Hi IDD,
So it's dictionaries at 20 paces ... (I’m doing it standing because of thesaurus).
I should preface my remarks by saying that I cannot locate any information which states that Akida 1000 production version produces 98% accuracy in analysing the NaNose results, but I did find a written reference to 98% accuracy for interpretation of mechanical vibration data (infra). (This may be in a video presentation, but I don’t know which one.)
As I understand it, you are interpreting “betterperformance than the original engineering samples” as meaning “faster”. While I agree that the improved performance would include an improvement in speed, I don’t consider that this would exclude an improvement in accuracy. The hardware improvements included both layout and design, the need for which had previously been identified.
Your argument is that the improvement in accuracy derives solely from the ML ability of Akida. My problem with that proposition is that it means that PvdM and the PhDs at BrainChip Research Institute (BRI) did not think of this when they first tested the NaNose readouts. Note that the 98% vibration results were achieved after 15 training epochs.
However, the press release from 24 Feb indicates that BRI did configure and train Akida to interpret the NaNose data using AI/ML.
https://brainchipinc.com/brainchip-nanose-successfully-detect-covid19/
BRAINCHIP INC. AND NANOSE MEDICAL SUCCESSFULLY DETECTCOVID-19 IN EXHALED BREATH WITH FAST HIGH-ACCURACY RESULTS
Akida AI chip ideal for hand-held diagnostic testingdevice
Aliso Viejo, California – February 24, 2021
NaNose Medical collected samples from 130 patients andsent nanomaterial sensor data to BrainChip’s Research Institute in Perth,Western Australia, which configured and trained its Akida™ neuromorphicprocessor to interpret the data using AI/ML.
By training, PhDs are meticuluous in performing lab tests.
I think that the BRI team would have used ML/AI to optimize Akida to interpret the NaNose data.
Anil Mankar:
https://telecomkh.info/?p=3818
An edge device that can perform incremental learning within the device itself, rather than send data to the cloud, can learn continuously.
Incremental or “one-shot” learning can begin with a very small set of samples, and grow its knowledge as more data is absorbed. The ability to evolve based on more data also results in higher accuracy. When retraining is done on the device’s hardware, instead of cloud retraining, the data and application remains private and secure. As Anil points out, accuracy can be improved when new data is added.
However, accuracy in relation to the same set of data wloud not change once Akida had been trained.
Taking into account the design changes and the thoroughness of our researchers, I still think that the improvement in accuracy can, at least in part, be attributed to the hardware changes in the production version of Akida 1000 SoC.
####################################################################
Vibration Test
https://brainchipinc.com/automotive/
The MetaTF Development Environment provides a high-level neural network APIto facilitate the development and emulation of Akida neural network models. TheADE is written in C ++, and largely inspired by the Keras API. The CWRU (Case Western Reserve University) data setwas used to build a reference networkfor vibration analysis. Data from each class originally comprising of more than480,000 continuous values wascollected and divided into segments of 1024 values.
Classification experiments were conducted onthe raw time-series data and continuous wavelet transform (CWT) of thetime-varying data. Continuous Wavelet Transform is a method to decompose a realsignal into a set of elementary waveforms that provide a way to analyze thesignal by examining different components related to its wavelets. The Akida model obtained an overallaccuracy of 98% for the classification of raw vibration data without anypreprocessing and with only 15 epochs for training.
In order to investigate the effects ofpreprocessing instead of using raw data, CWT features were extracted. Withoutchanging any additional parameters, the classification accuracy of theAkida model improved by 1% thus resulting in an overall accuracy of 99%. Thisis an excellent result that provides a point of reference for application ofthe Akida Neuromorphic processor in bearing fault detection.
(20min delay)
