SubTuning Selective Layer Finetuning for Multi-Task Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for training artificial intelligence models, particularly in autonomous vehicles, face challenges with resource-intensive data requirements and inefficient deployment of models for new tasks, especially in scenarios with scarce or corrupted data, where traditional finetuning methods are costly and suboptimal.
Innovation Solution
The proposed solution involves a parameter-efficient finetuning method called SubTuning, which selectively trains a carefully chosen subset of layers in a neural network while keeping the rest of the weights frozen, using a finetuning profile to identify the optimal layers for each task, thereby reducing computational costs and achieving performance comparable to full finetuning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional finetuning is used to train models for new tasks, then model performance on the new task is improved, but computational cost and resource requirements increase significantly
Solution Approach 1:
The patent segments the neural network into different layers and selectively finetunes only specific layers (e.g., later layers) rather than the entire network. This segmentation allows the model to adapt to new tasks while preserving computational efficiency by freezing earlier layers that capture general features.
Solution Approach 2:
The patent applies local quality by treating different layers of the network differently - some layers are finetuned while others are frozen. This local differentiation optimizes the balance between adapting to new tasks and maintaining computational efficiency, rather than applying uniform finetuning across all layers.
2Reliability
If multiple models are trained for different tasks in autonomous vehicles, then task performance is improved, but data storage and processing requirements increase tremendously
Solution Approach 1:
The patent enables a single pretrained model to perform multiple tasks through selective finetuning. By freezing most layers and only training specific layers for each new task, the model can be efficiently adapted to multiple functions without requiring separate full-model training for each task, reducing data storage and processing requirements.
3Adaptability or versatility
If selective layer finetuning is performed on small datasets, then adaptation to distribution shifts is improved, but the lack of strategic layer selection techniques limits effectiveness
Solution Approach 1:
The patent performs preliminary action by pre-identifying which layers should be finetuned based on the type of distribution shift or corruption expected. This strategic layer selection is determined before finetuning, allowing the system to efficiently adapt to small datasets with corrupted or shifted distributions without requiring complex runtime layer selection mechanisms.
Data Source
AI summary
Systems and methods generate an extended trained model. In one implementation, a method includes obtaining a preexisting trained model, the preexisting trained model including a plurality of preexisting weights, wherein each of the plurality of preexisting weights is associated with a preexisting value; identifying a subset of the plurality of preexisting weights; generating a plurality of extended weights based on a training process using duplicates of the subset of the plurality of preexisting weights; and generating the extended trained model, wherein the extended trained model includes the plurality of preexisting weights and the plurality of extended weights.