Aggregated Image Representations for Accurate Compact CNN Inference

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional techniques for compressing machine learning models face a tradeoff between the amount of compression and accuracy, with compressed models still being too large or complex for devices with limited resources, and existing methods fail to maintain accuracy when compressed beyond a certain point.

Innovation Solution

Generate a simplified representation of a trained machine learning model by creating a mapping between a set of images and their output classes, using aggregated or logical representations of pixel values, allowing inference operations to be performed efficiently without executing the full model.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a machine learning model is compressed using conventional techniques, then resource overhead and latency are reduced, but accuracy is adversely impacted

Engineering Contradiction:

Improveinference speedVSAvoidprediction accuracy

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The patent creates simplified representations (copies) of the trained machine learning model that capture essential classification behavior without replicating the full model complexity. These simplified models use aggregated image representations and decision rules to mimic the original model's predictions for specific output classes, enabling accurate inference without executing the complete complex model.

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The patent extracts specific classification knowledge from the trained machine learning model by identifying representative images for each output class and creating simplified decision rules based on these extracted patterns. This extraction process separates essential classification logic from the full model complexity, allowing the simplified representation to handle inference for certain classes without requiring the entire original model.

Inventive Principle:

Principle #2Taking out (Extraction)

2Weight of stationary object

If a machine learning model is compressed too much, then the model size is reduced, but accuracy becomes adversely impacted

Engineering Contradiction:

Improvemodel sizeVSAvoidprediction accuracy

Core Design Contradiction:

Weight of stationary objectVSMeasurement precision

Solution Approach 1:

The patent transforms the model representation by changing parameters from full model weights and architectures to simplified structures consisting of aggregated image representations, pixel value comparisons, and decision rules. This parameter transformation dramatically reduces model size while preserving classification accuracy for specific output classes by maintaining the essential decision-making patterns in a compact form.

Inventive Principle:

Principle #35Parameter changes

3Use of energy by moving object

If conventional compression techniques are applied, then resource overhead is reduced, but the compressed model remains too large for devices with limited resources

Engineering Contradiction:

Improvecomputational resourcesVSAvoidmodel size

Core Design Contradiction:

Use of energy by moving objectVSWeight of stationary object

Solution Approach 1:

The patent creates extremely lightweight simplified representations that can be deployed to resource-constrained devices. These simplified models use minimal memory storage (aggregated representations rather than full model weights) and require negligible computational resources during inference, making them suitable for deployment on devices with limited RAM, processing power, and energy capacity.

Inventive Principle:

Principle #27Cheap short-living objects (Disposable)

4Productivity

If a simplified representation is generated using aggregated representations, then inference operations become faster and less resource-intensive, but the complexity of generating the simplified representation increases

Engineering Contradiction:

Improveinference speedVSAvoidsimplification process complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The patent performs the complex work of creating simplified representations during an offline training phase, before deployment to resource-constrained devices. This preliminary action includes training the full model, identifying representative images for each class, generating aggregated representations, and creating decision rules. Once created, these simplified representations can be deployed and executed efficiently on devices with limited computational resources, as the complex generation process has already been completed.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS12380676B2Simplifying convolutional neural networks using aggregated representations of images

Publication Date: 2025.08.05 VIAN SYSTEMS INC

AI summary

One embodiment of the present invention sets forth a technique for simplifying a trained machine learning model. The technique includes determining a first set of images associated with a first output class predicted by the trained machine learning model. The technique also includes generating a first aggregated representation of the first set of images, wherein the first aggregated representation includes a first plurality of representative pixel values for a plurality of pixel locations included in the first set of images. The technique further includes generating a simplified representation of the trained machine learning model that includes a first mapping of the first aggregated representation to the first output class, wherein the first mapping indicates that the trained machine learning model predicts the first output class for one or more input images.