Stochastic Bottleneck Neural Network for Adaptable Dimensionality Reduction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural networks face challenges in achieving flexible regularization and adaptable dimensionality reduction, as they often require cumbersome re-training and lack the ability to seamlessly adjust latent dimensionality based on varying distortion requirements, and determining the optimal size of neural networks is difficult due to the curse of dimensionality and overfitting issues.

Innovation Solution

The proposed method employs a stochastic bottleneck architecture with non-uniform dropouts across hidden layers to prioritize dominant features, allowing for flexible adjustment of neural network depth and width without re-training, using techniques like tail dropping and multi-dimensional dropout rates to achieve adaptable dimensionality reduction and reduce computational complexity.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If traditional AE with deterministic bottleneck is used, then dimensionality reduction is achieved, but adaptability to adjust latent dimensionality without re-training is lost

Engineering Contradiction:
Improveadaptability to adjust latent dimensionalityVSAvoidmodel re-training requirement
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent applies dynamics by replacing the static deterministic bottleneck with a dynamic stochastic bottleneck that uses dropout mechanisms. This allows the effective dimensionality of the latent space to be dynamically adjusted during inference without re-training, as the dropout rates can be modified to change the number of active latent variables while maintaining the same trained model architecture.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes the parameter of dropout rate to control the effective dimensionality of the latent space. By varying the dropout rate parameter during inference, the system can adaptively adjust the number of active latent variables from the trained model without requiring model re-training, thus achieving flexible dimensionality reduction.

Inventive Principle:
Principle #35Parameter changes

2Manufacturing precision

If uniform dropout is used across all layers, then regularization is applied, but inability to prioritize dominant features is achieved

Engineering Contradiction:
Improvefeature prioritization accuracyVSAvoidnon-uniform dropout configuration
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent applies local quality by implementing non-uniform dropout rates that vary across different layers and positions in the neural network. Specifically, higher dropout rates are applied to later layers and to less important latent variables, while lower dropout rates are applied to earlier layers and more dominant features. This localized differentiation enables the system to prioritize dominant features during regularization while maintaining the ability to capture essential data characteristics.

Inventive Principle:
Principle #3Local quality

3Reliability

If deep and wide neural networks are used, then modeling capacity is improved, but requirement for large amount of training data increases

Engineering Contradiction:
Improvemodeling capacityVSAvoidtraining data requirement
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent applies partial action by using an over-complete neural network architecture that is intentionally designed with more latent variables than the final effective dimensionality. During training, the network uses this excessive capacity to learn robust representations, but during inference, dropout mechanisms activate to reduce the effective number of active latent variables to the required dimensionality. This allows the model to achieve high modeling capacity during training while requiring less data than would be needed for a network of equivalent final size.

Inventive Principle:
Principle #16Partial or excessive action

4Adaptability or versatility

If deterministic bottleneck architecture is used, then dimensionality reduction is fixed, but flexibility for different distortion requirements is lost

Engineering Contradiction:
Improveflexibility for distortion requirementsVSAvoiddimensionality adjustment operation
Core Design Contradiction:
Adaptability or versatilityVSEase of operation

Solution Approach 1:

The patent transforms the static deterministic bottleneck into a dynamic stochastic system where the effective dimensionality can be adjusted in real-time based on distortion requirements. By modifying dropout rates during inference, the system can adaptively change the number of active latent variables to match different application requirements without requiring multiple fixed models, thus achieving both flexibility and ease of operation.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3973463B1Non-uniform regularization in artificial neural networks for adaptable scaling
Publication Date: 2024.01.03 MITSUBISHI ELECTRIC CORP
  • EP3973463B1 patent drawingFigure 1A~1B
  • EP3973463B1 patent drawingFigure 1C
  • EP3973463B1 patent drawingFigure 1D

AI summary

A system for flexible regularization and adaptable scaling of an artificial neural network is provided. The system includes a memory to store an artificial neural network and training data, a processor and interface to submit signals and training data into the neural network having a sequence of layers, each layer includes a set of neuron nodes, wherein a pair of nodes from neighboring layers are mutually connected with a plural of trainable parameters to pass the signals from the previous layer to next layer, a random number generator to modify the output signal of each neuron nodes for regularization in a stochastic manner following a multi-dimensional distribution across layer depth and node width directions of the neural network, wherein at least one layer has non-identical profile across neuron nodes, a training operator to update the neural network parameters by using the training data such that the output of neural network provides better values in a plural of objective functions; and an adaptive truncator to prune the output of neuron nodes at each layer in a compressed size of the neural network to reduce the computational complexity on the fly in downstream testing phase for any new incoming data.