Sparse autoencoders inserted into VLMs and trained only for reconstruction can reliably detect adversarial attacks on images, including unseen domains and attack types.
Tool reference
Title resolution pending
Tool reference. 100% of classified Pith citations use this work as a method, library, or software dependency, not as a substantive claim.
18
Pith papers citing it
30.2k
external citations · Crossref
Method reference
100% of classified citations
citation-role summary
method 5
citation-polarity summary
authors
co-cited works
roles
method 5polarities
use method 5representative citing papers
LG-CoTrain, an LLM-guided co-training method, outperforms classical semi-supervised baselines for crisis tweet classification in low-resource settings with 5-25 labeled examples per class.
Neural point-forms are introduced as permutation-invariant neural layers that output learned form-comparison matrices for point clouds, with a claimed consistency proof under sampling and manifold assumptions and competitive results on synthetic and biological data.
A framework for cross-validation optimal feature selection in linear SVM classification is developed by reformulating the bilevel problem into a single-level mixed-integer optimization using LS-SVM, with simulation results indicating competitive performance.
Meta-learning with 24 classical complexity metrics predicts the optimal quantum encoding circuit among 9 candidates with up to 85.7% top-3 accuracy.
Quantum kernel methods show no statistically significant edge over strong classical baselines on tabular classification tasks, with current feature maps failing to match the spectral properties of the best classical kernel.
The synthetic prior for tabular foundation models covers only a narrow part of real table distributions, but this mismatch does not degrade model generalization.
No evidence found; excludes semiclassical black holes below 8.4-11.4 TeV and string balls below 9.0-10.7 TeV at 95% CL, and caps the quark-quark sphaleron fraction above 9 TeV at 0.0034.
Proposes a framework for collaborative dataset construction and smart-contract-hosted ML models on blockchain, with financial and gamified incentives to sustain accuracy.
Opcode-sequence generative models produce synthetic malware data that raises minor-class classification accuracy by up to 60% and overall detection to 96%.
Nested cross-validation reveals optimistic bias in standard validation for EEG alcoholism classification, with AdaBoost reaching 78.3% accuracy and most model differences not statistically significant per McNemar's test.
Higher-quality automatic speech recognition transcripts enable simple lexical models to achieve better Alzheimer's disease detection performance on the ADReSSo dataset.
A modular EEG-based BCI with S4D deep learning classifier achieves 84% offline accuracy and enables real-time control for a tetraplegic user, with 73% success in post-competition validation.
Joint modeling of multiple subjects' fMRI data produces low-dimensional embeddings that outperform raw high-dimensional voxel spaces on music genre and language topic classification while increasing semantic richness.
A preprocessing pipeline for resting-state and motor-task EEG is described to support future machine learning models that predict treatment efficacy in chronic neck pain.
A hybrid deep learning plus classical ML pipeline for waste image classification reaches up to 100% accuracy on TrashNet and a corrected household dataset while cutting feature dimensionality by over 95%.
DistilBERT achieves 84.78% accuracy and 84.75% F1-score on binary sentiment classification of Indonesian student opinions about AI in higher education, outperforming SVM at 82.14%.
citing papers explorer
-
Sparse Autoencoders as Plug-and-Play Firewalls for Adversarial Attack Detection in VLMs
Sparse autoencoders inserted into VLMs and trained only for reconstruction can reliably detect adversarial attacks on images, including unseen domains and attack types.
-
LLM-guided Semi-Supervised Approaches for Social Media Crisis Data Classification
LG-CoTrain, an LLM-guided co-training method, outperforms classical semi-supervised baselines for crisis tweet classification in low-resource settings with 5-25 labeled examples per class.
-
Neural Point-Forms
Neural point-forms are introduced as permutation-invariant neural layers that output learned form-comparison matrices for point clouds, with a claimed consistency proof under sampling and manifold assumptions and competitive results on synthetic and biological data.
-
Cross-validation-based optimal feature selection for linear SVM classification
A framework for cross-validation optimal feature selection in linear SVM classification is developed by reformulating the bilevel problem into a single-level mixed-integer optimization using LS-SVM, with simulation results indicating competitive performance.
-
Towards Automated Selection of Quantum Encoding Circuits via Meta-Learning
Meta-learning with 24 classical complexity metrics predicts the optimal quantum encoding circuit among 9 candidates with up to 85.7% top-3 accuracy.
-
Benchmarking Quantum Kernel Support Vector Machines Against Classical Baselines on Tabular Data: A Rigorous Empirical Study with Hardware Validation
Quantum kernel methods show no statistically significant edge over strong classical baselines on tabular classification tasks, with current feature maps failing to match the spectral properties of the best classical kernel.
-
Mind the Gap? A Distributional Comparison of Real and Synthetic Priors for Tabular Foundation Models
The synthetic prior for tabular foundation models covers only a narrow part of real table distributions, but this mismatch does not degrade model generalization.
-
A search for microscopic black holes, string balls, and sphalerons in proton-proton collisions at $\sqrt{s}$ = 13 TeV
No evidence found; excludes semiclassical black holes below 8.4-11.4 TeV and string balls below 9.0-10.7 TeV at 95% CL, and caps the quark-quark sphaleron fraction above 9 TeV at 0.0034.
-
Decentralized & Collaborative AI on Blockchain
Proposes a framework for collaborative dataset construction and smart-contract-hosted ML models on blockchain, with financial and gamified incentives to sustain accuracy.
-
Generating Synthetic Malware Samples Using Generative AI
Opcode-sequence generative models produce synthetic malware data that raises minor-class classification accuracy by up to 60% and overall detection to 96%.
-
Impact of Validation Strategy on Machine Learning Performance in EEG-Based Alcoholism Classification
Nested cross-validation reveals optimistic bias in standard validation for EEG alcoholism classification, with AdaBoost reaching 78.3% accuracy and most model differences not statistically significant per McNemar's test.
-
Impact of automatic speech recognition quality on Alzheimer's disease detection from spontaneous speech: a reproducible benchmark study with lexical modeling and statistical validation
Higher-quality automatic speech recognition transcripts enable simple lexical models to achieve better Alzheimer's disease detection performance on the ADReSSo dataset.
-
Improving motor imagery decoding methods for an EEG-based mobile brain-computer interface in the context of the 2024 Cybathlon
A modular EEG-based BCI with S4D deep learning classifier achieves 84% offline accuracy and enables real-time control for a tetraplegic user, with 73% success in post-competition validation.
-
Low-dimensional Embodied Semantics for Music and Language
Joint modeling of multiple subjects' fMRI data produces low-dimensional embeddings that outperform raw high-dimensional voxel spaces on music genre and language topic classification while increasing semantic richness.
-
A Machine Learning Framework for EEG-Based Prediction of Treatment Efficacy in Chronic Neck Pain
A preprocessing pipeline for resting-state and motor-task EEG is described to support future machine learning models that predict treatment efficacy in chronic neck pain.
-
Towards Accurate and Efficient Waste Image Classification: A Hybrid Deep Learning and Machine Learning Approach
A hybrid deep learning plus classical ML pipeline for waste image classification reaches up to 100% accuracy on TrashNet and a corrected household dataset while cutting feature dimensionality by over 95%.
-
Sentiment Analysis of AI Adoption in Indonesian Higher Education Using Machine Learning and Transformer-Based Models
DistilBERT achieves 84.78% accuracy and 84.75% F1-score on binary sentiment classification of Indonesian student opinions about AI in higher education, outperforming SVM at 82.14%.
- Do Larger Models Really Win in Drug Discovery? A Benchmark Assessment of Model Scaling in AI-Driven Molecular Property and Activity Prediction