PRIME: Deep Imbalanced Regression with Proxies
Authors: Jongin Lim, Sucheol Lee, Daeho Um, Sung-Un Park, Jinwoo Shin
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the effectiveness and broad applicability of PRIME, achieving state-of-the-art performance on four real-world regression benchmark datasets across diverse target domains. |
| Researcher Affiliation | Collaboration | 1AI Center, Samsung Electronics 2Korea Advanced Institute of Science and Technology (KAIST). |
| Pseudocode | No | The paper describes the proposed method using mathematical formulations and descriptive text, but it does not contain any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | We will release the code after publication. |
| Open Datasets | Yes | We conduct experiments on four real-world imbalanced regression benchmarks introduced by (Yang et al., 2021): (i) Age DB-DIR is a facial age estimation dataset derived from Age DB (Moschoglou et al., 2017). (ii) IMDBWIKI-DIR is an age estimation dataset constructed from IMDB-WIKI (Rothe et al., 2018). (iii) NYUD2-DIR is derived from the NYU Depth Dataset V2 (Silberman et al., 2012) for depth prediction from RGB indoor scenes. (iv) STS-B-DIR is a natural language dataset based on STSB (Cer et al., 2017; Wang, 2018), providing continuous similarity scores between pairs of sentences. |
| Dataset Splits | Yes | For all datasets, we report results for four subsets: All, Many, Median, and Few. All refers to the entire test set. Based on the number of training samples per label, Many includes labels with over 100 samples, Median covers those with 20 to 100 samples, and Few consists of labels with fewer than 20 samples. [Table 10: Overall dataset statistics. Age DB-DIR: # Training 12,208, # Val. 2,140, # Test 2,140; IMDB-WIKI-DIR: # Training 191,509, # Val. 11,022, # Test 11,022; NYUD2-DIR: # Training 50,688, # Val. , # Test 654; STS-B-DIR: # Training 5,249, # Val. 1,000, # Test 1,000] |
| Hardware Specification | Yes | To analyze the computational efficiency of PRIME, we compute the average wall-clock training time (in seconds) using four NVIDIA Tesla V100 GPUs. |
| Software Dependencies | No | The paper mentions using ResNet50, Bi LSTM, and GloVe word embeddings, but does not specify version numbers for any software libraries or frameworks (e.g., Python, PyTorch, TensorFlow, CUDA). |
| Experiment Setup | Yes | The number of proxies C is empirically determined for each dataset. Proxy embeddings {zp i }C i=1 are initialized with He initialization (He et al., 2015) and trained jointly with the model. The Proxy lr refers to the multiplication factor applied to the learning rate of the proxy. The hyperparameters λp, λa, τf, τt, and α are set empirically. ... Tables 11, 12, 13, and 14 summarize the implementation details for Age DB-DIR, IMDB-WIKI-DIR, NYUD2-DIR, and STS-B-DIR, respectively. ... For Age DB-DIR and IMDB-WIKI-DIR, we use Res Net50 ... Epoch 80, Batch size 64, Learning rate 2.5e-4, Weight decay 1.0e-4, Optimizer Adam, Scheduler Step LR (60/0.1). |