Finedeep: Mitigating Sparse Activation in Dense LLMs via Multi-Layer Fine-Grained Experts

Large language models have demonstrated exceptional performance across a wide range of tasks. However, dense models usually suffer from sparse activation, where many activation values tend towards zero (i.e., being inactivated). We argue that this could restrict the efficient exploration of model representation space. To mitigate this issue, we propose Finedeep, a deep-layered fine-grained expert architecture for dense models. Our framework partitions the feed-forward neural network layers of traditional dense models into small experts, arranges them across multiple sub-layers. A novel routing mechanism is proposed to determine each expert's contribution. We conduct extensive experiments across various model sizes, demonstrating that our approach significantly outperforms traditional dense architectures in terms of perplexity and benchmark performance while maintaining a comparable number of parameters and floating-point operations. Moreover, we find that Finedeep achieves optimal results when balancing depth and width, specifically by adjusting the number of expert sub-layers and the number of experts per sub-layer. Empirical results confirm that Finedeep effectively alleviates sparse activation and efficiently utilizes representation capacity in dense models.

HeadInfer: Memory-Efficient LLM Inference by Head-wise Offloading

Transformer-based large language models (LLMs) demonstrate impressive performance in long context generation. Extending the context length has disproportionately shifted the memory footprint of LLMs during inference to the key-value cache (KV cache). In this paper, we propose HEADINFER, which offloads the KV cache to CPU RAM while avoiding the need to fully store the KV cache for any transformer layer on the GPU. HEADINFER employs a fine-grained, head-wise offloading strategy, maintaining only selective attention heads KV cache on the GPU while computing attention output dynamically. Through roofline analysis, we demonstrate that HEADINFER maintains computational efficiency while significantly reducing memory footprint. We evaluate HEADINFER on the Llama-3-8B model with a 1-million-token sequence, reducing the GPU memory footprint of the KV cache from 128 GB to 1 GB and the total GPU memory usage from 207 GB to 17 GB, achieving a 92% reduction compared to BF16 baseline inference. Notably, HEADINFER enables 4-million-token inference with an 8B model on a single consumer GPU with 24GB memory (e.g., NVIDIA RTX 4090) without approximation methods.

DSMoE: Matrix-Partitioned Experts with Dynamic Routing for Computation-Efficient Dense LLMs

As large language models continue to scale, computational costs and resource consumption have emerged as significant challenges. While existing sparsification methods like pruning reduce computational overhead, they risk losing model knowledge through parameter removal. This paper proposes DSMoE (Dynamic Sparse Mixture-of-Experts), a novel approach that achieves sparsification by partitioning pre-trained FFN layers into computational blocks. We implement adaptive expert routing using sigmoid activation and straight-through estimators, enabling tokens to flexibly access different aspects of model knowledge based on input complexity. Additionally, we introduce a sparsity loss term to balance performance and computational efficiency. Extensive experiments on LLaMA models demonstrate that under equivalent computational constraints, DSMoE achieves superior performance compared to existing pruning and MoE approaches across language modeling and downstream tasks, particularly excelling in generation tasks. Analysis reveals that DSMoE learns distinctive layerwise activation patterns, providing new insights for future MoE architecture design.

Tensor-GaLore: Memory-Efficient Training via Gradient Tensor Decomposition

We present Tensor-GaLore, a novel method for efficient training of neural networks with higher-order tensor weights. Many models, particularly those used in scientific computing, employ tensor-parameterized layers to capture complex, multidimensional relationships. When scaling these methods to high-resolution problems makes memory usage grow intractably, and matrix based optimization methods lead to suboptimal performance and compression. We propose to work directly in the high-order space of the complex tensor parameter space using a tensor factorization of the gradients during optimization. We showcase its effectiveness on Fourier Neural Operators (FNOs), a class of models crucial for solving partial differential equations (PDE) and prove the theory of it. Across various PDE tasks like the Navier Stokes and Darcy Flow equations, Tensor-GaLore achieves substantial memory savings, reducing optimizer memory usage by up to 75%. These substantial memory savings across AI for science demonstrate Tensor-GaLore's potential.

Towards Combating Frequency Simplicity-biased Learning for Domain Generalization

Domain generalization methods aim to learn transferable knowledge from source domains that can generalize well to unseen target domains. Recent studies show that neural networks frequently suffer from a simplicity-biased learning behavior which leads to over-reliance on specific frequency sets, namely as frequency shortcuts, instead of semantic information, resulting in poor generalization performance. Despite previous data augmentation techniques successfully enhancing generalization performances, they intend to apply more frequency shortcuts, thereby causing hallucinations of generalization improvement. In this paper, we aim to prevent such learning behavior of applying frequency shortcuts from a data-driven perspective. Given the theoretical justification of models' biased learning behavior on different spatial frequency components, which is based on the dataset frequency properties, we argue that the learning behavior on various frequency components could be manipulated by changing the dataset statistical structure in the Fourier domain. Intuitively, as frequency shortcuts are hidden in the dominant and highly dependent frequencies of dataset structure, dynamically perturbating the over-reliance frequency components could prevent the application of frequency shortcuts. To this end, we propose two effective data augmentation modules designed to collaboratively and adaptively adjust the frequency characteristic of the dataset, aiming to dynamically influence the learning behavior of the model and ultimately serving as a strategy to mitigate shortcut learning. Code is available at AdvFrequency (https://github.com/C0notSilly/AdvFrequency).

