Demonstrators programme

As data volumes or message numbers surge and decision windows shrink, processing high-velocity data streams in near real time has become increasingly challenging in modern computing. This demonstrator shows how neuromorphic computing solutions can be used to deliver fast, energy efficient technologies.

Use case #1

ASTRON’s largest low-frequency radio telescope (LOFAR) enables world-leading radio astronomical research. This multipurpose sensor network is capable of handling extremely large data volumes for different research purposes, such as constructing detailed images yet seen of galaxies beyond our own. With the upcoming update to LOFAR that will increase data rates and bandwidth, a new way is needed to process these massive amounts of data.

Our researchers at CogniGron will develop neuromorphic computing chips capable of processing data at the receivers significantly more energy-efficient before the data is sent for central processing.

And what happens at this central location? Data from all 52 LOFAR antennas then needs to be correlated, fed into a calibration pipeline, and translated into images. Our goal is to facilitate, with neuromorphic computing, data to be processed faster while being online without the need for expensive data storage.

Processing of radio signals happens at many places in our modern world. One example is processing of radar signals. By working on radio-astronomy challenges, CogniGron aims to open the door to a wider range of real-life problems that call for faster and more energy-efficient compute capabilities.

This demonstrator explores how neuromorphic technologies can provide dedicated hardware solutions for continuous sensory processing in resource-constrained environments. Event-based vision sensors are already a success in neuromorphic engineering: this technology offers many advantages to traditional systems such as ultra-low latency for fast measurements and feedback, high data efficiency and minimal bandwidth usage, and energy-efficient, real-time processing at the edge. Our demonstrator programme is working towards two specific use cases for neuromorphic event-based sensing:

Use case #1

Robots and remotely operated vehicles (ROVs) must know their position and orientation in 3D space quickly and accurately in order to move safely and effectively. This use case focuses on estimating a system’s own motion (ego-motion) almost instantly and with high precision.

To do this, we will develop a biologically inspired algorithm that processes data from event-driven sensors, which report changes in the environment rather than continuous measurements. Many existing systems rely mainly on inertial measurement units (IMUs), but these sensors accumulate errors over time. By combining IMU data with event-based sensing in a multi-sensor (multi-modal) setup, the system can reduce drift, improve localization accuracy, and remain robust even during fast or complex motion.

Use case #2

Extensive monitoring of structures and materials can prevent disasters and help companies to plan efficiently. In order to deploy a sufficient number of sensors/robots, these need to be energy efficient, process the data real-time and securely. Current solutions require sending data to the cloud, which prevents processing data on-the-fly and compromises data privacy. A tactile-based feature extraction pipeline, based on neuromorphic spike detection and recognition, will be developed for the inspection of surfaces. 

Solutions include a tactile system (sensor devices, circuit design and recognition algorithm), that can be used to forecast initial stages of decay or degradation in structures, facilitating objective and measurable status data, and leading to data-driven predictions of lifetimes. Recognition of fabrics in the textile industry or real-time determination of the stiffness of cement in the construction industry, are further specific examples that could be tackled with this approach.

Neuromorphic computing technologies offer promising solutions to traditional computing challenges by enabling highly energy efficient, massively parallel, and near real-time processing, taking inspiration from the brain. This demonstrator focuses on three use cases to show the advantages of neuromorphic computing in medicine and healthcare.

Use case #1 - Ultra-low-power smart wearables for health monitoring

Wearable devices for health monitoring are becoming increasingly more common. Think of devices that can help detect, for example, stress signals or early seizure signs. Currently, these devices still mostly rely on batteries and cannot process much data themselves, so the data is sent to external servers. 

Neuromorphic computing enables ultra-low-power, event-driven, on-device processing of biosignals (for example, by using ambient energy) making it possible to perform personalized signal classification in the wearable device itself, minimizing heat, battery usage, and data transmission.

Use case #2 - Hardware security supported federated learning in health care

In healthcare, maintaining privacy of patient data while providing personalized models is crucial. In cases where health data is limited or fragmented, training machine learning models locally does not always work. To address this, this use case focuses on hardware-security-supported federated learning, a model training method that enables transfer learning-based aggregation across many devices or institutions without collecting data in one place.

Use case #3 - High-throughput analysis of omics data

Modern medical and life-science research increasingly relies on omics data: large, complex data sets describing genes, proteins, and metabolites. These datasets are used for understanding biological processes, discovering new biomarkers, and accelerating innovation in healthcare and biotechnology. Yet analysing multi-dimensional and multi-omics data efficiently remains a major technical challenge. 

At CogniGron, we explore how brain-inspired computing architectures can transform high-throughput analysis of omics data. Neuromorphic systems, designed for massive parallelism and ultra-low latency, offer a promising new route for detecting patterns and features in large biological data sets, far beyond what conventional computing architectures can efficiently deliver.