Integrated Gradients
PulseAugur coverage of Integrated Gradients — every cluster mentioning Integrated Gradients across labs, papers, and developer communities, ranked by signal.
5 day(s) with sentiment data
-
Saliency maps in mammography classifiers show bias against dense breasts
A new research paper published on arXiv investigates the fairness of saliency maps in mammography classifiers, specifically focusing on dense breasts where cancers can be more difficult to detect. The study found that s…
-
Transformer model predicts cesarean section risk using 3D body scans
Researchers have developed a novel multi-view Transformer network called MvBody, designed to predict the risk of cesarean sections using 3D optical body scans and self-reported medical data. This model aims to provide a…
-
Deep learning models benchmarked for breast cancer diagnosis
Researchers have evaluated ten deep learning models, including CNNs like DenseNet121 and transformer models such as ViT and Swin Transformer, for their effectiveness in diagnosing breast cancer from medical images. Dens…
-
New post-hoc FSL enhances neural network interpretability
Researchers have developed a post-hoc Feature Selection Layer (FSL) to improve the interpretability of pre-trained neural networks. This trainable module integrates with frozen models to identify and highlight the featu…
-
New metric measures AI explanation fairness across demographics
Researchers have developed a new metric, Plausibility-EFG, to measure disparities in how AI models explain their decisions across different demographic groups, specifically in hate speech detection. Using the HateXplain…
-
New Causal Methods Enhance AI Feature Attribution Scalability
Researchers have developed Causal Integrated Gradients (CIG) and Causal Expected Gradients (CEG), new methods for feature attribution in machine learning that incorporate prior causal structure among input features. The…
-
RNNs outperform traditional models in forest species classification
Researchers have developed a Recurrent Neural Network (RNN) model that demonstrates superior performance in classifying forest species, particularly in distinguishing between pure and mixed Norway spruce stands using Se…
-
SHAP attribution methods show weak causal faithfulness in language models
Researchers have evaluated the effectiveness of SHAP as an attribution method for sparse autoencoder (SAE) features in language models. Their study, using sentiment classification on the SST-2 benchmark with GPT-2 Small…
-
Graph Neural Networks Enhance Football Tactical Analysis
Researchers have developed a novel framework using graph neural networks (GNNs) combined with a sequential model to classify tactical phases in football matches on a second-by-second basis. This approach, detailed in a …
-
WOMBAT benchmark tests GNN explainers on molecular data
Researchers have developed WOMBAT, a new benchmark designed to test the accuracy of explainability tools for Graph Neural Networks (GNNs) when applied to molecular data. WOMBAT consists of 14 GNNs with manually defined …
-
New NObSP framework enhances neural network interpretability
Researchers have introduced NObSP (Nonlinear Oblique Subspace Projections), a novel framework designed to enhance the interpretability of deep neural networks. This method decomposes network predictions into explicit pe…
-
New AI frameworks improve Alzheimer's diagnosis using MRI and clinical data
Researchers have developed new deep learning frameworks for diagnosing Alzheimer's disease using multimodal data. One study focuses on grounding image-based models with anatomical references and addressing label leakage…
-
AI models struggle with multilingual and meme-based hate speech detection
Researchers are exploring advanced methods to improve AI's ability to detect hate speech, particularly in multilingual and multimodal contexts. One study focuses on training-time explainability to align AI reasoning wit…
-
New BERT-style model enhances EHR analysis with lab data and explainability
Researchers have developed BERT-LER, a novel BERT-style model designed for analyzing structured electronic health records (EHRs). This model effectively encodes laboratory test results as discrete tokens while preservin…
-
2D CNNs improve plant trait retrieval from spectral images
Researchers have developed a new method for plant trait retrieval using hyperspectral spectroscopy by transforming 1D spectral data into 2D images. This approach, utilizing convolutional neural networks (CNNs) like Effi…
-
New method enhances interpretability of microbiome transformer models
Researchers have developed a new method for interpreting the decisions made by transformer models, specifically those used in microbiome analysis like BiomeGPT. This approach, called Signed Integrated Gradients, address…
-
New Spectral Integrated Gradients method improves AI feature attribution
Researchers have introduced Spectral Integrated Gradients (SIG), a novel feature attribution method designed to improve upon existing techniques like Integrated Gradients (IG). SIG addresses the limitations of IG's stan…
-
Interpretable ML predicts traffic congestion impacted by COVID-19
Researchers have developed interpretable machine learning models to predict traffic congestion in Alameda County, California, considering the unique impacts of the COVID-19 pandemic. By incorporating variables related t…
-
New framework fools AI explainability auditors by embedding evasion logic
Researchers have developed a new framework called "Crushing the Evidence" that can fool white-box explainable AI (XAI) auditors. This dual-penalty evasion technique embeds evasion logic directly into model parameters, a…
-
New research tackles LLM sycophancy with benchmarks and attribution methods
Two new research papers address the issue of sycophancy in large language models, where models tend to agree with users even when it contradicts factual information. The first paper, MedPRESS, introduces a multi-turn be…