SynFER: Towards Boosting Facial Expression Recognition with Synthetic Data

Facial expression datasets remain limited in scale due to privacy concerns, the subjectivity of annotations, and the labor-intensive nature of data collection. This limitation poses a significant challenge for developing modern deep learning-based facial expression analysis models, particularly foundation models, that rely on large-scale data for optimal performance. To tackle the overarching and complex challenge, we introduce SynFER (Synthesis of Facial Expressions with Refined Control), a novel framework for synthesizing facial expression image data based on high-level textual descriptions as well as more fine-grained and precise control through facial action units. To ensure the quality and reliability of the synthetic data, we propose a semantic guidance technique to steer the generation process and a pseudo-label generator to help rectify the facial expression labels for the synthetic images. To demonstrate the generation fidelity and the effectiveness of the synthetic data from SynFER, we conduct extensive experiments on representation learning using both synthetic data and real-world data. Experiment results validate the efficacy of the proposed approach and the synthetic data. Notably, our approach achieves a 67.23% classification accuracy on AffectNet when training solely with synthetic data equivalent to the AffectNet training set size, which increases to 69.84% when scaling up to five times the original size. Our code will be made publicly available.

CLIP Multi-modal Hashing for Multimedia Retrieval

Multi-modal hashing methods are widely used in multimedia retrieval, which can fuse multi-source data to generate binary hash code. However, the individual backbone networks have limited feature expression capabilities and are not jointly pre-trained on large-scale unsupervised multi-modal data, resulting in low retrieval accuracy. To address this issue, we propose a novel CLIP Multi-modal Hashing (CLIPMH) method. Our method employs the CLIP framework to extract both text and vision features and then fuses them to generate hash code. Due to enhancement on each modal feature, our method has great improvement in the retrieval performance of multi-modal hashing methods. Compared with state-of-the-art unsupervised and supervised multi-modal hashing methods, experiments reveal that the proposed CLIPMH can significantly improve performance (a maximum increase of 8.38% in mAP).

Optimizing Placement and Power Allocation in Reconfigurable Intelligent Sensing Surfaces for Enhanced Sensing and Communication Performance

In this letter, we investigate the design of multiple reconfigurable intelligent sensing surfaces (RISSs) that enhance both communication and sensing tasks. An RISS incorporates additional active elements tailored to improve sensing accuracy. Our initial task involves optimizing placement of RISSs to mitigate signal interference. Subsequently, we establish power allocation schemes for sensing and communication within the system. Our final consideration involves examining how sensing results can be utilized to enhance communication, alongside an evaluation of communication performance under the impact of sensing inaccuracies. Numerical results reveal that the sensing task reaches its optimal performance with a finite number of RISSs, while the communication task exhibits enhanced performance with an increasing number of RISSs. Additionally, we identify an optimal communication spot under user movement.

Multi-SIGATnet: A multimodal schizophrenia MRI classification algorithm using sparse interaction mechanisms and graph attention networks

Schizophrenia is a serious psychiatric disorder. Its pathogenesis is not completely clear, making it difficult to treat patients precisely. Because of the complicated non-Euclidean network structure of the human brain, learning critical information from brain networks remains difficult. To effectively capture the topological information of brain neural networks, a novel multimodal graph attention network based on sparse interaction mechanism (Multi-SIGATnet) was proposed for SZ classification was proposed for SZ classification. Firstly, structural and functional information were fused into multimodal data to obtain more comprehensive and abundant features for patients with SZ. Subsequently, a sparse interaction mechanism was proposed to effectively extract salient features and enhance the feature representation capability. By enhancing the strong connections and weakening the weak connections between feature information based on an asymmetric convolutional network, high-order interactive features were captured. Moreover, sparse learning strategies were designed to filter out redundant connections to improve model performance. Finally, local and global features were updated in accordance with the topological features and connection weight constraints of the higher-order brain network, the features being projected to the classification target space for disorder classification. The effectiveness of the model is verified on the Center for Biomedical Research Excellence (COBRE) and University of California Los Angeles (UCLA) datasets, achieving 81.9\% and 75.8\% average accuracy, respectively, 4.6\% and 5.5\% higher than the graph attention network (GAT) method. Experiments showed that the Multi-SIGATnet method exhibited good performance in identifying SZ.

MINI-SEQUENCE TRANSFORMER: Optimizing Intermediate Memory for Long Sequences Training

We introduce Mini-Sequence Transformer (MsT), a simple and effective methodology for highly efficient and accurate LLM training with extremely long sequences. MsT partitions input sequences and iteratively processes mini-sequences to reduce intermediate memory usage. Integrated with activation recomputation, it enables significant memory savings in both forward and backward passes. In experiments with the Llama3-8B model, with MsT, we measure no degradation in throughput or convergence even with 12x longer sequences than standard implementations due to our careful memory optimizations. MsT is fully general, implementation-agnostic, and requires minimal code changes to integrate with existing LLM training frameworks.