Conventionally, RMDs require a large amount of motion pattern data to create models, and data creation has been a challenge. In this PoC, ISID's "Act Sense" solution for extracting human movement and posture is combined with time series processing technology based on QuantumCore's proprietary reservoir computing technology Qore (patented internationally), We have confirmed that RMD with less training data can be achieved with higher accuracy than 90%.

This makes this technology effective for analyzing time-series features obtained from videos, and enables inference of human behavior patterns with high accuracy even with a small amount of training data, making it possible to introduce this technology into fields where it is difficult to create a large number of behavior patterns in advance, such as in the healthcare field where a wide variety of behavior patterns appear. In addition, the system can detect abnormal behavior in the event of actions other than the patterns learned in advance, making it possible to apply the system to the construction field, such as detecting dangerous behavior at work sites.
In addition to motion pattern estimation, inference of the next motion pattern is expected to be applied to automatic driving safety systems, such as pedestrian behavior prediction.
What is reservoir computing?
Reservoir computing refers to computing using various materials with dynamics (noise sources), such as laser wavelengths and wave-moving water surfaces, and Recurrent Neural Networks (RNNs) that take advantage of this technology have recently gaining attention as a new machine learning method. It is a supervised learning method that consists of three layers: an input layer, an intermediate layer (reservoir layer), and an output layer (readout neuron layer).
Unlike deep learning, this method performs feature extraction by using the intermediate layer as a reservoir to perform calculations. Therefore, there is no need to strengthen the feature extraction function that was necessary in deep learning through learning, and there is no need to update the weights of the intermediate layer during learning, which significantly reduces the amount of data and computational power required for calculations during learning. Note that a variety of things can be used for the reservoir as long as they have dynamics (noise sources), and currently mechanisms that use the bodies of robots and octopuses as noise sources for computation are being explored. In this narrow sense, it is a new artificial brain-type computer that uses artificial neural circuits to prepare various noise sources from which information can be extracted, added, and calculated as needed.

About QuautumCore Technology
The features of reservoir computing are described above, but it had a challenge of being less accurate than deep learning. We succeeded in solving this technical problem with our proprietary technology (patent in progress), and have succeeded in developing a multivariate time series processing (Recurrent Neural Network, RNN) solution that achieves accuracy, cost, and speed that overwhelmingly exceed the performance of deep learning (Long Short Term Memory, LSTM): RNN (Recurrent Neural Network) solution Qore series, which offers overwhelmingly higher accuracy, cost, and speed than deep learning (Long Short Term Memory: LSTM) performance.
Qore's unique feature is that it "efficiently captures data waveforms, enabling classification that exceeds LSTM with a small amount of data," and it is particularly powerful in domains where the effects of individual, environmental, and temporal differences are large (i.e., domains where rule-based inference models are difficult to apply).
For example, in the case of anomaly detection, a large amount of normal data can be collected to build an inference model, but very little anomaly data can be obtained. In such cases, Qore can be used to derive an effective inference model in real time from a small amount of anomaly data. Moreover, Qore does not require complicated parameter tuning, which has been a problem with deep learning in the past.
WebQore is a dedicated API that allows anyone to use Qore online.
(WebQore product URL:https://www.qcore.co.jp/#sec2)
Furthermore, they have succeeded in developing Qore, also known as EdgeQore, which runs on the microcomputer Raspberry Pi. Using this EdgeQore, anyone can now realize "real-time analysis using small amounts of data," which is impossible to achieve even with deep learning on-premise. For example, EdgeQore can be used in a wide range of applications that could not be realized before, such as applications where privacy protection is important, use in environments without Internet access, and applications that require high latency (response time).
About Act Sense
ISID's solution for extracting specific human movements and postures from videos uses OpenPose, a technology developed by Carnegie Mellon University (CMU) in the United States that uses deep learning to detect characteristic points of human body postures, as the video analysis engine.
About QuantumCore Inc.
QuantumCore Inc. develops and provides API and edge functions based on Qore. QuantumCore offers solutions for data that could not be fully utilized by deep learning and for needs that could not be met by conventional deep learning, through Qore's features of high accuracy, no parameter tuning, real-time learning, and low cost.
| Company Name | QuantumCore Corporation |
| launch | April 2018 |
| address (e.g. of house) | 2-14-13 Nishi-Gotanda, Shinagawa-ku, Tokyo |
| representative director (i.e. someone chosen by the board of directors from among the directors to actually represent the company) | Shingo Akiyoshi |
| Company HP | https://www.qcore.co.jp/ |
Distributed by: PR TIMES
INNOVATIVE PORT Editors INNOVATIVE PORT" is an open innovation media operated by Creww, Inc. that connects startups, entrepreneurs, entrepreneurs in recovery, and companies aiming to create new businesses on the theme of social issues.






