Signal Processor Using Deterministic Neural Network for Generative Modeling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing signal processing systems, such as those employing neural networks and Boltzmann machines, face challenges in computational intensity and slow learning rates, particularly in generating output examples that match previously learned distributions, and require significant processing power.

Innovation Solution

The implementation of a deterministic neural network with a shared mapping system that translates between probability and category vectors, allowing for efficient data compression and facilitating faster learning by enabling closed-form weight calculation, along with a chain of signal processors that refine output examples based on context vectors.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If Boltzmann machines or deep generative models are used to generate output examples from learned distributions, then the system can provide generative capabilities, but the processing power required becomes excessively high

Engineering Contradiction:
Improvegenerative capabilityVSAvoidprocessing power
Core Design Contradiction:
Adaptability or versatilityVSPower

Solution Approach 1:

The patent segments the generative model into distinct functional components: a probability vector generation system that computes output probabilities, a category vector system that represents compressed categories, and a mapping system that connects them. This segmentation allows each component to be optimized independently, reducing overall computational requirements while maintaining generative capabilities.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes the parameter representation by introducing category vectors as compressed representations of probability vectors. Instead of working directly with full probability distributions requiring extensive computation, the system uses condensed category vectors that capture essential information with fewer parameters, thereby reducing processing power requirements.

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If Helmholtz machines are used to provide generative models, then the system can learn from training examples, but the learning rate becomes unacceptably slow

Engineering Contradiction:
Improvelearning capabilityVSAvoidlearning time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent replaces the slow, iterative learning mechanism of traditional Helmholtz machines with a more efficient deterministic neural network-based probability vector generation system. This substitution enables faster weight calculations and quicker convergence to learned distributions, dramatically reducing learning time while preserving the ability to learn from training examples.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

3Ease of operation

If traditional neural networks are used without a shared mapping system, then the system can process data, but the number of parameters to be learned becomes excessively large

Engineering Contradiction:
Improvedata processing capabilityVSAvoidnumber of parameters
Core Design Contradiction:
Ease of operationVSDevice complexity

Solution Approach 1:

The patent introduces a shared mapping system that serves multiple functions: it maps category vectors to probability vectors, enables compression of probability distributions into category representations, and facilitates efficient parameter sharing across different parts of the system. This universal mapping mechanism dramatically reduces the total number of parameters that need to be learned while maintaining full data processing capability.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP2973241B1Signal processing systems
Publication Date: 2020.10.21 DEEPMIND TECH LTD
  • EP2973241B1 patent drawingFigure 1a
  • EP2973241B1 patent drawingFigure 1b
  • EP2973241B1 patent drawingFigure 1c

AI summary

We describe a signal processor, the signal processor comprising: a probability vector generation system, wherein said probability vector generation system has an input to receive a category vector for a category of output example and an output to provide a probability vector for said category of output example, wherein said output example comprises a set of data points, and wherein said probability vector defines a probability of each of said set of data points for said category of output example; a memory storing a plurality of said category vectors, one for each of a plurality of said categories of output example; and a stochastic selector to select a said stored category of output example for presentation of the corresponding category vector to said probability vector generation system; wherein said signal processor is configured to output data for an output example corresponding to said selected stored category.