Compressed LSTM Layer Using Structured Matrices

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current neural network architectures, particularly recurrent neural networks (RNNs), face challenges in efficient data processing and storage due to their large parameter matrices, which hinders their performance and scalability.

Innovation Solution

The implementation of a compressed Long Short-Term Memory (LSTM) layer using Toeplitz-like structured matrices or re-defining gate parameter matrices with compressed and projection matrices, reducing memory requirements and accelerating training and processing.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional full-size LSTM layers are used in recurrent neural networks, then model accuracy and performance are maintained, but memory storage requirements and computational complexity increase significantly

Engineering Contradiction:
Improvemodel accuracyVSAvoidmemory storage requirements
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent applies parameter changes by modifying the gate parameter matrices from full-size dense matrices to structured matrices (Toeplitz, circulant, or low-rank decomposed forms). This structural transformation reduces the number of parameters while maintaining the functional capabilities of the LSTM gates, thereby reducing memory storage requirements while preserving model accuracy.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent segments the gate parameter matrices into structured components (e.g., decomposing into low-rank factors or separating Toeplitz structure parameters). This segmentation allows the parameter matrix to be represented more compactly through fewer independent parameters, reducing storage requirements while maintaining the essential transformation capabilities.

Inventive Principle:
Principle #1Segmentation

2Adaptability or versatility

If traditional full-size parameter matrices are used in LSTM layers, then comprehensive data transformation capability is maintained, but computational complexity and processing time increase

Engineering Contradiction:
Improvedata transformation capabilityVSAvoidcomputational complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent transforms the parameter matrices from dense full-size formats to structured formats (Toeplitz, circulant, or low-rank decomposed). This parameter change reduces computational complexity by exploiting the structured properties that allow for more efficient matrix-vector multiplications and reduced parameter updates during training, while maintaining data transformation capability.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent uses low-rank decomposition to represent full-size parameter matrices as products of smaller factor matrices. This copying approach creates compressed representations that can be computationally expanded when needed, reducing storage and computational complexity while preserving the essential transformation functionality through the factor matrices.

Inventive Principle:
Principle #26Copying

3Productivity

If compressed parameter matrices are used in LSTM layers, then memory efficiency and processing speed improve, but model performance and accuracy may deteriorate

Engineering Contradiction:
Improveprocessing speedVSAvoidmodel performance
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent employs parameter changes by adopting structured matrix formats (Toeplitz, circulant, or low-rank decomposed) that maintain sufficient representational capacity. These structured parameters are trained to preserve the essential functional relationships needed for accurate sequence processing, ensuring that model performance remains comparable to full-size networks while achieving compression rates of 75-83%.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3955168B1Compressed recurrent neural network models
Publication Date: 2024.07.10 GOOGLE LLC
  • EP3955168B1 patent drawingFigure 1
  • EP3955168B1 patent drawingFigure 2A
  • EP3955168B1 patent drawingFigure 2B

AI summary

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for implementing long-short term memory layers with compressed gating functions. One of the systems includes a first LSTM layer having gates that are configured to, for each of multiple time steps, generate a respective intermediate gate output vector by multiplying a gate input vector and a gate parameter matrix. The gate parameter matrix for at least one of the gates is a structured matrix or is defined by a compressed parameter matrix and a projection matrix. By including the compressed LSTM layer in the recurrent neural network, the recurrent neural network is configured to be able to process data more efficiently and use less data storage. A recurrent neural network having a compressed LSTM layer can be effectively trained to achieve word error rates comparable to full size, e.g. uncompressed, recurrent neural networks.