Neuton.AI joins Arm’s AI ecosystem, empowering users to create incredibly compact and explainable models automatically, without any coding or profound ML expertise. This opens up immense opportunities for tiny devices to perform more complex and powerful computing on the edge without running inferences via the cloud.

Arm invited Neuton.AI to join its partner ecosystem after Neuton.AI presented its novel approach to building incredibly compact and accurate models at the TinyML EMEA Technical Forum 2021.

«We designed Neuton TinyML to bring AI power to everyone, and we deeply appreciate the opportunity to become partners with Arm and enable the global TinyML community to take advantage of AI potential at scale, regardless of their data science expertise,» said Kate Vasilenko, Neuton’s Product Director.

Unlike other algorithms, Neuton TinyML doesn’t use quantization, pruning, and distillation, thus doesn’t compress the model once it’s created. Due to its own global optimization algorithm, Neuton builds models up to 1000 times smaller in size and in the number of coefficients and neurons than competitor frameworks like TensorFlow.

As such, users of any tech level have the opportunity to embed predictive models even into the smallest microcontrollers with the following characteristics: energy - 10s-100s mAh, processor < 100 MHz, memory < 100 Kb.

«The Arm AI Partner Program was designed to enable companies to work together to create AI solutions for the next decade of compute,» said Philip Lewer, director of ecosystem and developer relations, Machine Learning Group at Arm. «Through joining this growing and collaborative program, Neuton will enable even more developers to unlock new opportunities through ML at the edge.»

Neuton Neural Framework is a patented technology that is not based on any existing neural network algorithms. The uniqueness of Neuton’s approach is based on the automatic building of predictive models, neuron by neuron, ensuring the creation of neural network structures which are always optimized in size and accuracy. It is also fair to mention that Neuton is equally effective with both small and large datasets.

As part of this partnership, Neuton held a webinar on the topic “Bottom-up Approach to Building Extremely Tiny Models” as part of Arm’s AI Virtual Tech Talks. Most neural networks created today contain too many coefficients and neurons, and are based on a predetermined architecture which often results in the unnecessary growth of network size.

During this tech talk, the Neuton team showcased how to overcome this issue leveraging a completely different approach that would help to create models that are 1000x smaller and faster than models created using other frameworks such as TensorFlow lite and PyTorch. The webinar recording is available here.