Combination Of Model Selection Research Articles

Comparative Evaluation of LLMs in Clinical Oncology.

As artificial intelligence (AI) tools become widely accessible, more patients and medical professionals will turn to them for medical information. Large language models (LLMs), a subset of AI, excel in natural language processing tasks and hold considerable promise for clinical use. Fields such as oncology, in which clinical decisions are highly dependent on a continuous influx of new clinical trial data and evolving guidelines, stand to gain immensely from such advancements. It is therefore of critical importance to benchmark these models and describe their performance characteristics to guide their safe application to clinical oncology. Accordingly, the primary objectives of this work were to conduct comprehensive evaluations of LLMs in the field of oncology and to identify and characterize strategies that medical professionals can use to bolster their confidence in a model's response. This study tested five publicly available LLMs (LLaMA 1, PaLM 2, Claude-v1, generative pretrained transformer 3.5 [GPT-3.5], and GPT-4) on a comprehensive battery of 2044 oncology questions, including topics from medical oncology, surgical oncology, radiation oncology, medical statistics, medical physics, and cancer biology. Model prompts were presented independently of each other, and each prompt was repeated three times to assess output consistency. For each response, models were instructed to provide a self-appraised confidence score (from 1 to 4). Model performance was also evaluated against a novel validation set comprising 50 oncology questions curated to eliminate any risk of overlap with the data used to train the LLMs. There was significant heterogeneity in performance between models (analysis of variance, P<0.001). Relative to a human benchmark (2013 and 2014 examination results), GPT-4 was the only model to perform above the 50th percentile. Overall, model performance varied as a function of subject area across all models, with worse performance observed in clinical oncology subcategories compared with foundational topics (medical statistics, medical physics, and cancer biology). Within the clinical oncology subdomain, worse performance was observed in female-predominant malignancies. A combination of model selection, prompt repetition, and confidence self-appraisal allowed for the identification of high-performing subgroups of questions with observed accuracies of 81.7 and 81.1% in the Claude-v1 and GPT-4 models, respectively. Evaluation of the novel validation question set produced similar trends in model performance while also highlighting improved performance in newer, centrally hosted models (GPT-4 Turbo and Gemini 1.0 Ultra) and local models (Mixtral 8×7B and LLaMA 2). Of the models tested on a standardized set of oncology questions, GPT-4 was observed to have the highest performance. Although this performance is impressive, all LLMs continue to have clinically significant error rates, including examples of overconfidence and consistent inaccuracies. Given the enthusiasm to integrate these new implementations of AI into clinical practice, continued standardized evaluations of the strengths and limitations of these products will be critical to guide both patients and medical professionals. (Funded by the National Institutes of Health Clinical Center for Research and the Intramural Research Program of the National Institutes of Health; Z99 CA999999.).

Adaptive Non-Parametric Instrumental Regression in the Presence of Dependence

We consider the estimation of a structural function which models a non-parametric relationship between a response and an endogenous regressor given an instrument in presence of dependence in the data generating process. Assuming an independent and identically distributed (iid.) sample it has been shown in Johannes, J. and M. Schwarz [2011] that a least squares estimator based on dimension reduction and thresholding can attain minimax-optimal rates of convergence up to a constant. As this estimation procedure requires an optimal choice of a dimension parameter with regard amongst others to certain characteristics of the unknown structural function we investigate its fully data-driven choice based on a combination of model selection and Lepski's method inspired by Goldenshluger, A. and O. Lepski [2011]. For the resulting fully data-driven thresholded least squares estimator a non-asymptotic oracle risk bound is derived by considering either an iid. sample or by dismissing the independence assumption. In both cases the derived risk bounds coincide up to a constant assuming sufficiently weak dependence characterised by a fast decay of the mixing coefficients. Employing the risk bounds the minimax optimality up to constant of the estimator is established over a variety of classes of structural functions.

Strategy of active learning support vector machine for image retrieval

This study proposes a new method for content-based image retrieval by finding an optimal classifier. The optimal classifier is achieved by a new active learning support vector machine (SVM) which combines the model selection with the active learning. The unlabelled samples close to the boundary of the SVM classifier are selected based on the feature similarity for the active learning, and the adaptive regularisation is used to select the optimal model. The combination of model selection with active learning accelerates the convergence of the classifier. The new method can improve the image retrieval accuracy and reduce the time consumption. The experimental results show that the proposed method has a better performance with fewer samples and less time consumption for image retrieval.

Adaptive estimation of linear functionals in functional linear models

We consider the estimation of the value of a linear functional of the slope parameter in functional linear regression, where scalar responses are modeled in dependence of randomfunctions. In Johannes and Schenk [2010] it has been shown that a plug-in estimator based on dimension reduction and additional thresholding can attain minimax optimal rates of convergence up to a constant. However, this estimation procedure requires an optimal choice of a tuning parameter with regard to certain characteristics of the slope function and the covariance operator associated with the functional regressor. As these are unknown in practice, we investigate a fully data-driven choice of the tuning parameter based on a combination of model selection and Lepski’s method, which is inspired by the recent work of Goldenshluger and Lepski [2011]. The tuning parameter is selected as theminimizer of a stochastic penalized contrast function imitating Lepski’smethod among a random collection of admissible values. We show that this adaptive procedure attains the lower bound for the minimax risk up to a logarithmic factor over a wide range of classes of slope functions and covariance operators. In particular, our theory covers pointwise estimation as well as the estimation of local averages of the slope parameter.

A note on the generalized degrees of freedom under the L1 loss function

Generalized degrees of freedom measure the complexity of a modeling procedure; a modeling procedure is a combination of model selection and model fitting. In this manuscript, we consider two definitions of generalized degrees of freedom for a modeling procedure under the L 1 loss function, and investigate the connections between those two definitions. We also propose the extended Akaike information criterion, the adaptive model selection, and the extended generalized cross-validation under the L 1 loss function. Finally, we extend the results to M-estimation.