It is widely acknowledged that the performance of Transformer models is logarithmically related to their number of parameters and computational complexity. While approaches like Mixture of Experts (MoE) decouple parameter count from computational complexity, they still face challenges in inference due to high memory access costs. This work introduces UltraMem, incorporating large-scale, ultra-sparse memory layer to address these limitations. Our approach significantly reduces inference latency while maintaining model performance. We also investigate the scaling laws of this new architecture, demonstrating that it not only exhibits favorable scaling properties but outperforms MoE. In experiments, the largest UltraMem we train has \textbf{20 million} memory slots. The results show that our method achieves state-of-the-art inference speed and model performance within a given computational budget, paving the way for billions of slots or experts.

We present hyper-connections, a simple yet effective method that can serve as an alternative to residual connections. This approach specifically addresses common drawbacks observed in residual connection variants, such as the seesaw effect between gradient vanishing and representation collapse. Theoretically, hyper-connections allow the network to adjust the strength of connections between features at different depths and dynamically rearrange layers. We conduct experiments focusing on the pre-training of large language models, including dense and sparse models, where hyper-connections show significant performance improvements over residual connections. Additional experiments conducted on vision tasks also demonstrate similar improvements. We anticipate that this method will be broadly applicable and beneficial across a wide range of AI problems.

Tokenization is a fundamental component of large language models (LLMs), yet its influence on model scaling and performance is not fully explored. In this paper, we introduce Over-Tokenized Transformers, a novel framework that decouples input and output vocabularies to improve language modeling performance. Specifically, our approach scales up input vocabularies to leverage multi-gram tokens. Through extensive experiments, we uncover a log-linear relationship between input vocabulary size and training loss, demonstrating that larger input vocabularies consistently enhance model performance, regardless of model size. Using a large input vocabulary, we achieve performance comparable to double-sized baselines with no additional cost. Our findings highlight the importance of tokenization in scaling laws and provide practical insight for tokenizer design, paving the way for more efficient and powerful LLMs.

In robotics and autonomous driving, accurate depth estimation is vital yet challenging under dynamic scenes and extreme lighting. Conventional frame-based cameras offer rich context but suffer from motion blur and limited dynamic range, while event cameras provide high temporal resolution and dynamic range but lack global scene structure. Therefore, recent studies explore frame-event fusion depth estimation methods to leverage these two complementary modalities to achieve robust performance. However, due to the mismatch in temporal and spatial resolution, there is an inherent contradiction between high spatial resolution frames captured at sparse temporal intervals and event streams characterized by spatial sparsity but high temporal resolution, rendering cross-modal feature fusion ineffective. Moreover, the limited availability of frame-event depth datasets further undermines the model's generalization capability across different scenes. To address the above challenges, we propose HAFUNet, a Hierarchical Attention Fusion Network for depth estimation via frame-event fusion. Our method contains: (1) a pre-trained Dual-Stream Encoder (DSEer) to extract complementary features from frame and event inputs; (2) a Cross-modal Feature Interaction Module (CFIM) that aligns and fuses spatial-channel features across modalities; and (3) a Hierarchical Attention Decoder (HADer) that progressively refines depth predictions via attention-guided convolution. Experiments on synthetic and real-world datasets show that HAFUNet surpasses existing methods in depth accuracy and robustness. These results demonstrate the strength of our fusion strategy in diverse environments. Code is available at https://github.com/SiYZhangwh/HAFUNet.

Open set recognition (OSR) requires models to classify known samples while detecting unknown samples for real-world applications. Existing studies show impressive progress using unknown samples from auxiliary datasets to regularize OSR models, but they have proved to be sensitive to selecting such known outliers. In this paper, we discuss the aforementioned problem from a new perspective: Can we regularize OSR models without elaborately selecting auxiliary known outliers? We first empirically and theoretically explore the role of foregrounds and backgrounds in open set recognition and disclose that: 1) backgrounds that correlate with foregrounds would mislead the model and cause failures when encounters 'partially' known images; 2) Backgrounds unrelated to foregrounds can serve as auxiliary known outliers and provide regularization via global average pooling. Based on the above insights, we propose a new method, Background Mix (BackMix), that mixes the foreground of an image with different backgrounds to remove the underlying fore-background priors. Specifically, BackMix first estimates the foreground with class activation maps (CAMs), then randomly replaces image patches with backgrounds from other images to obtain mixed images for training. With backgrounds de-correlated from foregrounds, the open set recognition performance is significantly improved. The proposed method is quite simple to implement, requires no extra operation for inferences, and can be seamlessly integrated into almost all of the existing frameworks.

Open set recognition enables deep neural networks (DNNs) to identify samples of unknown classes, while maintaining high classification accuracy on samples of known classes. Existing methods based on auto-encoder (AE) and prototype learning show great potential in handling this challenging task. In this study, we propose a novel method, called Class-Specific Semantic Reconstruction (CSSR), that integrates the power of AE and prototype learning. Specifically, CSSR replaces prototype points with manifolds represented by class-specific AEs. Unlike conventional prototype-based methods, CSSR models each known class on an individual AE manifold, and measures class belongingness through AE's reconstruction error. Class-specific AEs are plugged into the top of the DNN backbone and reconstruct the semantic representations learned by the DNN instead of the raw image. Through end-to-end learning, the DNN and the AEs boost each other to learn both discriminative and representative information. The results of experiments conducted on multiple datasets show that the proposed method achieves outstanding performance in both close and open set recognition and is sufficiently simple and flexible to incorporate into existing frameworks.