Contrastive Learning Model Two-Phase Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The performance of contrastive learning models varies significantly based on the order of training samples, leading to inconsistent training results due to conflicting roles of intra-video negative samples in learning appearance and dynamics.
Innovation Solution
A two-phase training approach is introduced, where the first phase focuses on learning appearance using negative samples from different data sequences, and the second phase learns dynamics using negative samples from the same data sequence, optimizing the selection of negative sample pairs for each objective.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If negative samples from the same data sequence are used during training, then the model can learn dynamics, but the learning of appearance and dynamics conflicts leading to inconsistent performance
Solution Approach 1:
The patent segments the training process into two distinct phases: a first training phase that uses only negative samples from different data sequences to learn appearance, and a second training phase that uses negative samples from the same data sequence to learn dynamics. This segmentation resolves the conflict between appearance and dynamics learning by separating them into different training stages, thereby improving training performance consistency while maintaining adaptability.
2Adaptability or versatility
If training samples are used in different batch orders, then various training configurations can be explored, but the model performance varies significantly
Solution Approach 1:
The patent applies preliminary action by establishing a fixed, predetermined batch order for training samples before the training process begins. This preliminary arrangement ensures that the model is trained in a consistent sequence across different runs, eliminating performance variations caused by different batch orders while still allowing flexibility in choosing between the two-phase training configurations.
3Loss of information
If intra-video negative samples are used, then dynamics information can be captured, but the conflicting roles lead to degraded appearance learning
Solution Approach 1:
The patent segments the learning objectives by dedicating the first training phase exclusively to appearance learning using negative samples from different data sequences, and the second training phase to dynamics learning using negative samples from the same data sequence. This segmentation ensures that appearance learning accuracy is not degraded by the presence of intra-video negative samples, while dynamics information is still captured in the second phase.
Solution Approach 2:
The patent applies preliminary action by first completing appearance learning with inter-video negative samples before introducing intra-video negative samples for dynamics learning. This preliminary establishment of appearance features ensures that when dynamics learning begins in the second phase, the appearance representation is already solidified and not subject to degradation.
Data Source
AI summary
A method, apparatus, device, and medium for managing a contrastive learning model are provided. In one method, in a first training phase, a first contrastive learning model is generated by training the contrastive learning model with a first training sample set, a negative sample pair of the first training sample set comprising only data segments from different data sequences. In a second training phase, a second contrastive learning model is generated by training the first contrastive learning model with a second training sample set, a negative sample pair in the second training sample set comprising data segments from the same data sequence. Knowledge in terms of the appearance of the samples can be fully obtained in the first training phase, and knowledge in terms of the appearance and dynamics of the samples can be fully obtained in the second training, with example implementations of the present disclosure.