The internal representations learned by deep networks are often sensitive to architecture-specific choices, raising questions about the stability, alignment, and transferability of learned structure across models. In this paper, we investigate how structural constraints--such as linear shaping operators and corrective paths--affect the compatibility of internal representations across different architectures. Building on the insights from prior studies on structured transformations and convergence, we develop a framework for measuring and analyzing representational alignment across networks with distinct but related architectural priors. Through a combination of theoretical insights, empirical probes, and controlled transfer experiments, we demonstrate that structural regularities induce representational geometry that is more stable under architectural variation. This suggests that certain forms of inductive bias not only support generalization within a model, but also improve the interoperability of learned features across models. We conclude with a discussion on the implications of representational transferability for model distillation, modular learning, and the principled design of robust learning systems.

Massive activations are scalar values in transformer hidden states that achieve values orders of magnitude larger than typical activations and have been shown to be critical for model functionality. While prior work has characterized these phenomena in fully trained models, the temporal dynamics of their emergence during training remain poorly understood. We present the first comprehensive analysis of massive activation development throughout transformer training, using the Pythia model family as our testbed. Through systematic analysis of various model sizes across multiple training checkpoints, we demonstrate that massive activation emergence follows predictable mathematical patterns that can be accurately modeled using an exponentially-modulated logarithmic function with five key parameters. We develop a machine learning framework to predict these mathematical parameters from architectural specifications alone, achieving high accuracy for steady-state behavior and moderate accuracy for emergence timing and magnitude. These findings enable architects to predict and potentially control key aspects of massive activation emergence through design choices, with significant implications for model stability, training cycle length, interpretability, and optimization. Our findings demonstrate that the emergence of massive activations is governed by model design and can be anticipated, and potentially controlled, before training begins.

Recent advancements in object detection rely on modular architectures with multi-scale fusion and attention mechanisms. However, static fusion heuristics and class-agnostic attention limit performance in dynamic scenes with occlusions, clutter, and class imbalance. We introduce Dynamic Class-Aware Fusion Network (DyCAF-Net) that addresses these challenges through three innovations: (1) an input-conditioned equilibrium-based neck that iteratively refines multi-scale features via implicit fixed-point modeling, (2) a dual dynamic attention mechanism that adaptively recalibrates channel and spatial responses using input- and class-dependent cues, and (3) class-aware feature adaptation that modulates features to prioritize discriminative regions for rare classes. Through comprehensive ablation studies with YOLOv8 and related architectures, alongside benchmarking against nine state-of-the-art baselines, DyCAF-Net achieves significant improvements in precision, mAP@50, and mAP@50-95 across 13 diverse benchmarks, including occlusion-heavy and long-tailed datasets. The framework maintains computational efficiency ($\sim$11.1M parameters) and competitive inference speeds, while its adaptability to scale variance, semantic overlaps, and class imbalance positions it as a robust solution for real-world detection tasks in medical imaging, surveillance, and autonomous systems.

Text embeddings, i.e. vector representations of entire texts, play an important role in many NLP applications, such as retrieval-augmented generation, sentiment analysis, clustering, or visualizing collections of texts for data exploration. Currently, top-performing embedding models are derived from pre-trained language models via extensive supervised fine-tuning using curated text pairs. This contrasts with computer vision, where self-supervised training based on data augmentations has demonstrated remarkable success. Here we systematically compare the two most well-known augmentation strategies for positive pair generation in contrastive learning of text embeddings. We assess embedding quality on MTEB and additional in-domain evaluations and show that cropping augmentation strongly outperforms the dropout-based approach. We find that on out-of-domain data, the quality of resulting embeddings is below the supervised SOTA models, but for in-domain data, self-supervised fine-tuning produces high-quality text embeddings after very short fine-tuning, sometimes only marginally below the supervised SOTA. Finally, we show that representation quality increases towards the last transformer layers, which undergo the largest change during fine-tuning; and that fine-tuning only those last layers is sufficient to reach similar embedding quality.

The hydrogen trapping behaviour of metallic alloys is generally characterised using Thermal Desorption Spectroscopy (TDS). However, as an indirect method, extracting key parameters (trap binding energies and densities) remains a significant challenge. To address these limitations, this work introduces a machine learning-based scheme for parameter identification from TDS spectra. A multi-Neural Network (NN) model is developed and trained exclusively on synthetic data to predict trapping parameters directly from experimental data. The model comprises two multi-layer, fully connected, feed-forward NNs trained with backpropagation. The first network (classification model) predicts the number of distinct trap types. The second network (regression model) then predicts the corresponding trap densities and binding energies. The NN architectures, hyperparameters, and data pre-processing were optimised to minimise the amount of training data. The proposed model demonstrated strong predictive capabilities when applied to three tempered martensitic steels of different compositions. The code developed is freely provided.

