Sequence Generation Retrieval Using Similar Reference Segments

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing machine learning models struggle to generate accurate output sequences by only processing current input sequences, lacking sufficient context and requiring significant computational resources.

Innovation Solution

A system that generates a final output sequence conditioned on both current output segments and k most similar reference segments, using an encoder and decoder neural network with cross-attention mechanisms to leverage greater context while reducing computational demands.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If the model processes only the current input sequence without reference segments, then the computational resources required are reduced, but the accuracy and descriptive detail of the generated output sequence deteriorate

Engineering Contradiction:

Improveoutput generation accuracyVSAvoidcomputational resource consumption

Core Design Contradiction:

ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent pre-processes and stores reference segments from the training data before the actual generation process. These reference segments are encoded and stored in a database, allowing the model to quickly retrieve relevant context during generation without performing full sequence processing each time, thus improving accuracy while controlling computational costs

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent introduces reference segments as an intermediary between the input sequence and the generation process. These reference segments serve as additional contextual information that mediates the relationship between input and output, providing richer context without requiring the model to process the entire training dataset during generation

Inventive Principle:

Principle #24Intermediary (Mediator)

2Loss of information

If the model accesses greater context through reference segments, then the descriptive detail and accuracy of the output sequence is improved, but the computational resources required increase

Engineering Contradiction:

Improvecontext information availabilityVSAvoidsystem computational complexity

Core Design Contradiction:

Loss of informationVSDevice complexity

Solution Approach 1:

The patent extracts only the necessary contextual information from the training data in the form of reference segments. Instead of using the entire training dataset during generation, the model extracts and stores pre-computed embeddings of relevant sequences, allowing efficient access to context without the complexity of processing large datasets during inference

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent segments the training data into discrete reference segments that can be independently stored and retrieved. Each reference segment represents a discrete unit of contextual information that can be selectively accessed during generation, reducing the complexity of handling large continuous datasets

Inventive Principle:

Principle #1Segmentation

3Manufacturing precision

If the system uses k most similar reference segments for each current output segment, then the quality of generated sequences is improved, but the processing time and computational load increase

Engineering Contradiction:

Improvesequence generation qualityVSAvoidprocessing time

Core Design Contradiction:

Manufacturing precisionVSLoss of time

Solution Approach 1:

The patent performs similarity computations and retrieves the k most similar reference segments in advance during the encoding phase. By pre-computing these similarities and storing them alongside the reference segment embeddings, the system avoids performing time-consuming similarity searches during the actual generation process, thus maintaining high quality while reducing processing time

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent dynamically adjusts the retrieval process by computing similarities between current output segments and reference segment embeddings on-the-fly during generation. This dynamic approach allows the system to adaptively select the most relevant reference segments for each generation step, optimizing the balance between quality and processing efficiency

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS12536439B2Large scale retrieval for sequence generation

Publication Date: 2026.01.27 GDM HOLDING LLC

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating a final output sequence. In one aspect, a method comprises: receiving a current output sequence comprising one or more current output segments; receiving a set of reference segments and a respective reference segment embedding of each reference segment that has been generated using an embedding neural network; for each current output segment: processing the current output segment using the embedding neural network to generate a current output segment embedding of the current output segment; and selecting k most similar reference segments to the current output segment using the reference segment embeddings and the current output segment embedding; and processing the current output sequence and the k most similar reference segments for each current output segment to generate an additional output segment that follows the current output sequence in the final output sequence.