Custom Neuton models
A smarter approach to building edge AI models
Challenge
Traditional neural networks face a fundamental challenge: they require manual architecture design and rely on methods that often produce bloated models with millions of parameters. Data scientists must painstakingly tune dozens of variables, from learning rates to layer depths, in a trial-and-error process that's both resource-intensive and imprecise.
The Neuton models solution
Custom Neuton models takes a radically different approach. Instead of starting with a predetermined structure, it grows neural networks neuron-by-neuron, automatically determining the optimal architecture as it learns. This granular construction process, combined with a patented global optimization algorithm that avoids the pitfalls of traditional gradient descent methods, produces remarkably compact models without sacrificing accuracy.
Result
A fully automated system that requires minimal expertise to use. Simply provide your data and target variable, Neuton handles structure creation, cross-validation, and overfitting control automatically. The framework delivers models that are dramatically smaller than conventional neural networks, enabling faster predictions and deployment on resource-constrained devices, all while maintaining excellent accuracy and generalization capabilities.Nordic Edge AI Lab
Nordic Edge AI Lab is your gateway to ultra-efficient edge intelligence. Powered by our patented neural network framework, it transforms your sensor data into compact, high-performance AI models ready to deploy on Nordic’s ultra-low-power SoCs and SiPs in just minutes. There is no data science knowledge or experience required, only a good dataset and a can-do attitude, the Edge AI Lab handles the rest.
Documentation for the Edge AI Lab is available here.
Go to the Nordic Edge AI Lab
Benchmarks
Bearing fault detection case-study
Models built on dataset from Case School of Engineering
| Total Footprint | LiteRT | Neuton | Neuton Advantages | |
|---|---|---|---|---|
| NVM |
TinyML framework (model + inference engine + DSP) |
61.7 | 6.6 |
9 times smaller model |
| RAM | TinyML framework (model + inference engine + DSP) | 11.2 | 1.2 |
9 times smaller model |
| Inference time (ms) | 360 | 1,46 |
246 times faster |
|
| Holdout validation accuracy | 0.79 | 0.98 | 19% higher accuracy | |
Test performed with both Neuton and LiteRT models running on an nRF52840 and tested on the same validation dataset.
"Magic Wand" gesture recognition
| Total Footprint | LiteRT | Neuton | Neuton Advantages | |
|---|---|---|---|---|
| NVM |
TinyML framework (model + inference engine + DSP) |
79.96 | 5.42 |
14 times smaller model43% reduction of total NVM use |
| Device drivers and business logic | 93.47 | 93.47 | ||
| RAM | TinyML framework (model + inference engine + DSP) | 18.2 | 1.72 |
10 times smaller model26% reduction of total RAM use |
| Device drivers and business logic | 45.69 | 45.69 | ||
| Inference time (µs) | 55,262 | 1,640 |
33 times faster |
|
| Holdout validation accuracy | 0.93 | 0.94 | 0.7% higher accuracy | |
Test performed with both Neuton and LiteRT models running on an nRF52840 and tested on the same holdout dataset.
-
Predictive maintenance and building automation systems
In building automation systems, embedded machine learning can be leveraged to monitor equipment health in real-time and anticipate failures before they occur. By analyzing sensor data directly on edge devices, these systems can detect patterns and anomalies in HVAC, lighting, and security operations without relying on constant cloud connectivity. This approach reduces downtime, optimizes energy use, and lowers maintenance costs, all while enabling smarter, more responsive building environments. -
Smart sensor networks with local data analysis on each node
Smart sensor networks can utilize embedded machine learning to process information directly at the source, reducing latency and bandwidth requirements. Each node can independently detect patterns, filter noise, and make real-time decisions, enabling more efficient and scalable systems. This decentralized approach enhances responsiveness and reliability, especially in applications like environmental monitoring, industrial automation, and smart buildings, where immediate insights and minimal data transmission are critical. -
Movement and gesture recognition for remote controls and wearable devices
For remote controls and wearable devices, embedded machine learning enables real-time interpretation of motion data directly on the device. By running lightweight ML models on embedded processors, these devices can accurately detect and classify user gestures without needing constant cloud access. This allows for intuitive, low-latency interactions in applications like touchless control, fitness tracking, and assistive technologies, while maintaining energy efficiency and data privacy. -
Health and activity monitoring for smart health wearables
Smart health wearables are increasingly powered by embedded machine learning, enabling real-time analysis of biometric and motion data directly on the device. These wearables can track vital signs, detect irregularities, and classify physical activities with minimal latency. By processing data locally, they improve user privacy, extend battery life, and provide immediate feedback, making them ideal for continuous, personalized health monitoring and early detection of potential medical issues.
Use cases
Demo videos