Softmax with the cross entropy loss is the standard configuration for current neural classification models. The gold score for a target class is supposed to be 1, but it is never reachable under the softmax schema. Such a problem makes the training process continue forever and leads to overfitting. Moreover, the "target-approach-1" training goal forces the model to continuously learn all samples, leading to a waste of time in handling some samples which have already been classified correctly with high confidence, while the test goal simply requires the target class of each sample to hold the maximum score. To solve the above weaknesses, we propose the Adaptive Sparse softmax (AS-Softmax) which designs a reasonable and test-matching transformation on top of softmax. For more purposeful learning, we discard the classes with far smaller scores compared with the actual class during training. Then the model could focus on learning to distinguish the target class from its strong opponents, which is also the great challenge in test. In addition, since the training losses of easy samples will gradually drop to 0 in AS-Softmax, we develop an adaptive gradient accumulation strategy based on the masked sample ratio to speed up training. We verify the proposed AS-Softmax on a variety of text multi-class, text multi-label, text token classification, image classification and audio classification tasks with class sizes ranging from 5 to 5000+. The results show that AS-Softmax consistently outperforms softmax and its variants, and the loss of AS-Softmax is remarkably correlated with classification performance in validation. Furthermore, adaptive gradient accumulation strategy can bring about 1.2x training speedup comparing with the standard softmax while maintaining classification effectiveness.

House price valuation remains challenging due to localized market variations. Existing approaches often rely on black-box machine learning models, which lack interpretability, or simplistic methods like linear regression (LR), which fail to capture market heterogeneity. To address this, we propose a machine learning approach that applies two-stage clustering, first grouping properties based on minimal location-based features before incorporating additional features. Each cluster is then modeled using either LR or a generalized additive model (GAM), balancing predictive performance with interpretability. Constructing and evaluating our models on 43,309 German house property listings from 2023, we achieve a 36% improvement for the GAM and 58% for LR in mean absolute error compared to models without clustering. Additionally, graphical analyses unveil pattern shifts between clusters. These findings emphasize the importance of cluster-specific insights, enhancing interpretability and offering practical value for buyers, sellers, and real estate analysts seeking more reliable property valuations.

Out-of-distribution (OOD) detection lies at the heart of robust artificial intelligence (AI), aiming to identify samples from novel distributions beyond the training set. Recent approaches have exploited feature representations as distinguishing signatures for OOD detection. However, most existing methods rely on restrictive assumptions on the feature space that limit the separability between in-distribution (ID) and OOD samples. In this work, we propose a novel OOD detection framework based on a pseudo-label-induced subspace representation, that works under more relaxed and natural assumptions compared to existing feature-based techniques. In addition, we introduce a simple yet effective learning criterion that integrates a cross-entropy-based ID classification loss with a subspace distance-based regularization loss to enhance ID-OOD separability. Extensive experiments validate the effectiveness of our framework.

Decentralized Federated Learning (DFL) struggles with the slow adaptation of late-joining delayed clients and high communication costs in asynchronous environments. These limitations significantly hinder overall performance. To address this, we propose DFedCAD, a novel framework for rapid adaptation via Centroid-Aligned Distillation. DFedCAD first employs Weighted Cluster Pruning (WCP) to compress models into representative centroids, drastically reducing communication overhead. It then enables delayed clients to intelligently weigh and align with peer knowledge using a novel structural distance metric and a differentiable k-means distillation module, facilitating efficient end-to-end knowledge transfer. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet show that DFedCAD consistently achieves state-of-the-art performance, attaining the highest accuracy across all evaluated settings while reducing communication overhead by over 86%. Our framework provides a scalable and practical solution for efficient decentralized learning in dynamic, real-world scenarios.

We propose a novel perspective on varied-density clustering for high-dimensional data by framing it as a label propagation process in neighborhood graphs that adapt to local density variations. Our method formally connects density-based clustering with graph connectivity, enabling the use of efficient graph propagation techniques developed in network science. To ensure scalability, we introduce a density-aware neighborhood propagation algorithm and leverage advanced random projection methods to construct approximate neighborhood graphs. Our approach significantly reduces computational cost while preserving clustering quality. Empirically, it scales to datasets with millions of points in minutes and achieves competitive accuracy compared to existing baselines.