Domain-Aware Fine-Tuning: Enhancing Neural Network Adaptability

Authors: Seokhyeon Ha, Sunbeom Jeong, Jungwoo Lee

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our method outperforms other baseline methods, demonstrating its effectiveness in not only improving performance but also mitigating feature distortion.
Researcher Affiliation Collaboration Seokhyeon Ha1, Sunbeom Jeong1, Jungwoo Lee1,2 1 Seoul National University 2 Hodoo AI Lab {aoxm1231, sb3991, junglee}@snu.ac.kr
Pseudocode Yes Algorithm 1: Batch Normalization Conversion
Open Source Code No The paper does not contain an explicit statement about releasing the source code for the methodology or a link to a code repository.
Open Datasets Yes CIFAR-10 (Krizhevsky et al. 2009): A dataset that contains 10 categories of objects. For the OOD dataset, we use two additional datasets, CIFAR-10.1 (Recht et al. 2018) and STL (Coates, Ng, and Lee 2011).
Dataset Splits Yes Consistent with LP-FT (Kumar et al. 2022), we use ℓ2-regularized logistic regression classifier for linear probing and also choose the best ℓ2-regularization hyperparameter based on ID validation accuracy. For fine-tuning, we employ an SGD classifier with a cosine learning rate schedule and a batch size of 64. We fine-tune for 20 epochs on CIFAR-10, Entity-30, and Living-17, but extend the fine-tuning to 50 epochs on Domain Net and f Mo W due to their limited number of images. Early stopping is applied based on the ID validation accuracy.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts) used for running its experiments, only general statements about computational resources.
Software Dependencies No The paper mentions software components like 'SGD classifier' and 'Deep Labv3+ framework' but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes We fine-tune for 20 epochs on CIFAR-10, Entity-30, and Living-17, but extend the fine-tuning to 50 epochs on Domain Net and f Mo W due to their limited number of images. For fine-tuning, we employ an SGD classifier with a cosine learning rate schedule and a batch size of 64. [...] For the segmentation task, we adopt the Deep Labv3+ (Chen et al. 2018) framework and initialize its backbone with a pre-trained Res Net-50 model from Mo Cov2. During fine-tuning, we use the SGD optimizer with a batch size of 16 and a polynomial learning rate schedule with power 0.9. We conduct the fine-tuning process for a total of 30,000 iterations, and the final model is evaluated thereafter.