Artificial intelligence processing method supporting online learning and processor

Abstract

The invention provides an artificial intelligence processing method supporting online learning and a processor. By adopting a double-judgment mode combining adaptive template matching, adaptive high / low threshold judgment and neural network judgment, the system has an online learning capability, can adapt to time sequence signal characteristics of using equipment and users, solves the problem thatthe common time sequence signal characteristics have individual differences, and effectively improves the automatic classification accuracy. Compared with a traditional neural network back propagation algorithm for achieving hardware implementation of online learning, the method is small in parameter quantity and operand, only a small amount of data is stored, and compared with online learning ofthe neural network, the power consumption and the needed storage quantity are greatly reduced. All parameters in the online learning hardware module can be freely configured according to different application requirements, so that an online learning algorithm can be flexibly deployed on a processor, and the online learning hardware module is suitable for application scenes of automatic classification of various time sequence signals.

Description

technical field [0001] The invention relates to an online learning technology of time series signals, in particular to an artificial intelligence processing method and a processor supporting online learning. Background technique [0002] In real-world scenarios, sensors often collect a variety of time-series signals, such as pickups collecting vocal signals, and wearable electrodes collecting muscle electrical signals, brain electrical signals, and electrocardiographic signals and other physiological electrical signals. This type of time series signal, according to the morphological characteristics of its waveform, can represent the different states of the signal source or reflect the different properties of the signal source, so as to extract useful information for signal classification processing. For example, in the human voice signal, the pronunciation waveforms of the same word are similar, but the pronunciation waveforms of different words are quite different. Therefor...

AI Technical Summary

Problems solved by technology

However, the manually extracted time-frequency domain features are often unable to summarize the various categories of signals well, the generalization ability is weak, and they do not have online learning ability and cannot adapt to the signal source differences of time series signals, so the accuracy rate will be relatively low.

The signal classification technology based on neural network uses a large amount of training data and labels to automatically learn the waveform characteristics of the signal. More parameters, higher power consumption and hardware resource consumption

And because the training data generally does not contain the data of the user or the equipment used, or these data are unknown, and because there are generally individual differences in timing signals, that is, different devices of the same nature / state, timing signals of users, and even There may be large differences in the signals of the same device and users at different times, which leads to the possibility that the neural network parameters may overfit the data in the training set, and extract unnecessary classification features, so that they can be directly applied in the use Automatically classify the signals of equipment and users, and the classification accuracy will drop. This is because the training set does not contain information about such signals of equipment and users, and the generalization ability of the model for time series signals of equipment and users is not strong.

In order to improve the generalization ability of the model, the algorithm needs to conduct online learning based on the data of the equipment and users, and adjust the relevant parameters of the algorithm to adapt to the individual differences of the users. The traditional online learning algorithm of neural network, such as the back propagation algorithm, involves A large number of vector matrix multiplication and addition operations will result in high processing power consumption and hardware resource consumption

Images