RNN Time-Window Topology Analysis for Robust Decisions

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Recurrent neural networks (RNNs) are susceptible to adversarial perturbations and noise, leading to inconsistent and unreliable decision-making, particularly when processing data that deviates from the training set.

Innovation Solution

Implementing a window definition unit in RNNs to define different windows of time with varying durations and start times, allowing for the identification and reinforcement of robust topological patterns of activity, and attenuation of irrelevant patterns, thereby enhancing decision-making robustness.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If recurrent neural networks process data that deviates from the training set, then the network can handle diverse input types, but the processing results become inconsistent and unreliable

Engineering Contradiction:

Improveability to process diverse input typesVSAvoidconsistency of processing results

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The patent applies preliminary action by defining multiple time windows and identifying topological patterns before making decisions. The system pre-processes the temporal dynamics by segmenting the input sequence into different time windows, identifying topological patterns in each window, and comparing these patterns across windows to determine decision robustness before final classification occurs. This preliminary analysis of temporal patterns helps ensure reliable processing even when input data deviates from training distributions.

Inventive Principle:

Principle #10Preliminary action

2Productivity

If the RNN uses temporal dynamic behavior to process information, then the network can capture sequential dependencies, but the network becomes susceptible to adversarial perturbations and noise

Engineering Contradiction:

Improveinformation processing capabilityVSAvoidsusceptibility to adversarial perturbations

Core Design Contradiction:

ProductivityVSObject-affected harmful factors

Solution Approach 1:

The patent converts the harmful effect of temporal dynamics (which amplifies adversarial perturbations) into a beneficial feature by using it to identify and compare topological patterns across multiple time windows. The system leverages the temporal evolution of patterns to distinguish between genuine input characteristics and adversarial noise, comparing pattern consistency across windows to detect and mitigate the impact of perturbations.

Inventive Principle:

Principle #22Blessing in disguise (Convert harm into benefit)

Solution Approach 2:

The patent introduces topological patterns as an intermediary between the raw input data and the final decision. Instead of directly processing raw inputs through temporal dynamics, the system first extracts topological patterns from different time windows, then compares these patterns to determine decision robustness. This intermediary representation filters out noise and adversarial perturbations while preserving meaningful sequential dependencies.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Reliability

If the RNN reinforces relevant processing results over time, then the decision robustness improves, but the processing time increases

Engineering Contradiction:

Improvedecision robustnessVSAvoidprocessing time

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-defining multiple time windows with different durations and start times before processing the input sequence. This allows the system to efficiently capture relevant temporal patterns at different scales without requiring excessive processing time, as the window structure is predetermined and optimized for the specific task.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent uses partial action by selectively comparing topological patterns only across the predefined time windows that are most relevant to the decision, rather than analyzing the entire temporal sequence in detail. This selective comparison approach maintains decision robustness while reducing overall processing time by focusing computational resources on critical temporal patterns.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS20250390742A1Interpreting and improving the processing results of recurrent neural networks

Publication Date: 2025.12.25 INAIT SA

AI summary

A method includes defining a plurality of different windows of time in a recurrent artificial neural network, wherein each of the different windows has different durations, has different start times, or has both different durations and different start times, identifying occurrences of topological patterns of activity in the recurrent artificial neural network in the different windows of time, comparing the occurrences of the topological patterns of activity in the different windows, and classifying, based on a result of the comparison, a first decision that is represented by a first topological pattern of activity that occurs in a first of the windows as less robust than a second decision that is represented by a second topological pattern of activity that occurs in a second of the windows.