Starting Small - Learning with Adaptive Sample Sizes

Authors: Hadi Daneshmand, Aurelien Lucchi, Thomas Hofmann

ICML 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Experimental Results We present experimental results on synthetic as well as real-world data, which largely confirms the above analysis.
Researcher Affiliation Academia Department of Computer Science, ETH Zurich, Switzerland
Pseudocode Yes Algorithm 2 DYNASAGA
Open Source Code No No concrete access to source code (specific repository link, explicit code release statement, or code in supplementary materials) for the methodology described in this paper.
Open Datasets Yes Table 2. Details of the real datasets used in our experiments. All datasets were selected from the LIBSVM dataset collection.
Dataset Splits Yes The training set includes 90% of the data. The vertical axis shows the suboptimality of the expected risk, i.e. log2 E10 RS(wt) RS(w T ) , where S is a test set which includes 10% of the data and w T is the optimum of the empirical risk on T .
Hardware Specification No No specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments are provided.
Software Dependencies No No specific ancillary software details (e.g., library or solver names with version numbers) are provided.
Experiment Setup Yes Throughout all the experiments we used the logistic loss with a regularizer λ = 1 n 3. The various parameters used for the baseline methods are described in Table 3. A critical factor in the performance of most baselines, especially SGD, is the selection of the step-size. We picked the best-performing step-size within the common range guided by existing theoretical analyses, specifically η = 1/L and η = C C+µt for various values of C.