Stationary Activations for Uncertainty Calibration in Deep Learning
Authors: Lassi Meronen, Christabella Irwanto, Arno Solin
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate these properties on classification and regression benchmarks and a radar emitter classification task. |
| Researcher Affiliation | Collaboration | Lassi Meronen Aalto University / Saab Finland Oy Espoo, Finland lassi.meronen@aalto.fi Christabella Irwanto Aalto University Espoo, Finland christabella.irwanto@aalto.fi Arno Solin Aalto University Espoo, Finland arno.solin@aalto.fi |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Example codes implementing the proposed methods in this paper are available at https://github. com/Aalto ML/stationary-activations. |
| Open Datasets | Yes | Illustrative toy examples In Fig. 1, we consider the binary Banana classification tasks under the presence of various GP priors. Benchmark classification tasks In Table 1, we consider UCI benchmark classification tasks (including one small-data example) where we compare classification accuracy and negative log predictive density (NLPD) that penalizes both misclassification and miscalibrated uncertainty. Out-of-distribution characterization with CIFAR-10 As a rule of thumb the uncertainty of OOD samples should be high and uncertainty of in-distribution samples should be low. |
| Dataset Splits | Yes | (10-fold cv) NLPD ACC n d c SVGP GPDNN SV-DKL Matérn act. Each model was trained with only images of five classes {plane, car, bird, cat, deer}. During testing, images from all 10 classes were present (now including also {ship, truck, frog, dog, horse}). |
| Hardware Specification | No | The paper mentions 'We acknowledge the computational resources provided by the Aalto Science-IT project' but does not specify any hardware details like GPU/CPU models or memory. |
| Software Dependencies | No | The experiments were implemented in GPflow [43] (GPs and GPDNN), GPy Torch [22] (SV-DKL), and the rest in Py Torch (see App. B). |
| Experiment Setup | No | The NN architectures in all methods are the same (a fully connected network with layers d-1000-1000-500-50-c). For all neural network models using the Matérn activation functions the length-scale parameter ℓis fixed as the preceding layer(s) take care of scaling the inputs, which serves the same purpose. App. B lists full details of all the experiments. |