Generative Neural Networks Using Hidden Canvas for Image Reconstruction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current neural network architectures face challenges in efficiently generating reconstructions of input data and performing one-shot generalization, requiring significant time and resources due to limitations in attention and canvas writing mechanisms.

Innovation Solution

The proposed neural network system incorporates a generative subsystem with a recurrent neural network and an inference subsystem, utilizing an attention-based writing mechanism and spatial transformers to iteratively update a hidden canvas across multiple time steps, enabling enhanced generative and inferential outcomes by processing input data sequentially and sampling latent variables for reconstruction and generalization.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If current neural network architectures use traditional attention and canvas writing mechanisms for generating reconstructions, then generative capability is achieved, but time and computational resources required are excessive

Engineering Contradiction:
Improvegeneration speedVSAvoidtime required for generation
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent divides the image generation process into discrete time steps where a recurrent neural network iteratively updates a hidden canvas. Each time step processes a portion of the reconstruction task, allowing parallel computation and reducing overall generation time while maintaining quality.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a dynamic hidden canvas that evolves over multiple time steps through recurrent processing. The canvas is continuously updated based on latent variables and attention mechanisms, enabling efficient iterative refinement of the generated reconstruction without requiring excessive computational resources at any single step.

Inventive Principle:
Principle #15Dynamics

2Adaptability or versatility

If traditional neural network architectures are used for one-shot generalization, then model capability is maintained, but computational resources required are significant

Engineering Contradiction:
Improveone-shot generalization capabilityVSAvoidcomputational resources
Core Design Contradiction:
Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

The patent pre-processes input data into latent variables that capture essential features before the generation phase. This preliminary encoding enables the recurrent network to perform one-shot generalization more efficiently by working with compressed representations rather than raw data, reducing computational resource requirements.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces a hidden canvas as an intermediary representation between the recurrent neural network and the final output. This canvas serves as a dynamic buffer that stores and refines generative information over time, enabling efficient one-shot generalization by mediating between latent variables and reconstructed output.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Manufacturing precision

If iterative canvas updating is performed across multiple time steps, then reconstruction quality is improved, but computational complexity increases

Engineering Contradiction:
Improvereconstruction qualityVSAvoidsystem complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent designs a universal recurrent neural network architecture that performs multiple functions: encoding input data, generating latent variables, updating the hidden canvas, and producing reconstructions. This multi-functional design reduces overall system complexity despite the iterative multi-step process by using a single versatile network component.

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

Solution Approach 2:

The patent maintains continuous useful action through the recurrent updating of the hidden canvas across time steps. Each iteration refines the reconstruction quality while the network maintains its internal state, allowing the system to build upon previous computations rather than restarting, thus improving quality without proportionally increasing complexity.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP3398119B1Generative neural networks for generating images using a hidden canvas
Publication Date: 2022.06.22 DEEPMIND TECH LTD
  • EP3398119B1 patent drawingFigure 1
  • EP3398119B1 patent drawingFigure 2
  • EP3398119B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for a neural network system. In one aspect, a neural network system includes a recurrent neural network that is configured to, for each time step of a predetermined number of time steps, receive a set of latent variables for the time step and process the latent variables to update a hidden state of the recurrent neural network; and a generative subsystem that is configured to, for each time step, generate the set of latent variables for the time step and provide the set of latent variables as input to the recurrent neural network; update a hidden canvas using the updated hidden state of the recurrent neural network; and, for a last time step, generate output data item using the updated hidden canvas for the last time step.