Melika Payvand, SNSF Assistant Professor at the University of Zurich (UZH), was awarded an SNSF Starting Grant. In February, she started working on her project “UNITE: Brain-inspired device-circuits-algorithm co-design for resource-constrained hardware on the edge” at the Institute of Neuroinformatics, UZH/ETHZ. The project aims to develop devices – no bigger than a coin – that autonomously learn from their local environment and run for at least five years. Such low-power consumption devices are finding applications for wearables and internet of things systems.
Melika, congratulations on your SNSF Starting Grant! What are the main goals of the “UNITE” project?
The main goal of our project is to develop a new class of devices that can adapt to the input they receive with minimal memory and power needs. This device can have a variety of applications, for example, in personalizing tiny health monitoring devices without compromising a user’s privacy. This is a challenging problem because current adaptation algorithms require a lot of memory and power, which will not fit on battery-operated wearables. We plan to solve this by “uniting” the exciting advances in memory, silicon technologies, and novel brain-inspired algorithms.
What was the reason for starting this research project?
As I said, designing tiny, low-power, autonomously adapting systems is challenging. Thankfully, there is proof of existence of an efficient learning system: our brains – where neurons and synapses use their space and energy wisely. To save energy, neurons only generate an output after integrating enough spatial-temporal information about an input. To save wiring and space, they mostly make synapses to neurons in their neighborhood. Similarly, synapses only change their strength while learning based on recent and local information. We will implement these principles on integrated circuits utilizing novel nano-scale memory technologies to build efficient learning systems.
Would you tell us more about what exactly you will do in “UNITE”?
Using nano-scale memory integrated into silicon technologies, we will start by implementing two brain-inspired principles for reducing memory and power footprints. The first one uses more accurate neuron models that use temporal and spatial parameters. This process will enable more computational capacity for the neural network of these neurons. In parallel, we work on applying brain-like local connectivity, enabling more accessible information transport. We will apply these principles to biomedical signal and speech processing datasets and assess their performance. These results are combined into a system that is augmented with local algorithms enabling the adaptation of the network parameters to new inputs. This will result in a final learning system with low power and memory footprint.
What do you hope will be the project’s impact on research and society?
Our digital society is evolving into an era of ubiquitous specialized and distributed smart devices. These devices must continuously process various signals, such as heartbeat, speech, or even the temperature of one’s body or engine. This raises some challenges, such as the exponential increase in data and thus power consumption, compromising the users’ privacy, and continuous adaptation of the devices in dynamic environments. With UNITE we will take an essential step towards solving these problems on the road to more sustainable AI (artificial intelligence) hardware.
Image: Gerd Altmann, Pixabay