Optimizing AI for embedded devices: innovations from iNEST Spoke 9
Laura Meneghetti from SISSA applies fluid dynamics techniques to make neural networks more efficient on hardware with limited resources
Household appliances, smartphones, and self-driving cars are increasingly relying on artificial intelligence for various functions, such as image recognition or natural language processing, as ChatGPT does. Specifically, these are embedded devices designed to perform specific tasks efficiently, despite having significantly lower computational capacity than a computer.
But how can a complex AI algorithm be designed to work with limited resources?
To integrate artificial intelligence into embedded devices, scientists use sophisticated tools for mathematical and numerical modeling. This is where the work of Laura Meneghetti, a SISSA researcher involved in Spoke 9 of iNEST, comes into play. Her research focuses on a specific type of AI algorithm—neural networks—to make them more easily implementable even in devices with reduced computational power. «A neural network consists of various layers, each with a specific function,» explains Meneghetti. For example, in image recognition, where neural networks must analyze and classify images based on specific characteristics, «each filter detects a particular feature within the image, such as the presence of horizontal or vertical lines,» the researcher continues. By combining the results from different filters, the network can then determine the subject captured in the image.
This layered structure is key to optimizing the neural network. «Each filter has a different level of importance compared to the others, determined by a series of parameters called weights,» Meneghetti adds. «To simplify the process, a parameter reduction method can be applied. By analyzing the weights, one can identify the less relevant filters—and consequently layers. By removing them, no essential information is lost, and the network achieves the same results while operating more quickly. The method first focuses on reducing the number of layers and then minimizing the number of remaining parameters, preserving the crucial information for the network’s functioning.»
The parameter reduction method developed by the SISSA researcher is based on techniques used in fluid dynamics, the branch of physics that studies the behavior of fluids. Meneghetti has been working on this for several years: «During my Ph.D., I successfully applied this method to image recognition,» she states. After completing her Ph.D., her work continued to evolve: the method was refined and its effectiveness was demonstrated on various datasets and networks commonly used to test these types of algorithms. Her team is currently working on a scientific paper to publish all these promising results.
Laura Meneghetti’s work has already found industrial applications, thanks to collaborations with companies such as Electrolux Professional and Indaco Project. «The next step will be to refine the method for more complex applications and see if it works correctly with more challenging tasks,» the researcher says, «for example, not just classifying but also locating objects in images.» Such an application could have a significant impact in sectors beyond embedded devices, paving the way for technologies capable of radically transforming our way of living and interacting with the world.
