Augmented Neural Network Sequence Processing for Output Accuracy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Neural networks face challenges in determining when a final output has been generated for a given input, leading to potential inaccuracies and inefficiencies, particularly for complex inputs, as they may produce incomplete or unnecessary additional outputs.

Innovation Solution

An augmented neural network system that includes a sequence processing subsystem to determine when a final output has been generated, allowing for more accurate and efficient processing by deciding whether to produce additional outputs based on designated portions of the neural network output, and utilizing an external memory for enhanced performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the neural network generates multiple additional outputs for complex inputs, then the accuracy of the final output is improved, but the processing time and computing resources increase

Engineering Contradiction:
Improveoutput accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system dynamically adjusts the number of processing passes based on input complexity. The sequence processing subsystem monitors intermediate outputs and determines whether to generate additional outputs or terminate processing early, making the processing depth adaptive rather than fixed. This resolves the contradiction by allowing multiple passes only when necessary for complex inputs while maintaining single-pass efficiency for simpler cases.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system changes the parameter of output generation from a fixed multiple outputs approach to a conditional single or multiple outputs approach. By using a designated portion of the neural network output to control whether additional outputs are generated, the system optimizes the balance between accuracy and processing time based on the specific input characteristics.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If the neural network generates additional outputs for all inputs, then the reliability of the final output is improved, but the computing resources required increase

Engineering Contradiction:
Improveoutput reliabilityVSAvoidcomputing resources
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The system applies different processing strategies to different inputs based on their characteristics. By examining a designated portion of the neural network output and making localized decisions about whether additional outputs are needed, the system ensures high reliability only where necessary rather than uniformly applying additional processing to all inputs, thus reducing overall computing resource requirements.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The system performs partial action by generating additional outputs only when the designated portion of the neural network output indicates it is necessary. This avoids the excessive action of generating multiple outputs for all inputs, thereby maintaining reliability when needed while reducing computing resource consumption for cases where full processing is not required.

Inventive Principle:
Principle #16Partial or excessive action

3Measurement precision

If the system waits for complete outputs before further processing, then the accuracy is improved, but the productivity decreases

Engineering Contradiction:
Improveoutput completenessVSAvoidprocessing throughput
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The system performs preliminary action by generating an initial output early in the processing sequence. This initial output can be used for preliminary processing or intermediate purposes while the decision about generating additional outputs is being made. This resolves the contradiction by allowing some processing to proceed in parallel rather than waiting for complete outputs, thus maintaining productivity while still ensuring accuracy when needed.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system dynamically determines the processing sequence based on input characteristics. By using the designated portion of the neural network output to control whether additional outputs are generated and when processing should terminate, the system optimizes the balance between output completeness and processing throughput, allowing faster processing when complete outputs are not strictly necessary.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3238144B1Augmenting neural networks to generate additional outputs
Publication Date: 2021.04.14 DEEPMIND TECH LTD
  • EP3238144B1 patent drawingFigure 1
  • EP3238144B1 patent drawingFigure 2
  • EP3238144B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for augmenting neural networks to generate additional outputs. One of the systems includes a neural network and a sequence processing subsystem, wherein the sequence processing subsystem is configured to perform operations comprising, for each of the system inputs in a sequence of system inputs: receiving the system input; generating an initial neural network input from the system input; causing the neural network to process the initial neural network input to generate an initial neural network output for the system input; and determining, from a first portion of the initial neural network output for the system input, whether or not to cause the neural network to generate one or more additional neural network outputs for the system input.