Federated Learning for Handwritten Character Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for recognizing handwritten characters using OCR transmit raw data to servers, leading to potential personal information leakage and difficulties in collecting large datasets, as they rely on centralized data processing rather than decentralized learning.
Innovation Solution
A federated learning method that allows terminals to receive and process handwritten characters, extract feature values, and update a common prediction model using user feedback for error correction, with weight parameters being transmitted to a server for model improvement without sharing raw data, enabling local data caching and continuous model refinement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If raw handwritten character data is transmitted to a server for centralized processing, then the recognition model can be trained using large datasets, but personal information may be leaked and data collection becomes difficult
Solution Approach 1:
The patent extracts only the necessary model training components (weight parameters and gradient information) from the terminal devices, while leaving the raw handwritten character data localized on user devices. This extraction approach enables model training without transmitting sensitive personal information to servers, resolving the contradiction between data quantity needs and privacy protection.
Solution Approach 2:
The patent introduces federated learning as an intermediary mechanism that enables collaborative model training across multiple terminals without direct data sharing. The server acts as a coordinator that aggregates model updates (weight parameters and gradients) from various terminals, facilitating large-scale training while maintaining data locality and user privacy.
2Measurement precision
If centralized data processing is used to train the recognition model, then model accuracy can be improved, but user privacy and data security are compromised
Solution Approach 1:
The patent extracts only the essential model training signals (weight parameters and gradient information) from terminal devices, eliminating the need to transmit or store raw handwritten data on servers. This extraction strategy maintains recognition accuracy by preserving model learning capabilities while enhancing data security through localized data processing.
Solution Approach 2:
Each terminal device performs local model training using its own handwritten character data, updating its local recognition model without requiring centralized data collection. This self-service approach enables each device to improve its own recognition accuracy while contributing to the overall model improvement through shared weight parameter updates, thereby maintaining both accuracy and security.
3Measurement precision
If user feedback for error correction is collected and stored, then model learning can be improved, but storage requirements and data management complexity increase
Solution Approach 1:
The patent extracts only the corrected text information from user feedback, converting it into gradient information that guides model improvement. This extraction approach transforms complex unstructured feedback data into structured training signals, improving model learning while simplifying data management by storing only essential correction information rather than complete datasets.
Solution Approach 2:
The patent transforms user feedback (corrected text) into gradient parameters that update the model's weight parameters. This parameter transformation converts human-readable correction information into mathematical gradients that directly improve model accuracy, thereby enhancing learning effectiveness while maintaining simple data storage requirements.
Data Source
AI summary
Provided is a method for recognizing handwritten characters in a terminal through federated learning. In the method, a first common prediction model for recognizing text from handwritten characters input from a user is applied, the handwritten characters are received from the user, feature values are extracted from an image including the handwritten characters, the feature values are input to the first common prediction mode, first text information is determined from an output of the first common prediction model, the first text information and a second text information received from the user for error correction of the first text information are cached, and the first common prediction model is learned using the image including the handwritten characters, the first text information, and the second text information. In this way, the terminal can determine the text from the handwritten characters input by the user, and can learn the first common prediction model through a feedback operation of the user.