What Makes Partial-Label Learning Algorithms Effective?

Authors: Jiaqi Lv, Yangfan Liu, Shiyu Xia, Ning Xu, Miao Xu, Gang Niu, Min-Ling Zhang, Masashi Sugiyama, Xin Geng

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments on FMNIST [41], CIFAR-100 [21] and mini-Image Net [32]. and Our findings reveal that high accuracy on benchmark-simulated datasets with PLs...
Researcher Affiliation Academia 1Southeast University 2The University of Queensland 3RIKEN Center for Advanced Intelligence Project 4The University of Tokyo 5Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China
Pseudocode No Definition 2.1 (Mini-batch PL purification). Mini-batch PL purification is a process where for each mini-batch B D selected at iteration t, the weights are updated such that the distinction among candidate labels contributions increases over iterations: wt+1(x; f, S) = g(model s confidence for x based on current and previous iterations), (2) with g being a strictly increasing function that increases the weight for more likely candidate labels according to the model s confidence. The model s parameters θt are updated by optimizing a weighted loss over B: θt+1 = θt ηt θ X(x,S) B ℓ(f(x; θt), S; wt+1(x)). (3)
Open Source Code No Answer: [NA] Justification: The paper focuses on understand existing algorithms rather than to fundamentally improve them.
Open Datasets Yes As benchmarking on partially labeled vision datasets has become standard practice in evaluating deep PLL methods, we conduct experiments on FMNIST [41], CIFAR-100 [21] and mini-Image Net [32].
Dataset Splits Yes We left out 10% of the corrupted training samples as a validation set, and searched the initial learning rate from {0.1, 0.07, 0.05, 0.03} with cosine learning rate scheduling.
Hardware Specification Yes The implementation was based on Py Torch [26] and experiments were carried out with Ge Force RTX 4090 D.
Software Dependencies No The implementation was based on Py Torch [26]
Experiment Setup Yes All the methods were trained for 500 epochs with a standard SGD optimizer [9] with a momentum of 0.9 and the batch size was 256 (128 for mini-Image Net). We left out 10% of the corrupted training samples as a validation set, and searched the initial learning rate from {0.1, 0.07, 0.05, 0.03} with cosine learning rate scheduling.