)
Neuromorphic Computing in Industrial Quality Control
A New Paradigm in Computation
Neuromorphic computing, a technology that seeks to replicate the brain's own efficiency and processing power, can improve demanding industrial applications, particularly in the realm of real-time visual quality control. Reply’s approach offers significant advantages in energy consumption, speed, and scalability.
Emulating the Brain's Blueprint
The term "neuromorphic" is chosen with precision, as this form of computation aims to replicate the intricate activity of biological neurons in every significant aspect. Unlike in classical deep learning, where the concept of a neuron is more abstract, neurons within a neuromorphic system possess their own distinct sense of timing and frequency.
The core operation centres on the "spike", or "firing event". Inside a neuromorphic chip, individual spiking neurons accumulate incoming signals over time, much like the build-up of a membrane potential in their biological counterparts. When this internal charge reaches a specific threshold, the neuron fires a spike, transmitting information to other nodes in the network. This event-driven process forms the foundation of Spiking Neural Networks (SNNs), which are inherently dynamic and recurrent systems adept at processing data where temporality is key.
Core hardware architectures
The hardware that enables this can be broadly categorised into two types: fully asynchronous and partially asynchronous systems. In a fully asynchronous system, every core within the chip operates with its own independent timing and frequency, a design that most faithfully mirrors the sparse and efficient nature of the brain's computations. This architecture is exemplified by Intel's Loihi 2, currently one of the world's most prominent neuromorphic chip. Conversely, partially asynchronous systems employ synchronous processing within the individual cores but use asynchronous protocols for communication between them. This hybrid model can yield significant gains in hardware efficiency and performance speed-up. This kind of hardware design is embedded into two prominent chips: SpiNNcloud's SpiNNaker 2 and IBM’s TrueNorth.
Unprecedented energy efficiency
A primary advantage of neuromorphic systems is their remarkable energy efficiency. In an environment where power consumption is a critical design constraint, this technology offers a significant breakthrough. Neuromorphic hardware can deliver energy savings of up to 100 times compared to conventional CPUs and around 30 times compared to GPUs, making it a viable solution for edge computing and large-scale AI deployments where power is at a premium.
Enhanced scalability and performance
The architecture of neuromorphic computing is inherently designed for scalability. These systems can be expanded to massive networks, with platforms like Intel's Hala Point managing over a billion neurons, a scale that poses a considerable challenge for traditional systems. This scalability is matched by a substantial leap in performance. For computer vision tasks, these systems can achieve an inference speed-up of 120 times, enabling complex, real-time decision-making.
Intelligent data acquisition via event cameras
Neuromorphic principles also extend to data acquisition through the use of event cameras. Unlike traditional cameras that capture entire frames at fixed intervals, often recording redundant data, event cameras operate asynchronously. They capture data only when a change in a pixel's luminosity is detected. This results in a sparse but highly informative data stream that drastically reduces storage and processing requirements. In one practical example, an event-based approach reduced a 32-gigabyte dataset to just 7 gigabytes.
Neuromorphic Computing in Quality Control
The demanding field of industrial visual quality control represents an ideal application for neuromorphic technology. Manufacturing environments require real-time, high-accuracy defect detection, a task whose stringent demands on speed and efficiency align perfectly with the capabilities of neuromorphic systems.
To address this need, established AI models are being adapted for the neuromorphic paradigm. An example is Spiking-YOLO, a re-engineered version of the well-known YOLO object detection framework. This model utilises spiking neurons to process visual data. The specific architecture used in a recent project is a highly optimised implementation that merges computational layers to improve efficiency for deployment on neuromorphic hardware.
Multi-object detection for autonomous driving
In a practical application focused on multi-object detection for autonomous driving, a dataset of driving footage was converted into an event-like format using a simulator. The results were striking: while the car was in motion, the event-based data captured the full context of the surrounding environment, including other vehicles and the landscape. However, the moment the car stopped, the static parts of the scene vanished from the data stream, demonstrating the system's inherent focus on relevant change.
The model was pre-trained on the Common Objects in Context (COCO) dataset, and then further trained on the BDD multi-object dataset, with initial inference tests conducted via a classical simulator. The next stage of development involves re-training this model directly on the converted event-based data accessing the capabilities of the neuromorphic hardware.
An outlook on Intelligent Automation
Neuromorphic Computing is perfect for deployment on edge devices, which are common in industrial environments but often have significant hardware and memory limitations. The combination of high performance, low energy use, and efficient memory management makes neuromorphic systems an ideal solution for tasks like real-time image recognition on a factory floor.
Reply envisions a production line where an event camera constantly monitors products, with a neuromorphic chip instantly identifying anomalies and alerting personnel, all while operating within the tight constraints of an edge device. While these algorithms are not yet prepared for highly complex cognitive challenges, Reply’s experience shows that their current capabilities already mark a significant step forward for intelligent industrial automation.
Data Reply, as part of the Reply Group, supports customers in becoming data-driven. Data Reply is active in various industries and business areas and works intensively with customers so that they can achieve meaningful results through the effective use of data. To this end, Data Reply focuses on the development of data analytics platforms, machine learning solutions and streaming applications - automated, efficient and scalable - without compromising on IT security.