Activation Compression With Non-Uniform Mantissas for Neural Network Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning models, particularly neural networks, are computationally expensive and inefficient for real-time processing due to their reliance on general-purpose processors, necessitating improvements in hardware and software to enhance computational efficiency.

Innovation Solution

Implementing neural networks with activation values stored in a compressed format using block floating-point representation with non-uniform mantissas, allowing for efficient storage and retrieval during forward and backward propagation, utilizing general-purpose processors, neural network accelerators, or reconfigurable logic devices like FPGAs.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If activation values are stored in full precision format during forward propagation, then computational accuracy is maintained, but memory overhead and storage requirements increase

Engineering Contradiction:
Improvecomputational accuracyVSAvoidmemory overhead
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent changes the precision parameter of activation values from full precision to compressed precision formats. During forward propagation, activation values are stored in a compressed format with reduced precision, and during backward propagation, they are restored to higher precision. This parameter transformation resolves the contradiction by allowing low memory overhead during forward pass while maintaining high computational accuracy during backward pass.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent implements dynamic precision switching for activation values based on the operational phase. The system dynamically adjusts the precision level of stored activation values - using compressed low-precision format during forward propagation and restored high-precision format during backward propagation. This dynamic adaptation resolves the contradiction between memory efficiency and computational accuracy by optimizing precision requirements for each phase.

Inventive Principle:
Principle #15Dynamics

2Ease of operation

If neural networks are executed on general-purpose processors, then flexibility and ease of implementation are maintained, but computational speed and energy efficiency deteriorate

Engineering Contradiction:
Improveease of implementationVSAvoidcomputational speed
Core Design Contradiction:
Ease of operationVSProductivity

Solution Approach 1:

The patent changes the data format parameter of activation values from standard floating-point to compressed block floating-point formats. This parameter transformation enables general-purpose processors to achieve higher computational efficiency by reducing the number of bits processed and stored, while maintaining compatibility with standard processor architectures. The compression reduces memory bandwidth requirements and improves cache utilization, thereby increasing computational speed without requiring specialized hardware.

Inventive Principle:
Principle #35Parameter changes

3Productivity

If activation values are compressed during storage, then memory bandwidth and storage efficiency improve, but retrieval and decompression complexity increase

Engineering Contradiction:
Improvememory bandwidth efficiencyVSAvoidcompression complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent changes the numerical representation parameter of activation values to block floating-point format with shared exponents. This parameter transformation enables efficient compression by grouping multiple activation values that share common exponent bits, thereby reducing storage requirements and improving memory bandwidth efficiency. The block floating-point format maintains relatively simple compression and decompression operations compared to other compression schemes.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent merges multiple activation values into block floating-point representations where common exponent bits are shared across multiple values. By combining the exponent information for groups of activation values, the system reduces redundant data storage and improves memory efficiency. This merging approach simplifies the compression process by identifying and consolidating common patterns in the exponent bits of adjacent activation values.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3915056B1Neural network activation compression with non-uniform mantissas
Publication Date: 2025.07.16 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP3915056B1 patent drawingFigure 1
  • EP3915056B1 patent drawingFigure 2
  • EP3915056B1 patent drawingFigure 3

AI summary

Apparatus and methods for training a neural network accelerator using quantized precision data formats are disclosed, and in particular for storing activation values from a neural network in a compressed format having lossy or non-uniform mantissas for use during forward and backward propagation training of the neural network. In certain examples of the disclosed technology, a computing system includes processors, memory, and a compressor in communication with the memory. The computing system is configured to perform forward propagation for a layer of a neural network to produced first activation values in a first block floating-point format. In some examples, activation values generated by forward propagation are converted by the compressor to a second block floating-point format having a non-uniform and/or lossy mantissa. The compressed activation values are stored in the memory, where they can be retrieved for use during back propagation.