Local learning method and device and medium

Abstract

The invention discloses a local learning method and device and a medium. The method comprises the following steps: receiving an input data stream through a TinyML model, reasoning the data stream through the TinyML model to obtain a reasoning data stream, and sending the reasoning data stream to a local learning module; inputting the reasoning data stream into an additional neural network layer of a local learning module, and performing reasoning through the additional neural network layer to obtain a prediction result; a label flow is obtained according to the data flow, an update matrix is determined according to the prediction result and the label flow, and the update matrix is sent to an update layer of a local learning module; and determining an update weight parameter according to the update matrix, and sending the update weight parameter to the additional neural network layer so as to update the additional neural network layer according to the update weight parameter. According to the invention, resource-limited MCU online learning is realized through the local learning system, the model weight is updated locally according to the collected data, and the constraint and limitation of the MCU are effectively solved.

Description

technical field [0001] The present application relates to the technical field of microcontrollers, and in particular, to a local learning method, device and medium. Background technique [0002] Microcontroller Unit (MCU), also known as Single Chip Microcomputer (Single Chip Microcomputer) or single chip microcomputer, is to appropriately reduce the frequency and specifications of the central processing unit (Central Process Unit, CPU), and reduce the frequency and specifications of the memory (memory), Counter (Timer), USB, A / D conversion, UART, PLC, DMA and other peripheral interfaces, and even the LCD driver circuit are integrated on a single chip to form a chip-level computer, which can be controlled in different combinations for different applications. [0003] Tiny Machine Learning (TinyML) is a subfield of machine learning that includes algorithms, hardware, and software that can analyze sensor data and run on extremely low-power devices. [0004] At present, through...

AI Technical Summary

Problems solved by technology

Static objects In the case of learning from new data, the model needs to be retrained from scratch and re-uploaded to the MCU, which makes the deployment of TinyML in an industrial environment a challenging task

Images