Compact Random Feature Maps

Authors: Raffay Hamid, Ying Xiao, Alex Gittens, Dennis Decoste

ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present experiments on multiple standard data-sets with performance competitive with state-of-the-art results. We now present CRAFTMaps results on multiple data-sets.
Researcher Affiliation Collaboration Raffay Hamid RAFFAY@CC.GATECH.EDU e Bay Research Laboratory Ying Xiao YING.XIAO@GATECH.EDU Georgia Institute of Technology Alex Gittens AGITTENS@EBAY.COM e Bay Research Laboratory Dennis De Coste DDECOSTE@EBAY.COM e Bay Research Laboratory
Pseudocode Yes Algorithm 1 RANDOM FEATURE MAPS (RFM) Algorithm 2 CRAFTMAPS USING RFM
Open Source Code No The paper does not contain any statements about releasing source code or provide links to a code repository.
Open Datasets Yes Figure 5 shows the normalized root mean square errors (NRMSE) for MNIST data obtained for an r = 7 and q = 1 kernel using random feature maps... Figure 6 where NRMSE for 6 different data-sets over a range of E are shown. Table 1 shows the test classification error using CRAFTMaps on random feature maps (Kar & Karnick, 2012) and tensor sketch (Pham & Pagh, 2013) compared to 5 alternate approaches over 4 different data-sets. (These datasets are MNIST, USPS, COIL100, PENDIGITS) Results on MNIST8M Data Figure 7 (b) shows the comparative performance of CRAFTMaps for a given sized E (2^14) as training size varies from 60K to 8.1M.
Dataset Splits No The paper mentions 'training data' and 'test points' and uses 'randomly selected' or 'randomly picked' points, but it does not specify explicit train/validation/test splits, percentages, or absolute counts for validation data.
Hardware Specification No The paper mentions 'These times were recorded for MNIST data using a 40-core machine.' This provides a general core count but lacks specific CPU models, GPU details, or other detailed hardware specifications.
Software Dependencies No The paper mentions 'BLAS3' in relation to exploiting multi-core processing power, but it does not list any specific software libraries or their version numbers.
Experiment Setup Yes Unless otherwise mentioned, we use the H-0/1 heuristic of (Kar & Karnick, 2012) for random feature maps (RFM) and CRAFTMaps on RFM. These results were obtained for an r = 7 and q = 1 kernel. Here r = 7, 7, 5, 5 and 9 respectively while q = 1. These results were obtained using a polynomial kernel with r = 7, q = 1, D = 2^17, E = 2^14, and ECOCs equal to 200.