Dynamic Bit Precision Adaptation for Deep Neural Networks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Deep neural networks (DNNs) face challenges in training times, memory requirements, processor availability, battery power consumption, and energy efficiency due to their exponential growth in parameter sizes and over-parameterization, which traditional regularization techniques fail to effectively address.

Innovation Solution

The approach involves dynamically quantizing the range and set of values that parameters can take by using low-precision weights and bit precision values, optimizing bit precision as part of the training procedure, and employing a penalty term to control the number of unique values parameters can have, allowing for flexible resource constraint specification and optimal bit-precision per layer.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If deep neural networks use high-precision parameters, then algorithm performance is improved, but memory size and energy consumption increase

Engineering Contradiction:
Improvealgorithm performanceVSAvoidmemory size
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent applies parameter changes by dynamically adjusting the bit precision of network parameters during training. The system transitions from fixed high-precision representations to adaptive low-precision representations, changing the parameter precision from 32-bit floating point to variable bit representations (e.g., 8-bit, 4-bit, or even 1-bit) based on the importance and distribution of each parameter. This resolves the contradiction by maintaining sufficient performance while significantly reducing memory requirements through precision adaptation.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent segments the parameter space by applying different bit precision levels to different parameters or parameter groups. Instead of uniformly quantizing all parameters, the system identifies and segments parameters based on their sensitivity and importance, applying higher precision to critical parameters and lower precision to less sensitive ones. This selective segmentation maintains algorithm performance for important parameters while reducing overall memory consumption.

Inventive Principle:
Principle #1Segmentation

2Reliability

If deep neural networks use high-precision parameters, then algorithm performance is improved, but energy consumption increases

Engineering Contradiction:
Improvealgorithm performanceVSAvoidenergy consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent reduces energy consumption by dynamically changing parameter precision during training and inference. By transitioning from fixed high-precision to adaptive low-precision representations, the system minimizes the computational energy required for parameter storage, memory access, and arithmetic operations while maintaining acceptable performance levels through intelligent precision allocation.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent employs low-precision approximations as 'cheap' alternatives to high-precision parameters. By using simplified parameter representations (such as quantized or discretized values) that require less computational resources and energy to process, the system achieves a trade-off where marginal performance loss is acceptable in exchange for significant energy savings, especially in resource-constrained environments.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

3Adaptability or versatility

If deep neural networks are over-parameterized, then model capacity is improved, but training time increases

Engineering Contradiction:
Improvemodel capacityVSAvoidtraining time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent accelerates training by dynamically changing the precision and representation of parameters during the training process. The system adapts parameter bit depth and quantization levels based on training progress and parameter importance, allowing faster convergence and reduced computational overhead per iteration, thereby decreasing overall training time while maintaining model capacity.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces dynamics into the parameter representation by making bit precision adaptive rather than static. The system continuously adjusts parameter precision based on training state, parameter sensitivity analysis, and resource constraints, enabling the model to maintain high capacity when needed while operating more efficiently during training and inference, thus reducing total training time.

Inventive Principle:
Principle #15Dynamics

4Adaptability or versatility

If deep neural networks are over-parameterized, then model capacity is improved, but device complexity increases

Engineering Contradiction:
Improvemodel capacityVSAvoidprocessing complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent reduces processing complexity by changing parameter representations from complex high-precision floating-point formats to simpler low-precision formats (such as integers or discrete levels). This parameter transformation simplifies arithmetic operations, reduces memory bandwidth requirements, and enables deployment on devices with limited computational capabilities while preserving essential model capacity through adaptive precision management.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3543917B1Dynamic adaptation of deep neural networks
Publication Date: 2024.01.03 SRI INTERNATIONAL
  • EP3543917B1 patent drawingFigure 1
  • EP3543917B1 patent drawingFigure 2
  • EP3543917B1 patent drawingFigure 3

AI summary

Techniques are disclosed for training a deep neural network (DNN) for reduced computational resource requirements. A computing system includes a memory for storing a set of weights of the DNN. The DNN includes a plurality of layers. For each layer of the plurality of layers, the set of weights includes weights of the layer and a set of bit precision values includes a bit precision value of the layer. The weights of the layer are represented in the memory using values having bit precisions equal to the bit precision value of the layer. The weights of the layer are associated with inputs to neurons of the layer. Additionally, the computing system includes processing circuitry for executing a machine learning system configured to train the DNN. Training the DNN comprises optimizing the set of weights and the set of bit precision values.