Search for: All records

We present an online post-hoc calibration method, called Online Platt Scaling (OPS), which combines the Platt scaling technique with online logistic regression. We demonstrate that OPS smoothly adapts between i.i.d. and non-i.i.d. settings with distribution drift. Further, in scenarios where the best Platt scaling model is itself miscalibrated, we enhance OPS by incorporating a recently developed technique called calibeating to make it more robust. Theoretically, our resulting OPS+calibeating method is guaranteed to be calibrated for adversarial outcome sequences. Empirically, it is effective on a range of synthetic and real-world datasets, with and without distribution drifts, achieving superior performance without hyperparameter tuning. Finally, we extend all OPS ideas to the beta scaling method. 

We study the problem of making calibrated probabilistic forecasts for a binary sequence generated by an adversarial nature. Following the seminal paper of Foster and Vohra (1998), nature is often modeled as an adaptive adversary who sees all activity of the forecaster except the randomization that the forecaster may deploy. A number of papers have proposed randomized forecasting strategies that achieve an ϵ-calibration error rate of O(1/sqrt T), which we prove is tight in general. On the other hand, it is well known that it is not possible to be calibrated without randomization, or if nature also sees the forecaster's randomization; in both cases the calibration error could be Ω(1). Inspired by the equally seminal works on the "power of two choices" and imprecise probability theory, we study a small variant of the standard online calibration problem. The adversary gives the forecaster the option of making two nearby probabilistic forecasts, or equivalently an interval forecast of small width, and the endpoint closest to the revealed outcome is used to judge calibration. This power of two choices, or imprecise forecast, accords the forecaster with significant power -- we show that a faster ϵ-calibration rate of O(1/T) can be achieved even without deploying any randomization. 

A multiclass classifier is said to be top-label calibrated if the reported probability for the predicted class -- the top-label -- is calibrated, conditioned on the top-label. This conditioning on the top-label is absent in the closely related and popular notion of confidence calibration, which we argue makes confidence calibration difficult to interpret for decision-making. We propose top-label calibration as a rectification of confidence calibration. Further, we outline a multiclass-to-binary (M2B) reduction framework that unifies confidence, top-label, and class-wise calibration, among others. As its name suggests, M2B works by reducing multiclass calibration to numerous binary calibration problems, each of which can be solved using simple binary calibration routines. We instantiate the M2B framework with the well-studied histogram binning (HB) binary calibrator, and prove that the overall procedure is multiclass calibrated without making any assumptions on the underlying data distribution. In an empirical evaluation with four deep net architectures on CIFAR-10 and CIFAR-100, we find that the M2B + HB procedure achieves lower top-label and class-wise calibration error than other approaches such as temperature scaling. 

Intellectual and Developmental Disabilities (IDDs), such as Down syndrome, Fragile X syndrome, Rett syndrome, and autism spectrum disorder, usually manifest at birth or early childhood. IDDs are characterized by significant impairment in intellectual and adaptive functioning, and both genetic and environmental factors underpin IDD biology. Molecular and genetic stratification of IDDs remain challenging mainly due to overlapping factors and comorbidity. Advances in high throughput sequencing, imaging, and tools to record behavioral data at scale have greatly enhanced our understanding of the molecular, cellular, structural, and environmental basis of some IDDs. Fueled by the “big data” revolution, artificial intelligence (AI) and machine learning (ML) technologies have brought a whole new paradigm shift in computational biology. Evidently, the ML-driven approach to clinical diagnoses has the potential to augment classical methods that use symptoms and external observations, hoping to push the personalized treatment plan forward. Therefore, integrative analyses and applications of ML technology have a direct bearing on discoveries in IDDs. The application of ML to IDDs can potentially improve screening and early diagnosis, advance our understanding of the complexity of comorbidity, and accelerate the identification of biomarkers for clinical research and drug development. For more than five decades, the IDDRC network has supported a nexus of investigators at centers across the USA, all striving to understand the interplay between various factors underlying IDDs. In this review, we introduced fast-increasing multi-modal data types, highlighted example studies that employed ML technologies to illuminate factors and biological mechanisms underlying IDDs, as well as recent advances in ML technologies and their applications to IDDs and other neurological diseases. We discussed various molecular, clinical, and environmental data collection modes, including genetic, imaging, phenotypical, and behavioral data types, along with multiple repositories that store and share such data. Furthermore, we outlined some fundamental concepts of machine learning algorithms and presented our opinion on specific gaps that will need to be filled to accomplish, for example, reliable implementation of ML-based diagnosis technology in IDD clinics. We anticipate that this review will guide researchers to formulate AI and ML-based approaches to investigate IDDs and related conditions.

