AI research papers and technical signals from arXiv surfaced by ChAIcked.
Research proposes a quantum-inspired strategy using mixture of experts for image classification, contrasting with classical CNNs that rely on diffusion-based downsampling and block pooling for feature extraction.
Blog · Computer Vision · Neural Networks · Research · — Open article
Study challenges the common practice of using plain MLPs as tabular encoders in multimodal systems, evaluating state-of-the-art tabular models as stronger alternatives for this underexplored modality combination.
Blog · Multimodal Learning · Tabular Data · Research · — Open article
Explains the prevalence of vanishingly small Hessian eigenvalues in neural network loss landscapes through approximate continuous symmetries, offering insight into local geometry governing training dynamics.
Blog · Theory · Optimization · Research · — Open article
A new method uses large language models to generate neural network improvements by constraining generation to same-family architectures, separating transfer learning effects from adaptation within a controlled source-guided protocol.
Blog · Models · Architecture Search · Development · — Open article
FourTune introduces a fully 4-bit quantization approach for post-training diffusion models, reducing memory requirements and training time to make fine-tuning on diverse downstream tasks more practical for resource-constrained settings.
Blog · Models · Optimization · Fine-Tuning · — Open article
Researchers apply bivariate bicycle codes to reduce error overhead in quantum convolutional neural networks, enabling practical execution on near-term quantum devices by lowering the error-correction threshold requirements.
Blog · Models · Quantum Computing · Optimization · — Open article
A parameter-efficient tri-branch neural network combines graph neural networks, deep networks, and SMILES embeddings to capture diverse topological and structural information for more accurate molecular property prediction.
Blog · Models · Data · Development · — Open article
A new defense strategy against data poisoning exploits the observation that models learn poisonous samples more readily than benign ones, using training loss thresholds to identify and remove backdoor triggers without task-specific fine-tuning.
Blog · Models · Security · Development · — Open article
Research applies econometric techniques like Heckman correction to handle selection on unobservables in training data, addressing limitations of standard covariate-shift and importance-weighting approaches for real-world data collection scenarios.
Blog · Data · Development · Methodology · — Open article