Lossless or Quantized Boosting with Integer Arithmetic
Authors: Richard Nock, Robert Williamson
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments display that the algorithm can achieve a fast convergence during the early boosting rounds compared to Ada Boost, even with a weight storage that can be 30+ times smaller. Experiments are carried out with 10-folds stratified cross validation on 25 UCI domains (Blake et al., 1998). Figure 1: Left pane: test error (y, standard deviation) on UCI domain Hardware vs iteration number t (x, logscale), for Ada Boost (Schapire & Singer, 1999) (blue) vs RATBOOST (green). We observe that RATBOOST tends to converge faster to good solutions during early iterations. |
| Researcher Affiliation | Collaboration | Richard Nock 1 2 3 Robert C. Williamson 2 1 1Data61, 2The Australian National University, Canberra & 3The University of Sydney, Sydney (Australia). Correspondence to: <richard.nock@data61.csiro.au, bob.williamson@anu.edu.au>. |
| Pseudocode | Yes | Algorithm 1 RATBOOST Input sample S = {(xi, yi), i = 1, 2, ..., m}, number of iterations T, ϱ N , a Q+ ; Step 1 : let α 0; Step 2 : let wi = 1/2, i = 1, 2, ..., m; // initial weights Step 3 : for t = 1, 2, ..., T Step 3.1 : let j WL(S, w) Step 3.2 : let η(j) (1/m) P i wiyihj(xi) Step 3.3 : let δj aη(j) Step 3.4 : for i = 1, 2, ..., m, let wi δjyihj(xi) wi ; (21) Step 3.5 : let αj αj + δj // update of αj Return HT = P |
| Open Source Code | Yes | Code available at: http://users.cecs.anu.edu.au/~rnock/code/ratboost.zip |
| Open Datasets | Yes | Experiments are carried out with 10-folds stratified cross validation on 25 UCI domains (Blake et al., 1998). |
| Dataset Splits | Yes | Experiments are carried out with 10-folds stratified cross validation on 25 UCI domains (Blake et al., 1998). |
| Hardware Specification | No | The paper states 'All algorithms are run for 10 000 iterations' but does not provide any specific details regarding the hardware used for these experiments (e.g., GPU models, CPU types, memory). |
| Software Dependencies | No | The paper describes the algorithms implemented and variants tested ('We implemented the following variants of RATBOOST'), but it does not specify any software dependencies with version numbers (e.g., programming languages, libraries, or frameworks like Python, PyTorch, scikit-learn with their specific versions). |
| Experiment Setup | Yes | In all algorithms, we fix ϱ = 1, estimate M for assumption (M) from the training sample and then compute a as in Theorem 7 with π = 0. |