arXiv:2505.21749v2 Announce Type: replace 
Abstract: The role of hidden units in recurrent neural networks is typically seen as modeling memory, with research focusing on enhancing information retention through gating mechanisms. A less explored perspective views hidden units as active participants in the computation performed by the network, rather than passive memory stores. In this work, we revisit bilinear operations, which involve multiplicative interactions between hidden units and input embeddings. We demonstrate theoretically and empirically that they constitute a natural inductive bias for representing the evolution of hidden states in state tracking tasks. These are the simplest type of tasks that require hidden units to actively contribute to the behavior of the network. We also show that bilinear state updates form a natural hierarchy corresponding to state tracking tasks of increasing complexity, with popular linear recurrent networks such as Mamba residing at the lowest-complexity center of that hierarchy.

تستعرض دراسة حديثة دور الوحدات المخفية في الشبكات العصبية المتكررة، مشيرة إلى أنها تشارك بنشاط في العمليات الحسابية بدلاً من أن تكون مجرد مخازن للذاكرة. هذه النظرة تغير التركيز من آليات التحكم التقليدية التي تهدف إلى تعزيز الاحتفاظ بالمعلومات إلى استكشاف العمليات الثنائية التي تتضمن تفاعلات مضاعفة بين الوحدات المخفية والمدخلات. هذه الأبحاث مهمة لأنها قد تؤدي إلى رؤى جديدة وتحسينات في تصميم ووظائف الشبكات العصبية، مما قد يعزز أدائها في تطبيقات متنوعة.

Un estudio reciente revisita el papel de las unidades ocultas en las redes neuronales recurrentes, sugiriendo que participan activamente en los cálculos en lugar de ser solo almacenes de memoria. Esta perspectiva cambia el enfoque de los mecanismos de compuerta tradicionales, que buscan mejorar la retención de información, hacia la exploración de operaciones bilineales que implican interacciones multiplicativas entre las unidades ocultas y las entradas. Esta investigación es significativa, ya que podría conducir a nuevos conocimientos y mejoras en el diseño y la funcionalidad de las redes neuronales, potencialmente mejorando su rendimiento en diversas aplicaciones.

Une étude récente réexamine le rôle des unités cachées dans les réseaux de neurones récurrents, suggérant qu'elles participent activement aux calculs plutôt que de servir uniquement de mémoires. Cette perspective déplace l'accent des mécanismes de contrôle traditionnels visant à améliorer la rétention d'informations vers l'exploration des opérations bilinéaires impliquant des interactions multiplicatives entre les unités cachées et les entrées. Cette recherche est significative car elle pourrait conduire à de nouvelles perspectives et améliorations dans la conception et la fonctionnalité des réseaux de neurones, améliorant potentiellement leurs performances dans diverses applications.

A recent study revisits the role of hidden units in recurrent neural networks, suggesting they actively participate in computations rather than merely serving as memory stores. This perspective shifts the focus from traditional gating mechanisms aimed at enhancing information retention to exploring bilinear operations that involve multiplicative interactions between hidden units and inputs. This research is significant as it could lead to new insights and improvements in the design and functionality of neural networks, potentially enhancing their performance in various applications.

Revisiting Bi-Linear State Transitions in Recurrent Neural Networks

arXiv:2512.04943v1 Announce Type: new 
Abstract: This study introduces a pioneering methodology for human action recognition by harnessing deep neural network techniques and adaptive fusion strategies across multiple modalities, including RGB, optical flows, audio, and depth information. Employing gating mechanisms for multimodal fusion, we aim to surpass limitations inherent in traditional unimodal recognition methods while exploring novel possibilities for diverse applications. Through an exhaustive investigation of gating mechanisms and adaptive weighting-based fusion architectures, our methodology enables the selective integration of relevant information from various modalities, thereby bolstering both accuracy and robustness in action recognition tasks. We meticulously examine various gated fusion strategies to pinpoint the most effective approach for multimodal action recognition, showcasing its superiority over conventional unimodal methods. Gating mechanisms facilitate the extraction of pivotal features, resulting in a more holistic representation of actions and substantial enhancements in recognition performance. Our evaluations across human action recognition, violence action detection, and multiple self-supervised learning tasks on benchmark datasets demonstrate promising advancements in accuracy. The significance of this research lies in its potential to revolutionize action recognition systems across diverse fields. The fusion of multimodal information promises sophisticated applications in surveillance and human-computer interaction, especially in contexts related to active assisted living.

تم تقديم منهجية جديدة للتعرف على الأفعال البشرية، تستفيد من الشبكات العصبية العميقة واستراتيجيات الدمج التكيفية عبر عدة أنماط مثل RGB، وتدفقات الضوء، والصوت، ومعلومات العمق. تستخدم هذه الطريقة آليات البوابة لتعزيز دمج البيانات ذات الصلة، بهدف تحسين الدقة والموثوقية في التعرف على الأفعال البشرية.

