Neural Network Training with Shared Encoder for Contrastive Vision Tasks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing training methods for neural network models, such as those described in NPL 1, are limited to classification tasks and are not applicable to other computer vision tasks like object detection, leading to inefficiencies in training and performance.
Innovation Solution
A training method for neural network models that includes a first network branch for representation learning using supervised contrastive learning and a second branch for training computer vision tasks, utilizing a shared encoder network model, data augmentation, and label processing to convert labels into applicable representations for tasks like object detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If supervised contrastive learning is used for representation learning, then classification accuracy is improved, but the method is limited to classification tasks and cannot be applied to other computer vision tasks like object detection
Solution Approach 1:
The neural network is divided into two separate branches: a first network branch dedicated to representation learning using supervised contrastive learning, and a second network branch for performing specific computer vision tasks. This segmentation allows each branch to be optimized for its specific function while maintaining overall system versatility across multiple task types.
Solution Approach 2:
The encoder network model serves as a shared component between both network branches, enabling it to provide universal feature representations for both representation learning and various computer vision tasks including classification, object detection, and segmentation. This multi-functionality resolves the contradiction by allowing a single trained model to serve multiple purposes.
2Reliability
If data augmentation is performed to increase training data, then model generalization is improved, but training time and computational resources increase
Solution Approach 1:
Data augmentation processing is performed in advance during the training phase to create an expanded training dataset. By pre-augmenting the data before training the encoder network model, the system achieves better model generalization without incurring additional computational costs during the actual training execution, as the augmented data is ready for immediate use.
3Productivity
If a shared encoder network model is used for both representation learning and task-specific training, then training efficiency is improved, but the complexity of coordinating multiple training objectives increases
Solution Approach 1:
The training process dynamically balances multiple objectives by simultaneously training the shared encoder network model with both representation learning losses and task-specific losses. The system adaptively adjusts to handle the complexity of coordinating multiple training objectives while maintaining training efficiency through the unified architecture.
Solution Approach 2:
The first network branch acts as an intermediary between the shared encoder network model and the second network branch, processing representations through supervised contrastive learning before they are used by task-specific models. This intermediary structure manages the complexity of multiple training objectives by providing a dedicated processing stage for representation learning.
Data Source
AI summary
A training method according to the present disclosure is a training method performed by a computer to train a neural network model including a first network branch for representation learning with use of supervised contrastive learning, and a second network branch for training of computer vision tasks including a classification task and a task other than the classification task. The training method includes: performing label processing for converting labels of M image data items into labels applicable to the representation learning, as labels of the computer vision tasks; and training an encoder network model and a first model with use of a first loss function for use in the supervised contrastive learning, the labels of the M image data items on which the label processing has been performed, and embedding vectors of the M image data items.