p-MoD: Building Mixture-of-Depths MLLMs via Progressive Ratio Decay

Despite the remarkable performance of multimodal large language models (MLLMs) across diverse tasks, the substantial training and inference costs impede their advancement. The majority of computation stems from the overwhelming volume of vision tokens processed by the transformer decoder. In this paper, we propose to build efficient MLLMs by leveraging the Mixture-of-Depths (MoD) mechanism, where each transformer decoder layer selects essential vision tokens to process while skipping redundant ones. However, integrating MoD into MLLMs is non-trivial. To address the challenges of training and inference stability as well as limited training data, we adapt the MoD module with two novel designs: tanh-gated weight normalization (TanhNorm) and symmetric token reweighting (STRing). Moreover, we observe that vision tokens exhibit higher redundancy in deeper layer and thus design a progressive ratio decay (PRD) strategy, which gradually reduces the token retention ratio layer by layer, employing a shifted cosine schedule. This crucial design fully unleashes the potential of MoD, significantly boosting the efficiency and performance of our models. To validate the effectiveness of our approach, we conduct extensive experiments with two baseline models across 14 benchmarks. Our model, p-MoD, matches or even surpasses the performance of the baseline models, with only 55.6% TFLOPs and 53.8% KV cache storage during inference, and 77.7% GPU hours during training.

Synthetic Aperture Radar Image Change Detection via Layer Attention-Based Noise-Tolerant Network

Recently, change detection methods for synthetic aperture radar (SAR) images based on convolutional neural networks (CNN) have gained increasing research attention. However, existing CNN-based methods neglect the interactions among multilayer convolutions, and errors involved in the preclassification restrict the network optimization. To this end, we proposed a layer attention-based noise-tolerant network, termed LANTNet. In particular, we design a layer attention module that adaptively weights the feature of different convolution layers. In addition, we design a noise-tolerant loss function that effectively suppresses the impact of noisy labels. Therefore, the model is insensitive to noisy labels in the preclassification results. The experimental results on three SAR datasets show that the proposed LANTNet performs better compared to several state-of-the-art methods. The source codes are available at https://github.com/summitgao/LANTNet