All your loss are belong to Bayes
Authors: Christian Walder, Richard Nock
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical experiments are provided in Section 5 before concluding with Section 6. and We provide illustrative examples and quantitative comparisons of the ISGP prior in univariate regression/classification, and inferring proper losses. |
| Researcher Affiliation | Collaboration | Christian Walder 1 2 Richard Nock 1 2 1 CSIRO Data61, Australia. 2 The Australian National University. |
| Pseudocode | No | The paper describes computational procedures (e.g., Laplace approximation, EM) but does not present them in a structured pseudocode or algorithm block. |
| Open Source Code | Yes | The code for the real world problems of learning the loss is available online.2 https://gitlab.com/cwalder/linkgistic |
| Open Datasets | Yes | The task was to regress each of the four real-valued features of the auto-mpg dataset [DG17] onto the target (car efficiency in miles per gallon). and We further bench-marked ISGP-LINKGISTIC against GP-LINKGISTIC and logistic regression (as the latter was the strongest practical algorithm in the experiments of [NM20]) on a broader set of tasks, namely the three MNIST-like datasets of [LC10, XRV17, CBIK+18]. |
| Dataset Splits | Yes | We partitioned the dataset into five splits. For the Small problems, we trained on one split and tested on the remaining four ; for the Large problems we did the converse. We repeated this all C5 1 = C5 4 = 5 ways and averaged the results to obtain Table 1. |
| Hardware Specification | No | The paper states 'The experiments took roughly two CPU months' but does not provide any specific details about the hardware used (e.g., CPU models, GPU models, memory, or cloud instance types). |
| Software Dependencies | No | Optimisation was performed with L-BFGS [BLNZ95]. Gradients were computed with automatic differentiation (with the notable exception of (13)) using PyTorch [PGM+19]. While PyTorch is mentioned, a specific version number is not provided, nor are versions for other software dependencies. |
| Experiment Setup | Yes | We fixed M = 64 in (14). and We chose set µ = 0 and γ = 0.01. We set the length-scale parameter a = 1.2 and chose b such that k(0, 0) = 1 (using (15)) by (10) this makes ν(x) = x the most likely source a priori, to roughly bias towards traditional logistic regression. |