Self-Paced Learning: An Implicit Regularization Perspective
Authors: Yanbo Fan, Ran He, Jian Liang, Baogang Hu
AAAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we implement SPL-IR to both supervised and unsupervised tasks, and experimental results corroborate our ideas and demonstrate the correctness and effectiveness of implicit regularizers. |
| Researcher Affiliation | Academia | 1National Laboratory of Pattern Recognition, CASIA 2Center for Research on Intelligent Perception and Computing, CASIA 3Center for Excellence in Brain Science and Intelligence Technology, CAS 4University of Chinese Academy of Sciences (UCAS) |
| Pseudocode | Yes | Algorithm 1 : Self-Paced Learning via Implicit Regularizers |
| Open Source Code | No | The paper does not contain any statement or link indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | The UCI Handwritten Digit dataset 1 is used in this experiment. It consists of 2,000 handwritten digits classified into ten categories (09). Each instance is represented in terms of six kinds of features (or views). Here we make use of all the six views for all the comparing algorithms. The baseline algorithms are standard k-means on each single view s representation (the best single view result is reported as BSV), and Con-MC (the features are concatenated on all views firstly, and then standard k-means is applied). Three commonly used metrics are adopted: clustering ac- 1https://archive.ics.uci.edu/ml/datasets |
| Dataset Splits | Yes | We use 10-fold cross validation for all the databases, and report both their mean values and their standard derivations. Classification accuracy is used for performance measure. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used to run the experiments (e.g., CPU, GPU models, or memory specifications). |
| Software Dependencies | No | The paper mentions "Liblinear (Fan et al. 2008) is used as the solver of LR" but does not specify a version number for Liblinear or any other software dependencies. |
| Experiment Setup | Yes | There are two hyper-parameter (λ, μ) that need to be tuned in Algorithm 1. We follow a standard setting in SPL (Kumar, Packer, and Koller 2010). That is, λ is initialized to obtain about half samples, then it is iteratively updated to involve more and more samples gradually. The practical updating direction depends on the specific minimizer function. For functions given in Table 1, λT +1 = λT /μ for L1-L2 while λT +1 = λT μ for Huber, Cauchy and Welsch, where μ > 1 is a step factor and T is an iteration number. Similar settings are adjusted for the competing SPL regularizers, including SPL-hard (Kumar, Packer, and Koller 2010) and SPL-mixture (Zhao et al. 2015). |