ACCon: Angle-Compensated Contrastive Regularizer for Deep Regression
Authors: Botao Zhao, Xiaoyang Qu, Zuheng Kang, Junqing Peng, Jing Xiao, Jianzong Wang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments and theoretical analysis demonstrate that our proposed angle-compensated contrastive regularizer not only achieves competitive regression performance but also excels in data efficiency and effectiveness on imbalanced datasets. We perform comprehensive experiments across three datasets from both computer vision and natural language processing fields, showcasing enhanced performance relative to current state-of-the-art techniques. |
| Researcher Affiliation | Industry | Ping An Technology (Shenzhen) Co., Ltd. EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1: Training of the deep regression with ACCon. |
| Open Source Code | No | The paper does not provide any explicit statement about releasing source code, a link to a code repository, or mention of code in supplementary materials. |
| Open Datasets | Yes | Datasets: To rigorously evaluate our method, we selected three diverse datasets: (1) Age DB (Moschoglou et al. 2017): An age estimation dataset comprising 16,488 facial images; (2) IMDB-WIKI (Rothe, Timofte, and Van Gool 2018): A large-scale facial age dataset containing 523,000 images with corresponding age labels; (3) STS-B (Cer et al. 2017; Wang et al. 2018): A natural language dataset consisting of 7,249 sentence pairs, extracted from the Semantic Textual Similarity (STS) Benchmark. |
| Dataset Splits | Yes | To ensure a comprehensive assessment of our proposed method, we employed 2 distinct sampling strategies for partitioning the datasets into training, validation, and test sets: (1) Balanced Sampling: Following the benchmark established by (Yang et al. 2021), we partitioned the Age DB dataset into 12,208 training samples, 2,140 validation samples, and 2,140 test samples. For IMDB-WIKI, we allocated 191,500 images for training and 11,000 images each for validation and testing. From STS-B, we sampled 1,000 pairs each for validation and testing. This benchmark ensures a balanced label distribution in the validation and test sets, as illustrated in Appendix Figure 1. Similar to Yang et al. s benchmark, we denote these balanced datasets as Age DB-DIR, IMDB-WIKI-DIR, and STS-B-DIR. (2) Natural Sampling: We used the same dataset splitting ratio but randomized the division of training, validation, and test datasets, thereby preserving similar label distributions across all three sets (Appendix Figure 1). |
| Hardware Specification | No | The paper mentions 'Res Net-50 was utilized as the backbone' and discusses experimental results but does not provide specific details about the hardware (e.g., GPU model, CPU type, memory) used for these experiments in the main text. It states 'Detailed experimental setup, including training protocols, dataset specifications, and hyperparameters, are provided in Appendix B.', implying hardware details are not in the main body. |
| Software Dependencies | No | The paper mentions 'Res Net-50 was utilized as the backbone' and 'Bi LSTM+Glo Ve word embeddings were employed', which are models and embeddings. However, it does not specify any software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions) in the main text. It states 'Detailed experimental setup, including training protocols, dataset specifications, and hyperparameters, are provided in Appendix B.', implying software details are not in the main body. |
| Experiment Setup | No | The paper mentions several hyperparameters by name in Algorithm 1, such as 'The temperature index, τ; Weight coefficient, γ; Max training epoch, T; Smoothing term, ϵ.' and discusses the loss function L = Lreg + γ LACCon. However, it explicitly states 'Detailed experimental setup, including training protocols, dataset specifications, and hyperparameters, are provided in Appendix B.' This indicates that the specific concrete numerical values for these hyperparameters or other system-level training settings are not present in the main text. |