From 92b2a7d364d68df6d4dbcb257121220083c783d4 Mon Sep 17 00:00:00 2001 From: Chris Endemann Date: Fri, 29 Nov 2024 09:58:49 -0600 Subject: [PATCH] Update 6-confidence-intervals.md --- episodes/6-confidence-intervals.md | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/episodes/6-confidence-intervals.md b/episodes/6-confidence-intervals.md index 23e476a7..9a43d584 100644 --- a/episodes/6-confidence-intervals.md +++ b/episodes/6-confidence-intervals.md @@ -94,10 +94,12 @@ To help guide method selection, here are rough definitions for **model size**, * #### Application: Hospital resource allocation **Aleatoric Uncertainty:** + - **Linear models:** Predictive variance captures inherent randomness in hospital occupancy patterns due to seasonal variability or unpredictable local events. See [Taylor et al., 2021](https://bmcmedinformdecismak.biomedcentral.com/articles/10.1186/s12911-022-01787-9) for regression-based modeling of hospital bed demand during flu seasons. - **Deep learning:** Heteroscedastic models account for input-dependent noise, such as variability in hospital reporting systems or random fluctuations in patient arrival rates during holidays or flu season. See [Rajkomar et al., 2018](https://www.nature.com/articles/s41746-018-0029-1.pdf) for applications of deep learning in hospital resource prediction. **Epistemic Uncertainty:** + - **Tree ensembles:** Capture uncertainty for underrepresented or novel conditions, such as predicting hospital demand during rare pandemics. For an example, see [Shahid et al., 2020](https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1010602), which uses ensemble models for pandemic demand forecasting. - **OOD detection:** Identifies anomalies in resource usage data, such as unexpected spikes in equipment demands or misreported occupancy. See [Pang et al., 2021](https://arxiv.org/abs/2110.11334) for a comprehensive survey on OOD detection in real-world tabular data. @@ -109,10 +111,12 @@ To help guide method selection, here are rough definitions for **model size**, * #### Application: Tumor detection in radiology images **Aleatoric Uncertainty:** + - **CNNs:** Heteroscedastic loss accounts for noise from imaging artifacts like low resolution or motion blur. For an example of modeling aleatoric uncertainty in medical imaging, see [Kendall & Gal, 2017](https://arxiv.org/abs/1703.04977). - **Data augmentation:** Synthetic noise during training improves robustness to real-world imperfections in medical images. See [Shorten & Khoshgoftaar, 2019](https://journalofbigdata.springeropen.com/articles/10.1186/s40537-019-0197-0) for a survey on data augmentation strategies in deep learning. **Epistemic Uncertainty:** + - **Monte Carlo dropout:** Samples multiple outputs to generate uncertainty maps for tumor boundaries in rare conditions. See [Leibig et al., 2017](https://www.nature.com/articles/s41598-017-17876-z.pdf) for an application of MC dropout to retinal disease detection. - **OOD detection:** Flags anomalous radiology images, such as previously unseen tumor types or imaging modalities. See [Hendrycks & Gimpel, 2017](https://arxiv.org/abs/1610.02136) for the foundational OOD detection method applied to medical imaging. :::::::::::::::::::::::::::::: @@ -123,10 +127,12 @@ To help guide method selection, here are rough definitions for **model size**, * #### Application: Chatbot intent recognition **Aleatoric Uncertainty:** + - **Logistic regression:** Predictive variance highlights ambiguities in user queries with mixed intents. See [Hazra et al., 2020](https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00483/111592/Uncertainty-Estimation-and-Reduction-of-Pre) for a discussion on uncertainty estimation in NLP tasks. - **Transformers:** Heteroscedastic models capture token-level uncertainty in noisy or ambiguous language inputs. See [Malinin & Gales, 2018](https://arxiv.org/pdf/1802.10501v3) for heteroscedastic neural network modeling in NLP. **Epistemic Uncertainty:** + - **Bayesian transformers:** Quantify uncertainty in rare or unseen query topics (e.g., domain-specific technical questions). See [Fort et al., 2020](https://openreview.net/pdf?id=CSXa8LJMttt) for a discussion on uncertainty in transformers. - **OOD detection:** Detects out-of-scope queries or previously unseen intents, such as rare idiomatic expressions or slang. See [Lin & Xu, 2019](https://aclanthology.org/2020.coling-main.125.pdf) for OOD detection in NLP systems. ::::::::::::::::::::::::::::::