Deep Neural Network Node Addition for Efficient Subclass Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep neural network (DNN) techniques require large amounts of data and time to learn, making it difficult to easily acquire a DNN capable of performing desired image recognition tasks, especially when transitioning from one subclass to another.
Innovation Solution
A calculation device that adds a new node to an existing DNN, calculates coupling coefficients between this new node and other nodes using a backpropagation method, and adjusts these coefficients to minimize errors in recognizing features of data from a second subclass, allowing for efficient learning and adaptation without requiring extensive new training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If learning is performed by using large amounts of data to acquire a DNN, then determination accuracy is improved, but loss of time and loss of substance increase
Solution Approach 1:
The patent applies preliminary action by pre-learning common features shared across multiple subclasses before performing subclass-specific learning. The calculation device first learns features common to all subclasses in a predetermined class, then uses this pre-learned knowledge as a foundation for subsequent subclass-specific learning. This preliminary learning of commonalities reduces the amount of data and time needed for each individual subclass learning task, directly resolving the contradiction between achieving high determination accuracy and reducing learning time.
2Measurement precision
If learning is performed by using large amounts of data to acquire a DNN, then determination accuracy is improved, but loss of substance increases
Solution Approach 1:
The calculation device performs preliminary learning of common features that are shared across multiple subclasses before conducting subclass-specific learning. By extracting and storing these commonalities in advance, the system reduces the amount of training data required for each individual subclass task, thereby resolving the contradiction between achieving high determination accuracy and reducing the quantity of training data needed.
Solution Approach 2:
The patent implements universality by creating a DNN structure that learns universal common features applicable to multiple subclasses simultaneously. The calculation device designs the network to identify and learn features that are common across different subclasses within a predetermined class, making the learned knowledge universally applicable. This multi-functional approach allows the same learned features to serve multiple subclass recognition tasks, reducing the overall data requirement while maintaining high determination accuracy.
3Reliability
If a DNN is acquired by learning with large amounts of data, then reliability is improved, but device complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the learning process into distinct stages: first learning common features shared across multiple subclasses, then performing subclass-specific learning based on the pre-learned commonalities. This segmented approach breaks down the complex task of learning multiple subclasses into manageable stages, reducing the overall complexity of the learning process while maintaining reliable DNN performance through systematic progressive learning.
Data Source
AI summary
A calculation device includes an adding unit configured to add at least one new node to a network, which has multiple nodes that output results of calculations on input data are connected and which learned a feature of data belonging to a first subclass contained in a predetermined class. The calculation device includes an accepting unit configured to accept, as input data, training data belonging to a second subclass contained in the predetermined class. The calculation device includes a calculation unit configured to calculate coupling coefficients between the new node added by the adding unit and other nodes to learn a feature of the training data belonging to the second subclass based on an output result obtained when the training data accepted by the accepting unit is input to the network.