Storage System Feature Extraction Layer for AI Workflows
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional storage systems are limited in their ability to participate in AI workflows as they are only used for storing 'data bytes' without a high-level understanding of data, hindering their development and requiring extensive computing resources for training deep neural network (DNN) models for AI applications.
Innovation Solution
Implementing a lightweight AI solution on a storage system by generating a feature extraction layer from a pre-trained DNN model, which is deployed within the storage system to extract features from images, and training an output layer to create an image classification model, thereby reducing the computational burden and enabling faster model generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional storage systems are used to store only data bytes, then storage function is simple and reliable, but storage systems cannot participate in AI workflows and require extensive computing resources for training DNN models
Solution Approach 1:
The patent merges storage systems with AI processing capabilities by integrating a feature extraction layer into the storage system. This allows the storage system to not only store data but also perform feature extraction for AI workflows, thereby participating in AI applications without requiring separate complex processing systems.
Solution Approach 2:
The patent segments the DNN model into two parts: a pre-trained feature extraction layer that is integrated into the storage system, and a task-specific output layer that is trained separately. This segmentation allows the storage system to provide general AI capabilities while maintaining simplicity, as only the necessary feature extraction functionality is embedded rather than the entire complex model.
2Productivity
If pre-trained DNN models are used for feature extraction, then model training speed is improved and computing resources are reduced, but the storage system must integrate and manage complex model components
Solution Approach 1:
The patent applies preliminary action by using a pre-trained feature extraction layer that has already been trained on general image data. This pre-trained layer is then integrated into the storage system and used for extracting features from images stored in the storage system, eliminating the need to train the entire DNN model from scratch and significantly reducing training time and computational resources.
Solution Approach 2:
The patent extracts the feature extraction functionality from the complete DNN model and integrates it into the storage system. By taking out only the necessary feature extraction layer and separating it from the task-specific output layer, the system achieves fast feature extraction without managing the complexity of training the entire model, as only the pre-trained portion needs to be deployed.
3Reliability
If extensive computing resources are allocated for training DNN models, then model accuracy is improved, but storage systems cannot facilitate quick model generation and AI application development
Solution Approach 1:
The patent segments the DNN model into a pre-trained feature extraction layer and a task-specific output layer. The feature extraction layer, which provides the foundation for accurate feature representation, is pre-trained and integrated into the storage system. The output layer, which handles task-specific classification, is trained quickly using only the extracted features and small training sets, thereby achieving both accuracy and fast model generation.
Solution Approach 2:
The patent applies preliminary action by pre-training the feature extraction layer on comprehensive datasets before integrating it into the storage system. This preliminary training ensures that the feature extraction capabilities are already optimized for accuracy. When new tasks are performed, only the output layer needs to be trained, which can be done quickly with small datasets, thus reducing model generation time while maintaining high accuracy.
Data Source
AI summary
Image processing is described. An example method includes generating a feature extraction layer portion of an image classification model based on a deep neural network (DNN) model, and extracting features of a group of images by using the feature extraction layer portion. The method further includes training an output layer portion of the image classification model according to the features of training images in the group of images and classification labels of the training images. The method further includes generating the image classification model by combining the feature extraction layer portion and the output layer portion. Embodiments of the present disclosure implement a lightweight artificial intelligence (AI) solution on a storage system. An expanded storage system can facilitate the generation of AI applications and assist in training an image classification model more quickly, and the obtained image classification model can also produce high accuracy on a small training set.