maker that uses AI templates designed to create AI-based applications, today launched v2.0 of its development platform, which it claims speeds the way for developers to build, optimize and deploy computer vision models. 

The term “speed” and AI application development are rarely used in the same sentence, but by using this platform, resulting AI models can be more swiftly prepared to run on any hardware and environment, including cloud, edge and mobile – with accuracy and high runtime performance, Deci CEO and co-founder Yonatan Geifman said in a media advisory. This is because much of the grunge work has been eliminated by the Deci series of DeciNet templates made available in the v2.0 platform.

Using Deci, the company says, AI developers can achieve improved inference performance and efficiency to enable effective deployments on resource-constrained edge devices, maximize hardware use and reduce training and inference cost, Geifman said. The entire development cycle is shortened – saving upfront costs – and the uncertainty of how the model will deploy on the inference hardware is eliminated, he said.

Deci’s platform, powered by its proprietary neural architecture search (NAS) engine called AutoNAC (Automated Neural Architecture Construction), is designed to enable AI developers to automatically build efficient computer vision models that provide previously tested accuracy for required inference hardware, speed, size and targets. DeciNet models generated by Deci outperform other known state-of-the-art architectures by a factor of three times to 10 times, Geifman said.

Addressing AI dev struggles 

AI developers generally have struggled to develop production-ready deep-learning models for deployment in a reasonable amount of time. These challenges can largely be attributed to the AI efficiency gap facing the industry, in which algorithms are growing more powerful and complex, but available compute power isn’t keeping pace with demand. This gap also creates financial barriers by making the deep-learning development and processing more cumbersome and expensive, Geifman said.

While NAS has been presented as a potential solution to automate the design of superior artificial neural networks that can outperform manually designed architectures, the resource requirements to operate such technology are excessive for most companies. So far, NAS has only been successfully implemented by tech giants with large AI teams such as Google, Facebook and Microsoft and in the academic community, indicating its impracticality for the vast majority of developers.

Developers can start their projects with the DeciNet pretrained and optimized models generated by the AutoNAC engine for a wide range of hardware and computer vision tasks or use the AutoNAC engine to generate more custom architectures that are tailored for their specific use-cases, Geifman said. 

In addition, the platform supports teams with the wide range of tools required to develop deep learning-based applications. These include a hardware-aware PyTorch model to easily select and benchmark models and hardware, SuperGradients — an open-source training library housed on GitHub with proven recipes for faster training, automated runtime optimizations, model packaging and more, Geifman said.

With Deci’s v2.0 platform, AI developers can accomplish the following:

• Benchmark models and inference hardware: With Deci’s hardware-aware model zoo, developers can measure the inference time of pretrained and optimized models on and various hardware including edge devices via Deci’s SaaS platform.
• Generate tailored SOTA CNN architectures: Automatically find accurate and efficient architectures tailored for the application, hardware and performance targets with Deci’s AutoNAC engine.
• Simplify training with SuperGradients: Use proven hyperparameter recipes and with Deci’s PyTorch-based open-source training library called SuperGradients.
• Automated runtime optimization: Automatically compile and quantize models and evaluate different production settings.
• Deploy with a few lines of code: Developers can deploy deep learning workloads in any environment with Deci’s Python-based inference engine.

Deci’s platform includes these three tiers:

• Free Community Tier: For data scientists and ML engineers looking to find the best models, simplify hardware evaluation and boost runtime performance.
• Professional Tier: For deep-learning teams looking to quickly achieve production-grade inference performance and shorten development time.
• Enterprise Tier: For deep-learning experts looking to meet specific performance goals for highly customized use cases.

Deci competes in the market against Datagen, Reverie, Simerse, Zumo Labs, CVEDIA, Masterful AI, Mostly AI, OneView, Synthesis AI and Sky Engine.