Sparse Predictive Hierarchy Network for Temporal Data Processing

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current systems for processing high-dimensional, high-velocity data require significant expertise, are costly, and often rely on supervised machine learning methods that demand extensive labeled data and high computational resources, making them inefficient and infeasible for many applications.

Innovation Solution

The development of a Sparse Predictive Hierarchy Network with autoencoder-like layers and recurrent connections that enable efficient processing and prediction of temporal data, allowing for online learning and adaptation without pretraining, and can be distributed across heterogeneous devices, reducing computational overhead.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If supervised machine learning methods are used to process high-dimensional, high-velocity data, then prediction accuracy can be improved, but computational costs and resource requirements increase significantly

Engineering Contradiction:
Improveprediction accuracyVSAvoidcomputational costs
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The system segments the data processing task by separating encoding operations (performed offline during training) from decoding operations (performed online during inference). The encoder-compressor processes data offline to create compact representations, while the decoder-predictor performs lightweight online prediction. This segmentation shifts computational burden from resource-constrained online devices to more powerful offline processing systems.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The invention uses simple, computationally inexpensive decoder models that can be deployed on resource-constrained devices. These decoders are designed to be lightweight and efficient, sacrificing some complexity to achieve low computational costs and fast inference on edge devices with limited resources.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

2Productivity

If complex hardware and software systems are developed to process sensor data, then processing capability is improved, but system cost and maintenance expenses increase

Engineering Contradiction:
Improveprocessing capabilityVSAvoidsystem complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The system is divided into distinct functional modules: encoder-compressor for offline processing and decoder-predictor for online processing. This modular segmentation allows each component to be optimized independently and deployed on different hardware platforms, reducing overall system complexity while maintaining high processing capability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The framework is designed to be application-agnostic and domain-independent, capable of processing various types of sensor data (video, audio, physiological signals, industrial sensor data) using the same core architecture. This universality reduces the need for application-specific customizations and lowers maintenance costs.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If experts are employed at every stage of system development and operation, then system performance is improved, but operational costs and time requirements increase

Engineering Contradiction:
Improvesystem performanceVSAvoidoperational time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs automatic online learning and adaptation without requiring expert intervention. The decoder-predictor continuously learns from incoming data streams and adapts to changing conditions autonomously, eliminating the need for manual tuning or expert operation while maintaining high performance.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

Complex processing and learning operations are performed in advance during the offline encoding phase. The encoder-compressor pre-processes data and creates compact representations that capture essential patterns, so that online inference requires minimal processing and no expert intervention.

Inventive Principle:
Principle #10Preliminary action

4Measurement precision

If supervised machine learning requires large amounts of labeled training data, then model accuracy is improved, but data preparation costs and expertise requirements increase

Engineering Contradiction:
Improvemodel accuracyVSAvoiddata preparation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

Instead of requiring labeled data to train the model, the system inverts the approach: the encoder is trained offline on unlabeled data to learn efficient representations, and the decoder is trained on the compact encoded representations rather than raw data. This inversion eliminates the need for expensive labeled data preparation while maintaining model accuracy.

Inventive Principle:
Principle #13The other way round (Inversion)

Data Source

PatentEP3465551B1Processing time-varying data streams using sparse predictive hierarchies
Publication Date: 2024.05.15 OGMA INTELLIGENT SYST CORP
  • EP3465551B1 patent drawingFigure 1
  • EP3465551B1 patent drawingFigure 2
  • EP3465551B1 patent drawingFigure 3

AI summary

A technique for processing streams of time-varying data provides first through N th processing stages, each processing stage having a respective encoder and a respective decoder. The encoder of each processing stage receives successive values of time-varying input data and generates therefrom encoded output data. Each encoder provides its encoded output data as input to its respective decoder, which generates a prediction of a next value of the time-varying input data that the encoder will receive. Each encoded output is based upon both (1) a current input value and (2) one or more previous input values, such that encodings are based at least in part on history. Encoders are coupled output-to-input, with the input of the first encoder receiving an overall processing input. Decoders are likewise coupled output-to-input, with each decoder providing feedback to the previous decoder and the first decoder providing an overall processing output.