transfer-learning

Tweet highlights the 4th edition of the classic AI textbook 'Artificial Intelligence: A Modern Approach,' covering deep learning, transfer learning, multi-agent systems, robotics, and causality, with a link to Amazon.

Presents NEXUS, a lightweight foundation model with ~3M parameters pre-trained on LHC collision data, demonstrating improved downstream performance and cross-domain transfer to gravitational waves, flood forecasting, and neural activity.

This paper investigates whether temporally drifting data streams can be partitioned into discrete regimes by fitting a hidden Markov model to the trajectory of neural network weights trained on successive time windows, showing that recovered latent states correlate with transfer performance across two datasets.

AgentGFM proposes a graph foundation model where each node acts as an agent with a shared trainable policy, enabling adaptive information-flow control through a predict–act–observe–correct process. It achieves strong performance across node-level, graph-level, and large-scale transfer tasks.

This paper proposes a diagnosis and optimization framework for cross-domain and cross-lingual generalization in offensive language detection, decomposing performance degradation into dataset and language effects, and quantifying the trade-off between multilingual capability and source-task performance.

Introduces Building2Building (B2B), a large-scale benchmark for studying generalization and transfer in reinforcement learning using realistic HVAC control environments built on EnergyPlus, compatible with Gymnasium.

Inertia-1 is a research project that systematically explores the full lifecycle of motion models—data, sensing, objectives, and scale—to produce a unified representation that transfers across body placements, devices, and tasks without retraining, leveraging self-supervised pretraining on 18 million hours of accelerometry data.

Proposes PEACE, a knowledge-guided framework for transferring adult ECG interpretation to pediatric populations using label-conditioned contrastive alignment, achieving significant improvements under limited supervision.

This paper introduces a framework for when transfer should be expected in continual learning and proposes Transfer-Selective Replay (TSR), which selects replay data predicted to benefit the incoming task rather than indiscriminately replaying past examples. TSR improves forward transfer while maintaining stability, outperforming existing replay baselines.

Proposes a decoupled training strategy that adapts normalization layers and uses precomputed features to reduce overhead in transfer learning, achieving competitive accuracy with significantly reduced training time and energy consumption.

This paper proposes two hybrid continual learning methods—Replay Augmented Elastic Weight Consolidation and Constraint Guided Knowledge Distillation—to adapt pretrained speech models for identifying low-resource Australian Aboriginal languages while mitigating catastrophic forgetting.

This paper investigates cross-lingual transfer for low-resource ASR in Warlpiri, proposing a similarity-based framework combining acoustic and linguistic features to select optimal source languages. Experiments show that acoustically similar languages like Assamese and Hindi significantly reduce word and character error rates.

This paper identifies two structural mechanisms causing multi-task Gaussian processes to misestimate cross-task correlation in Bayesian optimization transfer learning, even for affinely related tasks. The authors propose three conservative remedies to mitigate these issues.

This article delves into transfer learning from the perspective of category theory, proposing deep manifold theory. It argues that neural networks learn relational structures through attribute-free numerical computation, thereby enabling cross-domain transfer, and explains the internal logic of classification.

StepFM is a foundation model that uses only step counter data for broad-spectrum health prediction, offering a privacy-preserving and scalable alternative to high-frequency sensor models.

LEMUR 2 introduces a large-scale dataset of over 14,000 neural network architectures and 750,000 training records across multimodal tasks, supporting NAS, AutoML, and deployment analysis.

This paper proposes a source-guided protocol where an LLM generates candidate modifications for a weak target model using a stronger same-family source model, showing substantial accuracy improvements on CIFAR-10 and SVHN benchmarks while disentangling transfer from adaptation effects.

This research investigates cross-lingual transfer learning from Sinhala to Dhivehi for automatic speech recognition, achieving significant improvements in word error rate compared to Dhivehi-only baselines.

This paper investigates whether explicit domain adaptation methods are beneficial for sentiment transfer when using frozen pre-trained language model backbones, finding that effectiveness depends on whether the backbone already possesses target-domain knowledge.

This study evaluates nine ECG foundation models for Brugada syndrome detection, finding that pre-training provides optimization stability but not transferable clinical knowledge, challenging assumptions about the benefits of large-scale pre-training for rare diseases.