Multi-Task Transformer Recovery of Lost Genomic Compression Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data compression techniques, particularly lossy compression, result in data loss that is challenging to recover, especially in correlated datasets like genomic and biological sequence data, where information dependencies are not fully leveraged for recovery.
Innovation Solution
A neural network system that integrates AI-based techniques for upsampling decompressed genomic data after lossy compression, utilizing a novel AI deblocking network with recurrent layers for feature extraction and a channel-wise transformer with attention to capture complex inter-channel dependencies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If lossy compression is applied to genomic data, then compression ratio is improved, but information loss increases
Solution Approach 1:
The patent introduces an intermediary neural network system that acts as a mediator between the compressed data and the original data. The neural network learns the mapping from compressed to original data, effectively recovering lost information without requiring the original uncompressed data to be transmitted or stored. This intermediary model enables information recovery that would otherwise be impossible with traditional lossy compression.
Solution Approach 2:
The patent changes the parameters of the data representation by transforming genomic data into a compressed latent space representation, then using neural networks to learn the inverse transformation. By changing how data is represented and processed (from direct compression to learned compression-decompression pairs), the system achieves both high compression ratios and effective information recovery.
2Device complexity
If traditional compression algorithms are used for genomic data, then processing simplicity is maintained, but ability to leverage cross-dataset dependencies is reduced
Solution Approach 1:
The patent replaces traditional mechanical compression algorithms with a neural network-based system. Instead of using fixed mathematical transformations, the system employs learned models that can adapt to the specific characteristics of genomic data and leverage cross-dataset dependencies. This substitution enables more sophisticated information recovery while maintaining practical usability through automated training and inference.
3Device complexity
If deep learning approaches treat each dataset independently, then model training is simplified, but ability to capture shared patterns across datasets is reduced
Solution Approach 1:
The patent merges multiple independent dataset processing streams into a unified neural network architecture. The model processes multiple correlated datasets simultaneously, allowing it to learn and exploit shared patterns and dependencies across datasets. This combining approach enables the system to recover information that would be invisible to independent processing of each dataset.
Solution Approach 2:
The patent creates a universal neural network model that can handle multiple different genomic datasets through a single unified architecture. The model learns generalizable patterns that apply across different datasets while maintaining the ability to capture dataset-specific characteristics. This multi-functional approach improves information recovery by leveraging correlations across diverse data sources.
Data Source
AI summary
A system and methods for upsampling of decompressed genomic data after lossy compression using a neural network integrates AI-based techniques to enhance compression quality. It incorporates a novel deep-learning neural network that upsamples decompressed data to restore information lost during lossy compression, taking advantage of cross-correlations between genomic data sets.