Domain Adaptation and Few-Shot Learning

Understanding Performance Plateaus and Representation Collapse in 3D MRI Transformers Trained to Classify Brain Disease

This paper investigates why performance plateaus in 3D MRI transformers for brain disease classification and finds that representation collapse, not task difficulty, causes the plateau.

Discovering Latent Scientific Concepts through Discrete Representation Learning

This paper introduces Kodiak, a discrete representation learning method for scientific imaging that improves representation quality and enables interpretable, reusable concepts across domains.

Supervised Domain Adaptation Mitigates Cross-Ethnicity Prediction Error in Neuroimaging Based Cognitive Prediction

This study uses domain adaptation methods to reduce cross-ethnic prediction errors in neuroimaging-based cognitive prediction using the ABCD dataset.

Biological Continued Pretraining Reshapes the Capability Profile of a Foundation Model Without Catastrophic Forgetting

This paper re-analyzes a foundation model's checkpoints to show biological continued pretraining improves general knowledge, code generation, and reasoning without catastrophic forgetting, but subsequent fine-tuning reverses these gains.

FEATMAP: Targeted Correction of Acquisition Signatures Harmonizes Medical Foundation Model Embeddings and Enables Robust Task Generalization

FEATMAP corrects acquisition signatures in medical foundation model embeddings to improve cross-condition similarity and task generalization without retraining.

Intra-African Geographic Domain Shift in Wildlife Camera Trap Species Classification: A Comparative Study of Supervised and Zero-Shot Foundation Models

This study evaluates how well wildlife camera trap species classification models perform across different African regions, comparing supervised and zero-shot models.

Protocol-Guided Cross-Domain Transfer Learning for Bovine Facial Pain Recognition under Weak Dairy-Farm Labels

This paper evaluates transfer learning for bovine facial pain recognition under weak labels, finding that protocol design and operating-point choices outperform pretraining.

Few-Shot Classification of C. elegans Developmental Stages via Explainable Hierarchical Hyperbolic Graph Embeddings

HyperDev uses hyperbolic geometry to classify C. elegans developmental stages with few labeled examples, achieving high accuracy and biological alignment.

Selecting Synthetic Data for Successful Simulation-Based Transfer Learning in Dynamical Biological Systems

The paper demonstrates how to select synthetic data for transfer learning in dynamical biological systems, showing performance improvements of up to 92% in mean absolute error.

Pretraining Objective Shapes Cross-Category Generalization in Affective Image Prediction: A Geometric Comparison of Vision Transformer Encoders

This paper compares how different pretraining objectives affect the generalization of Vision Transformers in affective image prediction, finding that contrastive learning outperforms classification pretraining in cross-category generalization.

Mistake gating leads to energy and memory efficient continual learning

This paper introduces mistake-gated learning, a method that reduces synaptic updates by focusing on classification errors, improving energy efficiency and memory usage in continual learning.

Learning sculpts orthogonal task manifolds for continual skill learning in recurrent networks

This paper shows that recurrent neural networks can learn new tasks without forgetting old ones by preserving task-specific latent dynamics through orthogonal task manifolds.

Complexity Begets Simplicity: Self-Supervised Learning for Palaeontological Images with Few or No Labels

This paper demonstrates that self-supervised learning can effectively classify paleontological images with few or no labels, achieving results comparable to traditional methods with reduced computational requirements.

Mixture models for domain-adaptive brain decoding

This paper introduces a mixture model framework for domain-adaptive brain decoding, achieving state-of-the-art performance with less training data by optimizing mixture weights via convex optimization.

MACS: Multi-Domain Adaptation Enables Accurate Connectomics Segmentation

MACS improves connectomics segmentation by adapting across multiple domains and using active learning, achieving significant performance gains over existing methods.

Leveraging synthetic data produced from museum specimens to train adaptable species classification models

The paper presents a pipeline using synthetic images from 3D scans of museum specimens to train adaptable species classification models, achieving high accuracy on real photographs.

Representation Transfer via Invariant Input-driven Neural Manifolds for Brain-inspired Computations

The paper introduces a brain-inspired framework for domain adaptation in AI, using recurrent neural networks to create transferable representations that enable rapid few-shot adaptation with fewer parameters.

Parameter Efficient Fine-tuning of Transformer-based Masked Autoencoder Enhances Resource Constrained Neuroimage Analysis

This study evaluates parameter-efficient fine-tuning methods for vision transformers in neuroimaging, showing they outperform full fine-tuning and training from scratch with fewer parameters.

CEL: A Continual Learning Model for Disease Outbreak Prediction by Leveraging Domain Adaptation via Elastic Weight Consolidation

This paper introduces CEL, a continual learning model for disease outbreak prediction that uses domain adaptation via Elastic Weight Consolidation, showing improved performance over existing models.

Hard-wired visual filters for environment-agnostic object recognition

Hard-wired Gabor filters enable environment-agnostic object recognition without overfitting.

Few-shot Algorithms for Consistent Neural Decoding (FALCON) Benchmark

The paper introduces FALCON, a benchmark suite for evaluating robustness of iBCI decoders using standardized datasets and evaluation metrics.

UNSUPERVISED HARMONIZATION OF BRAIN MRI USING 3D CYCLE GANS AND ITS EFFECT ON BRAIN AGE PREDICTION

This paper harmonizes brain MRI data from multiple sources using 3D CycleGANs and shows improved accuracy in brain age prediction compared to raw data.

A biologically inspired architecture with switching units can learn to generalize across backgrounds

The paper introduces a bio-inspired neural network architecture with switching units that improves generalization across different visual backgrounds.

Abstract representations emerge naturally in neural networks trained to perform multiple tasks

This paper shows that training neural networks on multiple tasks leads to the emergence of abstract representations, which support generalization and few-sample learning.

Informing generative replay for continual learning with long-term memory formation in the fruit fly

The paper introduces a computational model of generative replay in fruit fly mushroom bodies to improve continual learning, showing a 20% performance increase on CIFAR-100.

wenda_gpu: fast domain adaptation for genomic data

wenda_gpu provides a fast domain adaptation method for genomic data using GPU acceleration, achieving comparable results to the original wenda while improving cancer mutation prediction on small datasets.

Effective Representation Learning via The Integrated Self-Supervised Pre-training models of StyleGAN2-ADA and DINO for Colonoscopy Images

The paper combines StyleGAN2-ADA and DINO for self-supervised pre-training to learn representations from colonoscopy images without annotated data.

From shallow to deep: exploiting feature-based classifiers for domain adaptation in semantic segmentation

This paper combines CNNs and Random Forests for domain adaptation in image segmentation, using pseudo-labels from a few annotations to improve performance across domains.

Memory Consolidation with Orthogonal Gradients for avoiding Catastrophic Forgetting

The paper proposes a regularization-based model to prevent catastrophic forgetting in deep networks by constraining parameter updates to directions orthogonal to previous task gradients, showing improved performance on standard datasets.

Using top-down modulation to optimally balance shared versus separated task representations

This paper evaluates modulation networks for balancing shared and separated task representations, finding that multiplicative adaptive modulation achieves the best performance.