The AI Revolution: Radically Different Reliability (Part Two)
by Rob Telson
In the previous blog, we highlighted how event-based Spiking Neural Networks (SNNs) are less artificial and more intelligent than conventional AI because they more closely resemble the biological brain.
The BrainChip Akida ™ processor has been used to classify images, identify odors and tastes, recognize breath data for disease classification, identify air quality, interpret LiDAR laser light data, recognize keywords, and detect cybersecurity attacks. It will be especially valuable in the evolution of smart ‘edge’ devices, anything outside of a traditional lab or data center that connects to the Internet/cloud, where low power (as in micro- to milli-watt), high performance, low cost, and small size are important.
This includes wearable medical devices, automotive collision avoidance systems, industrial automation, voice-operated remote-control devices, and many other beneficial AI-based tools.
AI Neuromorphic Computing is showing huge promise in real-world applications of the Akida technology. It’s truly beneficial.
AI leveraging Neuromorphic Architecture can be used to identify disease markers in blood or breath, and is capable of detecting abnormalities in X-rays, CT scans, and MRIs. It’s been tested for cancer diagnostics: when the patient exhales, the sensors record volatile organic compounds (VOCs) in their breath. The sensor data is then analyzed for specific VOCs that are thought to be biomarkers of 17 different types of cancer, including cancer of the lung, breast, prostate, stomach, liver, leukemia, and melanoma.
Breath sensor data is also being used for various virus detections such as COVID-19, as well as monitoring of people currently infected. This data can then be used as a possible way to control epidemic outbreaks. BrainChip’s Akida neuromorphic processor, running in the event domain, has detected the instance of disease between a disease group and a healthy control group with 98.7% accuracy.
Traditional CNN-based AI solutions are not robust enough to support the low power and data efficiency required to expedite the testing and receipt of accurate results.
Switching gears (pun intended), SNN-based AI solutions can also be used to make transportation systems safer. When applying this technology to monitor wheel bearing noise for signs of wear, as an example, safety for both passenger trains and heavy iron ore trains increases because early notification and preventative maintenance on wheel bearings reduces the chances of critical failure.
As technology drives the world of smart transportation, the Akida processor will be at the forefront of providing more efficient and effective solutions throughout our vehicles. As the demand for smart technology increases in our vehicles, we will be able to drive intelligent applications in the automotive market by reducing power demand and cloud dependency. Voice commands for in cabin operations, object detection and classification for front and rear camera systems, and gesture recognition for in cabin controls are just a few of the possibilities where neuromorphic computing will make a positive impact on our vehicle experience in the near future.
Radically Different Reliability
For practical, real-world products and scenarios like these mentioned above, AI must deliver insight, sequential knowledge, and advanced learning capabilities. This is entirely different from what traditional convolutional neural networks can do. We need chips for general intelligence that behave like biological neurons, instead of pocket calculators.
Until now, edge AI ‘processing’ was performing only minimal preprocessing. Data was preprocessed, transmitted to the cloud for the heavy lifting, and the results were sent back to the device for action. This method has obvious disadvantages: Internet service availability, latency issues (especially as the number of devices that are connected to the network increases), and major security issues (particularly for medical, personal, or image data).
Today’s Akida chips perform AI processing and produce results within the edge device itself. This allows for radically different reliability. Think about it, the demand for computing on the cloud is decreased. Security on the edge device is enhanced. The potential for delayed computing challenges is reduced. This is huge!
The Akida Platform is Different
BrainChip’s Akida event domain neuromorphic processor is extremely power-efficient, scalable, high in memory capacity, easy to configure and reconfigure as application needs change. It can also execute prior-generation CNNs, as well as future generation new SNNs, on the same hardware – concurrently, if needed. This last feature has proven to be an important differentiator because it can perform tasks where CNNs typically shine, such as extraction of features from given data , but do it far more efficiently by running the inference in the event domain.
Like the biological brain, Akida chips can learn in real-time, and keep learning. With this capability, plus its ease of reconfiguration, products built with Akida can switch tasks quickly, in the order of milliseconds. Once learned, a task can be copied from one device to many other devices that also keep learning.
Neuromorphic computation is far more intricate and sophisticated than conventional AI.
This level of sophistication is needed as new hardware and software technologies are enabling more intricate and advanced tasks. Spike-based learning has always been capable of recognizing repeating patterns, but now, in combination with features learned through backpropagation algorithms, it can instantly learn to recognize previously unknown objects, predict the next sensory stimulus such as odor, and other capabilities like the biological brain. Neuromorphic hardware has improved dramatically, and in future generations we can expect even more advanced neuromorphic processors that will work on the principles of the human neocortex. Next, knowledge, and eventually, intelligence!
Brainchip’s event domain neuromorphic processing technology, aka Akida, will continue to add new features that mimic, as closely as possible, the biological brain’s neural processes. With this, we can build machines capable of more complicated and more important tasks such as sequential networks, natural language processing, and other higher level brain functions . It’s AI, but it’s less artificial and more intelligent. Most importantly, it’s extremely beneficial.
BrainChip: This is our Mission
Rob Telson, Vice President of World Wide Sales
Rob brings over 20 years of sales expertise in licensing intellectual property and selling EDA technology across multiple vertical markets. Rob has had success developing sales and support organizations at small, midsize, and large companies. At ARM, a global semiconductor and software design company, Mr. Telson was Vice President of Foundry Sales worldwide and prior was Vice President of Sales for the Americas. Most recently at Synopsys, he was responsible for building and developing a business focused on disruptive technologies in the semiconductor space. Rob holds a BS in political science from the University of Arizona and a Program for Leadership Development certificate from Harvard Business School.