What a wonderful read, although my comprehension of all the new capabilities is somewhat lacking.
However, I think most of us understand real world comparisons and benchmarking.
Hence, I think this extract from the article is particularly powerful:
......
And so we come to the old proverb that states, “The proof of the pudding is in the eating.” Just how well does the Akida perform with industry-standard, real-world benchmarks?
Well, the lads and lasses at
Prophesee.ai are working on some of the world’s most advanced neuromorphic vision systems. From their website we read: “Inspired by human vision, Prophesee’s technology uses a patented sensor design and AI algorithms that mimic the eye and brain to reveal what was invisible until now using standard frame-based technology.”
According to the paper
Learning to Detect Objects with a 1 Megapixel Event Camera, Gray.Retinanet is the latest state-of-the-art in event-camera based object detection. When working with the Prophesee Event Camera Road Scene Object Detection Dataset at a resolution of 1280×720, the Akida achieved 30% better precision while using 50X fewer parameters (0.576M compared to 32.8M with Gray.Retinanet) and 30X fewer operations (94B MACs/sec versus 2432B MACs/sec with Gray.Retinanet). The result was improved performance (including better learning and object detection) with a substantially smaller model (requiring less memory and less load on the system) and much greater efficiency (a lot less time and energy to compute).
As another example, if we move to a frame-based camera with a resolution of 1352×512 using the KITTI 2D Dataset, then ResNet-50 is kind of a standard benchmark today. In this case, Akida returns equivalent precision using 50X fewer parameters (0.57M vs. 26M) and 5X fewer operations (18B MACs/sec vs. 82B MACs/sec) while providing much greater efficiency (75mW at 30 frames per second in a 16nm device). This is the sort of efficiency and performance that could be supported by untethered or battery-operated cameras.
Another very interesting application area involves networks that are targeted at 1D data. One example would be processing raw audio data without the need for all the traditional signal conditioning and hardware filtering.
Consider today’s generic solution as depicted on the left-hand side of the image below. This solution is based on the combination of Mel-frequency cepstral coefficients (MFCCs) and a depth-wise separable CNN (DSCNN). In addition to hardware filtering, transforms, and encoding, this memory-intensive solution involves a heavy software load.
Raw audio processing: Traditional solution (left) vs. Akida solution (right)
(Source: BrainChip)
By comparison, as we see on the right-hand side of the image, the raw audio signal can be fed directly into an Akida TENN with no additional filtering or DSP hardware. The result is to increase the accuracy from 92% to 97%, lower the memory (26kB vs. 93kB), and use 16X fewer operations (19M MACs/sec vs. 320M MACs/sec). All of this basically returns single inference while consuming two microjoules of energy. Looking at this another way, assuming 15 inferences per second, we’re talking less than 100µW for always-on keyword detection.
Similar 1D data is found in the medical arena for tasks like vital signs prediction based on a patient’s heart rate or respiratory rate. Preprocessing techniques don’t work well with this kind of data, which means we must work with raw signals. Akida’s TENNs do really well with raw data of this type.
In this case, comparisons are made between Akida and the state-of-the-art S4 (SOTA) algorithm (where S4 stands for structured state space sequence model) with respect to vital signs prediction based on heart rate or respiratory rate using the Beth Israel Deaconess Medical Center Dataset. In the case of respiration, Akida achieves ~SOTA accuracy with 2.5X fewer parameters (128k vs. 300k) and 80X fewer operations (0.142B MACs/sec vs. 11.2B MACs/sec). Meanwhile, in the case of heart rate, Akida achieves ~SOTA accuracy with 5X fewer parameters (63k vs. 600k) and 500X fewer operations (0.02B MACs/sec vs. 11.2B MACs/sec).
It’s impossible to list all the applications for which Akida could be used. In the case of industrial, obvious apps are robotics, predictive maintenance, and manufacturing management. When it comes to automotive, there’s real-time sensing and the in-cabin experience. In the case of health and wellness, we have vital signs monitoring and prediction; also, sensory augmentation. There are also smart home and smart city applications like security, surveillance, personalization, and proactive maintenance. And all of these are just scratching the surface of what is possible.