 

UV‐ranged micro‐LEDs are being explored for numerous applications due to their high stability and power efficiency. However, previous reports have shown reduced external quantum efficiency (EQE) and increased leakage current due to the increase in surface‐to‐volume ratio with a decrease in the micro‐LED size. Herein, the size‐related performance for UV‐A micro‐LEDs, ranging from 8 × 8 to 100 × 100 μm², is studied. These devices exhibit reduced leakage current with the implementation of atomic layer deposition‐based sidewall passivation. A systematic EQE comparison is performed with minimal leakage current and a size‐independent on‐wafer EQE of around 5.5% is obtained. Smaller sized devices experimentally show enhanced EQE at high current density due to their improved heat dissipation capabilities. To the best of authors’ knowledge, this is the highest reported on‐wafer EQE demonstrated in <10 μm dimensioned 368 nm UV LEDs.

 

Gene regulatory networks underpin stress response pathways in plants. However, parsing these networks to prioritize key genes underlying a particular trait is challenging. Here, we have built the Gene Regulation and Association Network (GRAiN) of rice ( Oryza sativa ). GRAiN is an interactive query-based web-platform that allows users to study functional relationships between transcription factors (TFs) and genetic modules underlying abiotic-stress responses. We built GRAiN by applying a combination of different network inference algorithms to publicly available gene expression data. We propose a supervised machine learning framework that complements GRAiN in prioritizing genes that regulate stress signal transduction and modulate gene expression under drought conditions. Our framework converts intricate network connectivity patterns of 2160 TFs into a single drought score. We observed that TFs with the highest drought scores define the functional, structural, and evolutionary characteristics of drought resistance in rice. Our approach accurately predicted the function of OsbHLH148 TF, which we validated using in vitro protein-DNA binding assays and mRNA sequencing loss-of-function mutants grown under control and drought stress conditions. Our network and the complementary machine learning strategy lends itself to predicting key regulatory genes underlying other agricultural traits and will assist in the genetic engineering of desirable rice varieties. 

Transcription factors (TFs) play a central role in regulating molecular level responses of plants to external stresses such as water limiting conditions, but identification of such TFs in the genome remains a challenge. Here, we describe a network-based supervised machine learning framework that accurately predicts and ranks all TFs in the genome according to their potential association with drought tolerance. We show that top ranked regulators fall mainly into two ‘age’ groups; genes that appeared first in land plants and genes that emerged later in the Oryza clade. TFs predicted to be high in the ranking belong to specific gene families, have relatively simple intron/exon and protein structures, and functionally converge to regulate primary and secondary metabolism pathways. Repeated trials of nested cross-validation tests showed that models trained only on regulatory network patterns, inferred from large transcriptome datasets, outperform models trained on heterogenous genomic features in the prediction of known drought response regulators. A new R/Shiny based web application, called the DroughtApp, provides a primer for generation of new testable hypotheses related to regulation of drought stress response. Furthermore, to test the system we experimentally validated predictions on the functional role of the rice transcription factor OsbHLH148, using RNA sequencing of knockout mutants in response to drought stress and protein-DNA interaction assays. Our study exemplifies the integration of domain knowledge for prioritization of regulatory genes in biological pathways of well-studied agricultural traits.