Se ha introducido una nueva metodología para el reconocimiento de acciones humanas, aprovechando redes neuronales profundas y estrategias de fusión adaptativas a través de múltiples modalidades como RGB, flujos ópticos, audio e información de profundidad. Este enfoque utiliza mecanismos de compuerta para mejorar la integración de datos relevantes, con el objetivo de mejorar la precisión y la robustez en el reconocimiento de acciones humanas.

Une nouvelle méthodologie pour la reconnaissance des actions humaines a été introduite, tirant parti des réseaux de neurones profonds et des stratégies de fusion adaptatives à travers plusieurs modalités telles que RGB, les flux optiques, l'audio et les informations de profondeur. Cette approche utilise des mécanismes de verrouillage pour améliorer l'intégration des données pertinentes, visant à améliorer la précision et la robustesse dans la reconnaissance des actions humaines.

A new methodology for human action recognition has been introduced, leveraging deep neural networks and adaptive fusion strategies across multiple modalities such as RGB, optical flows, audio, and depth information. This approach utilizes gating mechanisms to enhance the integration of relevant data, aiming to improve accuracy and robustness in recognizing human actions.

Towards Adaptive Fusion of Multimodal Deep Networks for Human Action Recognition

arXiv:2512.04189v1 Announce Type: new 
Abstract: Binary Neural Networks (BNNs), which constrain both weights and activations to binary values, offer substantial reductions in computational complexity, memory footprint, and energy consumption. These advantages make them particularly well suited for deployment on resource-constrained devices. However, training BNNs via gradient-based optimization remains challenging due to the discrete nature of their variables. The dominant approach, quantization-aware training, circumvents this issue by employing surrogate gradients. Yet, this method requires maintaining latent full-precision parameters and performing the backward pass with floating-point arithmetic, thereby forfeiting the efficiency of binary operations during training. While alternative approaches based on local learning rules exist, they are unsuitable for global credit assignment and for back-propagating errors in multi-layer architectures. This paper introduces Binary Error Propagation (BEP), the first learning algorithm to establish a principled, discrete analog of the backpropagation chain rule. This mechanism enables error signals, represented as binary vectors, to be propagated backward through multiple layers of a neural network. BEP operates entirely on binary variables, with all forward and backward computations performed using only bitwise operations. Crucially, this makes BEP the first solution to enable end-to-end binary training for recurrent neural network architectures. We validate the effectiveness of BEP on both multi-layer perceptrons and recurrent neural networks, demonstrating gains of up to +6.89% and +10.57% in test accuracy, respectively. The proposed algorithm is released as an open-source repository.

يقدم مقال جديد خوارزمية انتشار الخطأ الثنائي (BEP) المصممة لتحسين تدريب الشبكات العصبية الثنائية (BNNs)، التي تتميز بأوزانها وتفعيلات ثنائية. تتناول هذه الخوارزمية التحديات المتعلقة بالتحسين القائم على التدرجات في الشبكات العصبية الثنائية، التي تعتمد تقليديًا على طرق التدريب الواعية بالتكميم والتي تؤثر على الكفاءة خلال عملية التدريب.

Un nuevo artículo presenta el algoritmo de Propagación de Errores Binarios (BEP) diseñado para mejorar el entrenamiento de las Redes Neuronales Binarias (BNNs), que se caracterizan por sus pesos y activaciones binarias. Este algoritmo aborda los desafíos de la optimización basada en gradientes en las BNNs, que tradicionalmente dependen de métodos de entrenamiento conscientes de la cuantización que comprometen la eficiencia durante el proceso de entrenamiento.

Un nouvel article présente l'algorithme de propagation d'erreur binaire (BEP) conçu pour améliorer l'entraînement des réseaux neuronaux binaires (BNNs), caractérisés par leurs poids et activations binaires. Cet algorithme aborde les défis de l'optimisation basée sur le gradient dans les BNNs, qui s'appuient traditionnellement sur des méthodes d'entraînement conscientes de la quantification qui compromettent l'efficacité pendant le processus d'entraînement.

A new paper presents the Binary Error Propagation (BEP) algorithm designed to enhance the training of Binary Neural Networks (BNNs), which are characterized by their binary weights and activations. This algorithm addresses the challenges of gradient-based optimization in BNNs, which traditionally rely on quantization-aware training methods that compromise efficiency during the training process.

Revisiting Bi-Linear State Transitions in Recurrent Neural Networks

Was this article worth reading? Share it

LucidQuery AI

Humanize AI

The Visualizer