Conditioned RNN Output Generation for Real-Time N-Bit Sampling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural network systems face challenges in generating accurate output examples efficiently, particularly in environments with limited computational resources such as mobile devices.

Innovation Solution

The system generates output examples using a recurrent neural network by splitting the output value into two halves, generating the values of the first half and then the second half conditioned on the first half, reducing the number of sequential matrix-vector product computations and optimizing for real-time implementation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If the system generates N-bit output values using a recurrent neural network in real-time on resource-constrained devices, then the computational and memory requirements become prohibitively high, but reducing the computational approach compromises output generation accuracy

Engineering Contradiction:
Improvereal-time output generation speedVSAvoidcomputational and memory resources
Core Design Contradiction:
ProductivityVSQuantity of substance

Solution Approach 1:

The patent divides the N-bit output value generation into two separate stages: first generating the most significant N/2 bits, then generating the least significant N/2 bits. This segmentation reduces the output space size from 2^N to 2^(N/2), thereby reducing the number of sequential matrix-vector product computations required at each time step while maintaining the ability to generate accurate N-bit output values through conditioned generation

Inventive Principle:
Principle #1Segmentation

2Productivity

If the system reduces the number of sequential matrix-vector product computations by splitting output generation into two halves, then computational efficiency improves, but the complexity of the generation process increases

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidgeneration process complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent performs preliminary generation of the most significant N/2 bits first, using these generated values as conditioning input for the subsequent generation of the least significant N/2 bits. This preliminary action structure allows the system to break down the complex N-bit generation task into two manageable stages, where each stage operates with reduced computational complexity while the conditioning mechanism ensures overall generation accuracy

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3698292B1Generating output examples using recurrent neural networks conditioned on bit values
Publication Date: 2025.04.02 GDM HOLDING LLC
  • EP3698292B1 patent drawingFigure 1
  • EP3698292B1 patent drawingFigure 2
  • EP3698292B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating output examples using neural networks. One of the methods includes, at each generation time step, processing a first recurrent input comprising an N-bit output value at the preceding generation time step in the sequence using a recurrent neural network and in accordance with a hidden state to generate a first score distribution; selecting, using the first score distribution, values for the first half of the N bits; processing a second recurrent input comprising (i) the N-bit output value at the preceding generation time step and (ii) the values for the first half of the N bits using the recurrent neural network and in accordance with the same hidden state to generate a second score distribution; and selecting, using the second score distribution, values for the second half of the N bits of the output value.