Latent-Space Confidence Estimation for Deep Learning Predictions

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Deep learning models lack the ability to assess uncertainty in their predictions, leading to reliability and trust issues in real-world applications, as they provide outputs with the same confidence for inappropriate inputs, and existing methods for uncertainty estimation are limited to specific probability distributions and require the entire training dataset.

Innovation Solution

Model the latent space of deep learning networks using probabilistic techniques to generate confidence predictions, allowing the rejection of inputs that lead to outputs below a predetermined threshold, thereby enhancing the robustness and reliability of deep learning models.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If deep learning models provide predictions for all input data, then productivity is improved, but reliability deteriorates due to inability to assess uncertainty

Engineering Contradiction:
Improveprediction throughputVSAvoidprediction trustworthiness
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent introduces latent variables as an intermediary between input data and final predictions. These latent variables serve as a probabilistic bridge that allows the model to assess uncertainty while maintaining prediction capabilities. The latent space represents the distribution of possible interpretations, enabling the system to distinguish between confident and uncertain predictions without reducing overall productivity.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If Bayesian modelling is used to assess uncertainty, then reliability is improved, but device complexity becomes intractable due to non-linearities and extreme dimensionality

Engineering Contradiction:
Improveuncertainty assessmentVSAvoidmodelling complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent extracts the probabilistic reasoning from the entire deep network and concentrates it in a small latent space. Instead of performing Bayesian inference across the full high-dimensional network, the method isolates uncertainty assessment to a low-dimensional latent representation. This extraction makes probabilistic modelling tractable while preserving reliability benefits.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the parameterization of the problem by representing uncertainty not in the original high-dimensional input space but in a transformed latent space with fewer parameters. This parameter reduction from extreme dimensionality to a compact latent representation makes Bayesian-style uncertainty assessment computationally feasible.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If confidence threshold filtering is applied, then reliability is improved, but productivity decreases due to rejection of valid predictions

Engineering Contradiction:
Improveprediction confidenceVSAvoidprediction volume
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent performs preliminary assessment of prediction confidence through latent variable sampling before final decision-making. By evaluating uncertainty in the latent space ahead of time, the system can identify low-confidence cases that would benefit from human review or additional processing, while allowing high-confidence cases to proceed automatically, thus minimizing productivity loss.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3983957B1Method and system for confidence estimation of a trained deep learning model
Publication Date: 2026.04.08 THINKSONO LTD
  • EP3983957B1 patent drawingFigure 1
  • EP3983957B1 patent drawingFigure 2
  • EP3983957B1 patent drawingFigure 3

AI summary

The present invention relates to a method and system of determining a measure of confidence for a trained deep learning model. Further, the present invention relates to a method and system of predicting a measure of confidence for a trained deep learning model using the latent variables of the trained deep learning network. Aspects and/or embodiments recite a method and/or system that, by modelling a portion of the latent space with probabilistic techniques, allows network prediction to be sampled and tested for robustness in order to derive a measure of certainty/uncertainty. This measure of certainty/uncertainty, according to aspects and/or embodiments, can be used to reject network inputs that will lead to outputs or predictions below a predetermined decision confidence threshold. Aspects and/or embodiments can augment substantially any deep learning model/network that uses an expressive and substantially small latent space (for example, having fewer than approximately 50,000 values) and substantially boost sensitivity and/or substantially boost specificity of these models/networks. Aspects and/or embodiments can provide a control mechanism for deep neural networks that can require a user-determined level of prediction confidence in the outputs of the deep neural networks.