Ignacio Martinez’s Post

An article I found helpful and keep coming back to is McGrath et al. (2020). Why? When I first learned about the g-formula, I found it quite interesting. However, I always learn better with an example. This paper gives examples! It also gives use cases, R code and sample data (that can be loaded with the package). This paper was instrumental in helping me learn the parametric g-formula (a great tool for causal inference). Highly recommend for anyone interested in it! [Link: https://lnkd.in/eAv2NWzp] #CausallyCurious #RealWorldEvidence #CausalInference #RealWorldData

I'm pleased to announce that our paper with Ángelo Gárate Barraza and Cibele Russo, "The gradient test statistic for outlier detection in generalized estimating equations", has just been published online in Statistics & Probability Letters! In this paper, we develop diagnostic tools for estimating equations, useful for the analysis of data with longitudinal structure. The gradient statistic introduced by Terrell (2002) is used to propose a case deletion measure, as well as a statistic for the detection of outlying observations using a mean-shift outlier model. The published version can be found at https://lnkd.in/e_gckp69

Congrats to Devesh Batra, Salvatore Mercuri, and Raad Khraishi for publishing their paper "Conformal Predictions for Longitudinal Data"! https://lnkd.in/esKzVviu From the paper: "...we proposed Longitudinal Predictive Conformal Inference (LPCI) as a distribution-free and model-agnostic uncertainty quantification approach for longitudinal time series data. Our method provides asymptotic theoretical guarantees across both cross-sectional and temporal dimensions. This is done without creating infinitely-wide prediction intervals at any point, which is a first in conformal predictions for longitudinal data. In experiments, we observed that LPCI was able to improve longitudinal coverage over the current state-of-the-art, whilst also exhibiting a greater adaptivity of widths."

NEW PAPER "Sample size requirements are not being considered in studies developing prediction models for binary outcomes: a systematic review" --> https://lnkd.in/eQDG3Hdc "Sample size calculation and justification is rarely reported in studies developing a prediction model for a binary outcome using logistic regression and studies often do not use enough data to meet minimum sample size requirements for their prediction model scenario. Models developed using insufficient data will lead to model instability and unreliable predictions, that if used to guide clinical decision making have the potential to cause harm."

📃Scientific paper: Time splitting method for nonlinear Schr\"odinger equation with rough initial data in $L^2$ Abstract: We establish convergence results related to the operator splitting scheme on the Cauchy problem for the nonlinear Schr\"odinger equation with rough initial data in $L^2$, $$ \left\\{ \begin\{array\}\{ll\} i\partial\_t u +\Delta u = \lambda \|u\|^\{p\} u, & \(x,t\) \in \mathbb\{R\}^d \times \mathbb\{R\}\_+, u \(x,0\) =\phi \(x\), & x\in\mathbb\{R\}^d, \end\{array\} \right. $$ where $\lambda \in \\{-1,1\\}$ and $p \>0$. While the Lie approximation $Z\_L$ is known to converge to the solution $u$ when the initial datum $\phi$ is sufficiently smooth, the convergence result for rough initial data is open to question. In this paper, for rough initial data $\phi\in L^2 \(\mathbb\{R\}^d\)$, we prove the convergence of the filtered Lie approximation $Z\_\{flt\}$ to the solution $u$ in the mass-subcritical range, $\max\left\\{1,\frac\{2\}\{d\}\right\\} \leq p \< \frac\{4\}\{d\}$. Furthermore, we provide a precise convergence result for radial initial data $\phi\in L^2 \(\mathbb\{R\}^d\)$, ;Comment: typos are fixed Discover the rest of the scientific article on es/iode ➡️https://etcse.fr/GjCh1

Having delved into the paper, "Data-Free Knowledge Distillation with Soft Targeted Transfer Set Synthesis," it's evident that the objective function presents a non-convex challenge. Intriguingly, attempts to render it convex led to an outcome of unstable Knowledge Distillation. In my pursuit to adeptly formalize convex optimization equations, I've taken the step of enrolling in "Optimization for the Information and Data Sciences" (EE-588, also crosslisted as CSCI-612). Commencing with foundational definitions, the course equips us with tools to elegantly convert seemingly intractable non-convex problems into manageable convex optimization scenarios, all within the grasp of contemporary technology. Eager to unravel the realm of convexity and its practical applications! 📚🔍 #ConvexOptimization #LifelongLearner

A common concern about #MachineLearning methods is that the results of an analysis are less interpretable than conventional statistical methods. When I teach I like to flip this myth on its head, and illustrate all of the rich approaches to interpretation with ML methods that are more limited or not possible with conventional statistical approaches. One example of this includes assessing the predictive contributions of variables based on methods that include automatic approaches to regularization and/or variable selection. Another approach is to check for the detection of, and then visualize, complex interactions between variables. One great resource that I'll illustrate in today's workshop session is the randomForestExplainer package in R https://lnkd.in/ekZF72ng, developed by Ola Paluszynska and colleagues. This package offers several utilities for understanding feature importance and the nature of effects from a random forest.

AIET 2024 will be published as a volume of the Springer book series Lecture Notes on Data Engineering and Communications Technologies.  Indexed by Ei Compendex, Scopus. Submission: https://aiet.org/sub.html Topics: https://aiet.org/cfp.html

We are happy to share ZLaP, "Label Propagation for Zero-shot Classification with Vision-Language Models", that will be presented at CVPR 2024. Work done with Yannis Kalantidis, and Giorgos Tolias. We tackle the case of zero-shot classification in the presence of unlabeled data. We leverage the graph structure of the unlabeled data and introduce ZLaP, a method based on label propagation (LP) that utilizes geodesic distances for classification. Our main contributions are: - We tailor label propagation to VLMs and zero-shot classification over bi-modal graphs, proposing per modality neighbor search and edge weight balancing.  - We propose an efficient way for performing inductive inference with label propagation via a dual solution and through sparsification. This not only improves the test-time efficiency of our method but also performance.  - We show that geodesic similarity aligns text and image embeddings from VLMs better than commonly used cosine similarity. - Our method complemented with the class proxies presented in concurrent work (InMaP) achieves state-of-the-art results for zero-shot on 14 common datasets. - We also show that ZLaP has potentially to be used even with web-crawled data from LAION-400M. If you will be in Seattle at CVPR, visit our poster at the fifth poster session on June 21th. Paper: https://lnkd.in/eFehav3v Code: https://lnkd.in/eNkQSDJU

Zero-shot visual recognition with Vision-Language Models is performed with a simple non-parametric classifier, i.e. nearest neighbor assignment based on cosine similarity. We are improving this by exploiting unlabeled images and a non-parametric classifier that operates with geodesic similarities derived via a bimodal graph. Despite CLIP being trained to bring text and images close, the two modalities are still far apart in the common feature space, which is not the case with our geodesic similarities (see the tSNE plots). Work to be presented next week at CVPR 2024.