paper / deviparikh / May 30
Mischaikow and Parikh address a fundamental limitation in data-driven dynamical systems analysis: finite data cannot exactly determine an underlying vector field, making classical bifurcation characterization provably impossible in general. Their approach sidesteps vector field reconstruction entirely, instead using the homological Conley index over a parameter space to identify the simplest dynamics consistent with observed data — an application of Occam's razor to bifurcation theory. The central result shows that for phase spaces of dimension ≥ 6, any vector field admitting an isolating block whose Conley index is consistent with a saddle-node bifurcation can be smoothly deformed into a canonical model with exactly one structurally stable saddle-node bifurcation, without altering the vector field outside the isolating block.
dynamical-systemsbifurcation-theorysaddle-node-bifurcationconley-indexdata-driven-dynamicstopologyordinary-differential-equations
“It is impossible, in general, to correctly characterize the underlying bifurcation dynamics from a finite dataset generated by an unknown ODE.”
paper / deviparikh / May 30
Label bias — group-conditional annotation errors that cause systematic performance disparities across demographic subgroups — is underexplored in image segmentation because detection typically requires clean, unbiased ground truth that is rarely available. This paper adapts Confident Learning to the segmentation domain, enabling label bias detection directly from biased training data by comparing provided labels against a model's confident predictions to isolate directional errors. A key finding is that label bias leaves a detectable signature in the encoder's feature space by influencing subgroup separability, which the authors exploit for mitigation rather than suppression. The framework is validated across synthetic and real-world datasets, achieving equitable performance without ever requiring clean annotations.
algorithmic-fairnessimage-segmentationlabel-biascomputer-visionbias-mitigationdata-centric-aidemographic-subgroups
“Standard overlap metrics like Dice fail to detect or quantify label bias in segmentation datasets.”
paper / deviparikh / May 30
Vision SSMs (e.g., Mamba-based) suffer from a fundamental limitation: flattening 2D images into 1D token sequences loses local relational structure before global state-space mixing. GraphScan addresses this by inserting a graph-based preprocessing step — constructing spatially bounded local graphs per token, learning feature-conditioned affinities with relative positional bias, and performing one-step message passing to semantically ground tokens before SSM processing. This reframes scanning not as geometric serialization but as learned local semantic routing, achieving state-of-the-art Vision SSM performance across classification, detection, and segmentation benchmarks with linear complexity preserved.
vision-transformersstate-space-modelsgraph-neural-networkscomputer-visionimage-segmentationdynamic-scanningmamba
“Existing Vision SSM scan operators primarily adapt paths or sampling locations rather than explicitly modeling which local patches should exchange information before global state-space mixing.”
paper / deviparikh / May 30
Current video-language models generate factually accurate descriptions but offer no control over tone, urgency, or style — a critical gap in communication-sensitive deployments like traffic management or emergency response. RoadTones addresses this with a full pipeline: a 51K human-validated dataset with tonal annotations, a Chain-of-Thought video-to-text model conditioned on tone, and a paired evaluation suite measuring both factual consistency and tone adherence. The system was further validated through a user study confirming caption quality across all three axes.
video-language-modelstext-generationtone-controlcomputer-visionvideo-captioningchain-of-thoughtdataset
“Existing video-language models cannot control the tone, urgency, or style of generated road event descriptions.”