Temporal Filter Quantization for Energy-Efficient Image Sequence AI

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Recurrent Neural Networks (RNNs) face challenges in energy efficiency and quantization, making them unsuitable for mobile computing units, as they require significant processing energy and cannot be effectively quantized without compromising identification accuracy.

Innovation Solution

Implementing neural networks with time folds, which allows for energy-efficient quantization of filter coefficients and activations, reducing bit requirements and enabling more compact hardware designs while maintaining performance, by applying filters to time sequences and incorporating temporal and spatial folding mechanisms.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If RNN parameters are quantized to reduce energy consumption and storage requirements, then energy efficiency and compactness improve, but identification accuracy deteriorates

Engineering Contradiction:
Improveenergy consumptionVSAvoididentification accuracy
Core Design Contradiction:
Use of energy by moving objectVSMeasurement precision

Solution Approach 1:

The patent changes the parameter representation from floating-point to quantized formats (binary, ternary, quaternary values). This parameter transformation enables significant energy reduction and storage compression while maintaining identification accuracy through optimized quantization schemes and training methodologies specifically designed for quantized RNNs.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent implements dynamic quantization strategies where quantization precision is adaptively adjusted based on the specific computational context, data characteristics, and operational requirements. This allows the system to maintain high accuracy when needed while achieving maximum energy efficiency in other scenarios, resolving the contradiction between precision and energy consumption.

Inventive Principle:
Principle #15Dynamics

2Ease of operation

If RNN parameters are quantized to enable mobile computing unit operation, then device portability improves, but processing capability deteriorates

Engineering Contradiction:
Improvemobile deployment capabilityVSAvoidprocessing capability
Core Design Contradiction:
Ease of operationVSProductivity

Solution Approach 1:

The patent transforms parameters into quantized formats suitable for mobile devices with limited resources. By using binary, ternary, or quaternary quantization levels, the system achieves compact representation that fits mobile computing units while maintaining adequate processing capability through optimized algorithms and architecture design.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent segments the RNN processing into quantized operational units that can be efficiently executed on mobile hardware. This segmentation allows the system to distribute computational tasks across mobile device resources while maintaining overall processing capability despite the constraints of quantized operations.

Inventive Principle:
Principle #1Segmentation

3Quantity of substance

If strong quantization is applied to RNN parameters, then storage requirements reduce, but computational precision deteriorates

Engineering Contradiction:
Improvestorage requirementsVSAvoidcomputational precision
Core Design Contradiction:
Quantity of substanceVSManufacturing precision

Solution Approach 1:

The patent applies parameter changes by transforming continuous floating-point parameters into discrete quantized values (binary, ternary, quaternary). This transformation dramatically reduces storage requirements while computational precision is preserved through specialized training procedures and quantization-aware operations that account for the discrete nature of the parameters.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent uses copying mechanisms where full-precision parameters are maintained during training phases, and quantized versions are generated for deployment. This allows the system to leverage high-precision training data and algorithms while deploying compact quantized models, effectively decoupling the precision requirements of training from the storage constraints of deployment.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP3736742B1Machine learning system, a method, a computer program and a device for creating the machine learning system
Publication Date: 2025.01.29 ROBERT BOSCH GMBH
  • EP3736742B1 patent drawingFigure 1
  • EP3736742B1 patent drawingFigure 2
  • EP3736742B1 patent drawingFigure 3

AI summary

The invention relates to a machine learning system (10) comprising at least one temporal filter (21). An input, comprising a temporal sequence of images, is processed by the machine learning system (10) using the filter (21). The machine learning system (10) is configured to apply the filter (21) to a sequence of pixels, all of which are located at the same coordinates of the images or at the same coordinates of intermediate results. The filter coefficients of the filter (21) are quantized. The invention further relates to a method, a computer program, and a device for creating the machine learning system (10).