🚀Happy to share our paper at @NeurIPSConf with @BertrandThirion @ogrisel @pneuvial: It introduces KOPI, a novel method for statistically controlled variable selection based on Knockoffs. Come check out our poster #1004 at #NeurIPS2023!. 🧵1/6.

🧠We’ve also employed KOPI on fMRI and genomics data. The aim of fMRI data analysis is to recover relevant brain regions for a given cognitive task as shown below. These results are reproducible using

📊We’ve put KOPI to the test against other state-of-the-art Knockoffs-based method on simulated data. KOPI controls the FDP in all settings and offers superior or equal power compared to other methods.

We can then build a JER controlling family, from which we obtain valid FDP upper bounds (see Blanchard et al., 2020). 🪄 This idea is directly transposable to the aggregated case!

🎯Our aim is twofold: first, obtaining FDP control rather than FDR control (expected FDP) and second; robustifying Knockoffs inference using aggregation. To compute FDP upper bounds, we need to characterize the null distribution of π-statistics.

📖 Knockoffs are a popular framework for statistically controlled variable selection – the idea of Knockoffs is to build noisy copies of the original variables that serve as controls in the variable selection process.

RT @CEA_Officiel: #IA 🤖 ⏐En direct du centre @CEAParisSaclay, à NeuroSpin, où nous accueillons les équipes de @FranceInter pour une journée….

RT @alexisthual: Our paper on using Optimal Transport to compare human cortical surfaces was accepted at #NeurIPS22 🥳.We implement a new OT….

All results are fully reproducible using which relies on @nilearn and sanssouci.python. Computing lower bounds takes ~1 minute per fMRI dataset. This work was done at/with @Parietal_INRIA @BertrandThirion @pneuvial @InstitutDATAIA @Inria_Saclay.

These gains are observed on 36 fMRI datasets from Neurovault that differ across several dimensions (number of subjects, nature of contrasts,.fMRI study, quantity of signal…) and for targeted FDPs:

Once this family is selected, TDP bounds can be computed. Notip can also be used to find the largest possible brain region with at least (say) 90% of truly activated voxels with a high probability; Notip yields substantially more detections than parametric methods:

Instead of using parametric - most often linear - family shapes for calibration, we propose to learn such families directly. From permuted p-values curves, we extract quantile curves that serve as a non-parametric set of families. An adequate family is then selected from this set

Once permuted p-values are computed, one can choose an adequate threshold family - i.e., the least conservative possible family that controls crosses at most α% of permuted p-values curves. This procedure is called calibration.

To compute TDP lower bounds, we need to estimate the null distribution of p-values (see e.g. Blanchard et al., Annals of Stats 2020) by using permutations.

Notip provides lower confidence bounds on the True Discovery Proportion (TDP) of any brain region, or equivalently upper bounds on the False Discovery Proportion (TDP = 1 - FDP). This guarantee is directly interpretable, unlike the expected proportion of False Discoveries (FDR).

How many active voxels are there in this cluster ? This question can be addressed using Notip: Non-parametric True Discovery Proportion control for brain imaging published in @NeuroImage_EiC. Pre-print: