Abstract:Mixture-of-Experts (MOE) has garnered significant attention for their ability to scale up neural networks while utilizing the same or even fewer active parameters. However, MoE does not relieve the massive memory requirements of networks, which limits their practicality in real-world applications, especially in the era of large language models (LLMs). While recent work explores the possibility of removing entire layers of MoE to reduce memory, the performance degradation is still notable. In this paper, we propose Condense-MoE (CD-MoE} that, instead of dropping the entire MoE layer, condenses the big, sparse MoE layer into a small but dense layer with only a few experts that are activated for all tokens. Our approach is specifically designed for fine-grained MoE with shared experts, where Feed-Forward Networks are split into many small experts, with certain experts isolated to serve as shared experts that are always activated. We demonstrate the effectiveness of our method across multiple MoE models such as DeepSeekMoE and QwenMoE on various benchmarks. Specifically, for the DeepSeekMoE-16B model, our approach maintains nearly 90% of the average accuracy while reducing memory usage by 30% and enhancing inference speed by 30%. Moreover, we show that with lightweight expert fine-tuning, the pruned model can achieve further improvements on specific tasks. Our code are available at https://github.com/duterscmy/CD-MoE/tree/main.
Abstract:In recent years, with the significant evolution of multi-modal large models, many recommender researchers realized the potential of multi-modal information for user interest modeling. In industry, a wide-used modeling architecture is a cascading paradigm: (1) first pre-training a multi-modal model to provide omnipotent representations for downstream services; (2) The downstream recommendation model takes the multi-modal representation as additional input to fit real user-item behaviours. Although such paradigm achieves remarkable improvements, however, there still exist two problems that limit model performance: (1) Representation Unmatching: The pre-trained multi-modal model is always supervised by the classic NLP/CV tasks, while the recommendation models are supervised by real user-item interaction. As a result, the two fundamentally different tasks' goals were relatively separate, and there was a lack of consistent objective on their representations; (2) Representation Unlearning: The generated multi-modal representations are always stored in cache store and serve as extra fixed input of recommendation model, thus could not be updated by recommendation model gradient, further unfriendly for downstream training. Inspired by the two difficulties challenges in downstream tasks usage, we introduce a quantitative multi-modal framework to customize the specialized and trainable multi-modal information for different downstream models.
Abstract:Aligning large language models (LLMs) with human intent is critical for enhancing their performance across a variety of tasks. Standard alignment techniques, such as Direct Preference Optimization (DPO), often rely on the binary Bradley-Terry (BT) model, which can struggle to capture the complexities of human preferences -- particularly in the presence of noisy or inconsistent labels and frequent ties. To address these limitations, we introduce the Tie-rank Oriented Bradley-Terry model (TOBT), an extension of the BT model that explicitly incorporates ties, enabling more nuanced preference representation. Building on this, we propose Tie-rank Oriented Direct Preference Optimization (TODO), a novel alignment algorithm that leverages TOBT's ternary ranking system to improve preference alignment. In evaluations on Mistral-7B and Llama 3-8B models, TODO consistently outperforms DPO in modeling preferences across both in-distribution and out-of-distribution datasets. Additional assessments using MT Bench and benchmarks such as Piqa, ARC-c, and MMLU further demonstrate TODO's superior alignment performance. Notably, TODO also shows strong results in binary preference alignment, highlighting its versatility and potential for broader integration into LLM alignment. The implementation details can be found in https://github.com/XXares/TODO.
Abstract:This paper addresses the problem of vision-based pedestrian localization, which estimates a pedestrian's location using images and camera parameters. In practice, however, calibrated camera parameters often deviate from the ground truth, leading to inaccuracies in localization. To address this issue, we propose an anchor-based method that leverages fixed-position anchors to reduce the impact of camera parameter errors. We provide a theoretical analysis that demonstrates the robustness of our approach. Experiments conducted on simulated, real-world, and public datasets show that our method significantly improves localization accuracy and remains resilient to noise in camera parameters, compared to methods without anchors.
Abstract:Many questions in science center around the fundamental problem of understanding causal relationships. However, most constraint-based causal discovery algorithms, including the well-celebrated PC algorithm, often incur an exponential number of conditional independence (CI) tests, posing limitations in various applications. Addressing this, our work focuses on characterizing what can be learned about the underlying causal graph with a reduced number of CI tests. We show that it is possible to a learn a coarser representation of the hidden causal graph with a polynomial number of tests. This coarser representation, named Causal Consistent Partition Graph (CCPG), comprises of a partition of the vertices and a directed graph defined over its components. CCPG satisfies consistency of orientations and additional constraints which favor finer partitions. Furthermore, it reduces to the underlying causal graph when the causal graph is identifiable. As a consequence, our results offer the first efficient algorithm for recovering the true causal graph with a polynomial number of tests, in special cases where the causal graph is fully identifiable through observational data and potentially additional interventions.
Abstract:Multi-modality spatio-temporal (MoST) data extends spatio-temporal (ST) data by incorporating multiple modalities, which is prevalent in monitoring systems, encompassing diverse traffic demands and air quality assessments. Despite significant strides in ST modeling in recent years, there remains a need to emphasize harnessing the potential of information from different modalities. Robust MoST forecasting is more challenging because it possesses (i) high-dimensional and complex internal structures and (ii) dynamic heterogeneity caused by temporal, spatial, and modality variations. In this study, we propose a novel MoST learning framework via Self-Supervised Learning, namely MoSSL, which aims to uncover latent patterns from temporal, spatial, and modality perspectives while quantifying dynamic heterogeneity. Experiment results on two real-world MoST datasets verify the superiority of our approach compared with the state-of-the-art baselines. Model implementation is available at https://github.com/beginner-sketch/MoSSL.
Abstract:Understanding causal relationships between variables is a fundamental problem with broad impact in numerous scientific fields. While extensive research has been dedicated to learning causal graphs from data, its complementary concept of testing causal relationships has remained largely unexplored. While learning involves the task of recovering the Markov equivalence class (MEC) of the underlying causal graph from observational data, the testing counterpart addresses the following critical question: Given a specific MEC and observational data from some causal graph, can we determine if the data-generating causal graph belongs to the given MEC? We explore constraint-based testing methods by establishing bounds on the required number of conditional independence tests. Our bounds are in terms of the size of the maximum undirected clique ($s$) of the given MEC. In the worst case, we show a lower bound of $\exp(\Omega(s))$ independence tests. We then give an algorithm that resolves the task with $\exp(O(s))$ tests, matching our lower bound. Compared to the learning problem, where algorithms often use a number of independence tests that is exponential in the maximum in-degree, this shows that testing is relatively easier. In particular, it requires exponentially less independence tests in graphs featuring high in-degrees and small clique sizes. Additionally, using the DAG associahedron, we provide a geometric interpretation of testing versus learning and discuss how our testing result can aid learning.
Abstract:With the growing use of large language models hosted on cloud platforms to offer inference services, privacy concerns are escalating, especially concerning sensitive data like investment plans and bank account details. Secure Multi-Party Computing (SMPC) emerges as a promising solution to protect the privacy of inference data and model parameters. However, the application of SMPC in Privacy-Preserving Inference (PPI) for large language models, particularly those based on the Transformer architecture, often leads to considerable slowdowns or declines in performance. This is largely due to the multitude of nonlinear operations in the Transformer architecture, which are not well-suited to SMPC and difficult to circumvent or optimize effectively. To address this concern, we introduce an advanced optimization framework called SecFormer, to achieve fast and accurate PPI for Transformer models. By implementing model design optimization, we successfully eliminate the high-cost exponential and maximum operations in PPI without sacrificing model performance. Additionally, we have developed a suite of efficient SMPC protocols that utilize segmented polynomials, Fourier series and Goldschmidt's method to handle other complex nonlinear functions within PPI, such as GeLU, LayerNorm, and Softmax. Our extensive experiments reveal that SecFormer outperforms MPCFormer in performance, showing improvements of $5.6\%$ and $24.2\%$ for BERT$_{\text{BASE}}$ and BERT$_{\text{LARGE}}$, respectively. In terms of efficiency, SecFormer is 3.56 and 3.58 times faster than Puma for BERT$_{\text{BASE}}$ and BERT$_{\text{LARGE}}$, demonstrating its effectiveness and speed.
Abstract:In the field of deep learning, Stochastic Gradient Descent (SGD) and its momentum-based variants are the predominant choices for optimization algorithms. Despite all that, these momentum strategies, which accumulate historical gradients by using a fixed $\beta$ hyperparameter to smooth the optimization processing, often neglect the potential impact of the variance of historical gradients on the current gradient estimation. In the gradient variance during training, fluctuation indicates the objective function does not meet the Lipschitz continuity condition at all time, which raises the troublesome optimization problem. This paper aims to explore the potential benefits of reducing the variance of historical gradients to make optimizer converge to flat solutions. Moreover, we proposed a new optimization method based on reducing the variance. We employed the Wiener filter theory to enhance the first moment estimation of SGD, notably introducing an adaptive weight to optimizer. Specifically, the adaptive weight dynamically changes along with temporal fluctuation of gradient variance during deep learning model training. Experimental results demonstrated our proposed adaptive weight optimizer, SGDF (Stochastic Gradient Descent With Filter), can achieve satisfactory performance compared with state-of-the-art optimizers.
Abstract:Learning causal structures from interventional data is a fundamental problem with broad applications across various fields. While many previous works have focused on recovering the entire causal graph, in practice, there are scenarios where learning only part of the causal graph suffices. This is called $targeted$ causal discovery. In our work, we focus on two such well-motivated problems: subset search and causal matching. We aim to minimize the number of interventions in both cases. Towards this, we introduce the $Meek~separator$, which is a subset of vertices that, when intervened, decomposes the remaining unoriented edges into smaller connected components. We then present an efficient algorithm to find Meek separators that are of small sizes. Such a procedure is helpful in designing various divide-and-conquer-based approaches. In particular, we propose two randomized algorithms that achieve logarithmic approximation for subset search and causal matching, respectively. Our results provide the first known average-case provable guarantees for both problems. We believe that this opens up possibilities to design near-optimal methods for many other targeted causal structure learning problems arising from various applications.