Recurrent Sequence Recognition with Prediction-Classification Feedback

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current sequence recognition systems, such as speech recognition systems, face challenges in accurately predicting and classifying sequences due to limitations in leveraging prediction information for improved classification results, particularly in handling large vocabulary tasks with numerous states.

Innovation Solution

A sequence recognition system incorporating a machine learning framework with a recurrent neural network architecture, featuring a prediction component and a classification component in a recurrent loop, where feedback from the classification component adapts the prediction component, and vice versa, to dynamically improve predictions and classifications, utilizing deep neural networks with bottleneck layers for dimension reduction and sigmoid functions for nonlinearity.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional sequence recognition systems are used, then the system structure is simple, but the accuracy of state predictions and classifications deteriorates

Engineering Contradiction:
Improveaccuracy of state predictionsVSAvoidsystem structure complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system is divided into distinct functional modules: a prediction component that generates predicted states, a classification component that performs classification based on predicted and actual states, and a training component that coordinates multi-objective training. This segmentation allows each module to specialize in specific tasks, improving overall prediction accuracy while maintaining manageable system complexity through modular design.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system implements feedback mechanisms where the classification component receives both actual states and predicted states, and the training component uses classification results to adjust prediction models. The recurrent loop allows the system to continuously refine predictions based on classification feedback, thereby improving measurement precision through iterative optimization.

Inventive Principle:
Principle #23Feedback

2Measurement precision

If deep neural networks with bottleneck layers are used, then the accuracy of classifications improves, but the computational complexity increases

Engineering Contradiction:
Improveaccuracy of classificationsVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The bottleneck layer extracts essential features from high-dimensional input data by projecting them into a lower-dimensional space. This extraction process retains the most discriminative information needed for accurate classification while discarding redundant features, thereby improving classification accuracy without proportionally increasing computational complexity.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system changes parameters dynamically during training and operation, adjusting the dimensionality of hidden layers and the number of training iterations based on performance metrics. This adaptive parameter adjustment allows the system to optimize classification accuracy while controlling computational complexity by scaling resources according to actual needs.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If recurrent neural network architecture with feedback loops is used, then the reliability of predictions improves, but the device complexity increases

Engineering Contradiction:
Improvereliability of predictionsVSAvoidarchitecture complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The prediction component performs preliminary predictions of future states before actual observations are fully processed. This preliminary action allows the system to prepare prediction hypotheses in advance, which are then refined by the classification component using actual data, thereby improving prediction reliability through proactive preparation rather than reactive processing.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The recurrent architecture maintains continuous operation where the prediction component continuously generates predicted states and the classification component continuously refines them based on incoming data. This continuous useful action ensures that the system constantly improves prediction reliability through ongoing feedback and adjustment, rather than operating in discrete, disconnected cycles.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP3218901B1Prediction-based sequence recognition
Publication Date: 2019.12.25 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP3218901B1 patent drawingFigure 1
  • EP3218901B1 patent drawingFigure 2
  • EP3218901B1 patent drawingFigure 3

AI summary

A sequence recognition system comprises a prediction component configured to receive a set of observed features from a signal to be recognized and to output a prediction output indicative of a predicted recognition based on the set of observed features. The sequence recognition system also comprises a classification component configured to receive the prediction output and to output a label indicative of recognition of the signal based on the prediction output.