Rehearsal Network for One-Shot Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural networks face challenges in learning from very few examples, such as one-shot learning, and traditional collaborative filtering does not account for individual user taste and preference variations, leading to inefficiencies in learning and resource consumption.
Innovation Solution
A rehearsal network service that uses biological memory indicators, including saliency, valence, and timestamp values, to select and generate new input data for training, allowing for generalized learning by recalculating novelty and effective salience values and applying data augmentation techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional neural networks are used to learn user preferences, then they can learn from data, but they require a large set of examples which increases data requirements and training resources
Solution Approach 1:
The patent creates synthetic training examples by copying and transforming existing data points through data augmentation techniques. The system generates new training samples by applying transformations to salient existing examples, allowing the neural network to learn from fewer original examples while maintaining high learning accuracy.
Solution Approach 2:
The system performs preliminary identification and selection of salient data points before full training occurs. By pre-processing data to identify highly relevant examples and preparing augmented versions in advance, the system reduces the amount of data needed during actual training while maintaining learning effectiveness.
2Adaptability or versatility
If traditional collaborative filtering is used, then it can process user data, but it does not account for individual user taste and preference variations leading to reduced personalization accuracy
Solution Approach 1:
The patent applies local quality by treating each user's preference data with specialized processing. The system identifies salient data points specific to individual users and applies targeted data augmentation and rehearsal techniques tailored to each user's preference patterns, rather than applying uniform processing to all users.
Solution Approach 2:
The system segments user preference data into distinct salient examples that represent different aspects of individual user tastes. By dividing the data into meaningful segments and applying selective rehearsal and augmentation to each segment, the system captures individual variations in user preferences more effectively.
3Reliability
If neural networks train with large datasets, then they achieve better learning performance, but it increases computational resources and training time
Solution Approach 1:
The patent extracts only the most salient and informative data points from the available dataset for training. By identifying and selecting only the highly relevant examples through salience measurement, the system eliminates the need to process entire large datasets, significantly reducing computational resource consumption while maintaining learning performance.
Solution Approach 2:
The system applies partial action by focusing training efforts on a selective subset of salient examples rather than processing all available data. The data augmentation techniques generate additional training samples from this selective subset, providing sufficient training material without requiring exhaustive processing of complete large datasets.
4Use of energy by moving object
If few examples are used for training, then resource consumption is reduced, but the neural network cannot learn effectively from very few examples
Solution Approach 1:
The patent compensates for the small number of original training examples by creating multiple synthetic copies through data augmentation. The system generates additional training samples by applying various transformations to the limited salient examples, effectively multiplying the training value of each original data point without requiring additional computational resources for data collection and processing.
Data Source
AI summary
A method, a device, and a non-transitory storage medium are described in which a rehearsal network service is provided that enables generalized learning for all types of input patterns ranging from one-shot inputs to a large set of inputs. The rehearsal network service includes using biological memory indicator data relating to a user and the input data. The rehearsal network service includes calculating a normalized effective salience for each input data, and generating a new set of input data in which the inclusion of input data is proportional to its normalization effective salience. The rehearsal network service provides the new set of input data to a learning network, such as a neural network or a deep learning network that can learn the user's taste or preference.