Research

Help evidential models keep learning

Evidential networks can estimate fine-grained uncertainty, but some activation functions give almost no learning signal for examples assigned little evidence. We analyzed how this happens across different activations and designed a family of activations and matching regularizers that lets the model continue updating its evidence.

This extends the ICML 2023 result beyond one training setting. Experiments cover image classification, few-shot recognition, and blind face restoration.

More reliable adaptation of vision models

Parameter-efficient fine-tuning can adapt a vision foundation model with very few examples, yet the resulting predictions may be too uncertain. Bayesian-PEFT strengthens the model's prior belief using knowledge acquired during pretraining and adds a diverse evidential ensemble to represent what remains uncertain.

Across few-shot settings and fine-tuning methods, the paper evaluates both predictive accuracy and how well confidence matches outcomes.

Learn from every training example

We traced the weak performance of some evidential classifiers to a specific training failure: an activation can map an example into a zero-evidence region, leaving the model unable to learn from its label. Our theoretical analysis led to a regularizer that restores a learning signal for those examples.

Experiments on challenging datasets show that an uncertainty-aware model need not discard useful training data to produce evidence-based predictions.

Separate kinds of uncertainty in few-shot prediction

With only a few observations, a model needs to distinguish what it has not learned from noise that is inherent in the data. Evidential Conditional Neural Processes replace the usual Gaussian prediction of a conditional neural process with a hierarchical Bayesian structure that separates those two sources of uncertainty.

The analysis explains the connection to standard conditional neural processes and the model's robustness to noisy training tasks; experiments test it on few-shot regression problems.

Choose the most useful few-shot tasks

Meta-learning often trains on randomly chosen tasks, even though labeling and processing every task is costly. We developed a multidimensional belief measure that estimates a task's uncertainty and uses it to select informative tasks for training.

A theoretical link between conflicting and incorrect belief makes this selection possible; a multi-query task design further improves label and compute efficiency in few-shot image classification.

Learn physical laws from sparse simulations

Equations of state must predict energy and pressure without contradicting thermodynamics. The team built a physics-informed model that learns an underlying Helmholtz free-energy function from simulated energy and pressure, so its predictions are linked by the same physical structure.

On warm dense polystyrene, the reported fit is within 1% relative error for both quantities and satisfies the Maxwell and Gibbs–Duhem relations.

Predict energy and pressure consistently

Equation-of-state tables cover wide ranges of physical conditions, but simulations are expensive and training samples are sparse. This work combines a meta-learning style deep regressor, a kernel-based uncertainty regularizer, and an ensemble to improve the stability of energy and pressure estimates.

The study verifies thermodynamic consistency with automatic differentiation, testing whether the learned table remains useful beyond an accurate numerical fit.

Discover patterns in behavior over time

Behavioral signals can be spread across time and different kinds of interaction. Deep Temporal Sets represent their temporal relationships, while Evidential Reinforced Attentions use uncertainty to explore which moments may be informative.

Using visual and touch behavior from child-computer interactions, the study distinguishes children with autism spectrum disorder from typically developing children and identifies characteristic patterns in the data.

Find the informative part of a long time series

A long vessel trajectory may contain only a short period that reveals its type. The proposed uncertainty-aware multiple-instance framework searches for that discriminative period instead of treating every observation as equally useful.

Its uncertainty estimate also flags less trustworthy trajectory predictions and guides fusion with synthetic-aperture radar imagery when that additional information is available.

Belief weighted average computation for classification problems in video applications

International patent application on belief-weighted computation for classification problems in video applications.