Recurrent Neural Networks (RNNs) are useful in temporal sequence tasks. However, training RNNs involves dense matrix multiplications which require hardware that can support a large number of arithmetic operations and memory accesses. Implementing online training of RNNs on the edge calls for optimized algorithms for an efficient deployment on hardware. Inspired by the spiking neuron model, the Delta RNN exploits temporal sparsity during inference by skipping over the update of hidden states from those inactivated neurons whose change of activation across two timesteps is below a defined threshold. This work describes a training algorithm for Delta RNNs that exploits temporal sparsity in the backward propagation phase to reduce computational requirements for training on the edge. Due to the symmetric computation graphs of forward and backward propagation during training, the gradient computation of inactivated neurons can be skipped. Results show a reduction of ∼80% in matrix operations for training a 56k parameter Delta LSTM on the Fluent Speech Commands dataset with negligible accuracy loss. Logic simulations of a hardware accelerator designed for the training algorithm show 2-10X speedup in matrix computations for an activation sparsity range of 50%-90%. Additionally, we show that the proposed Delta RNN training will be useful for online incremental learning on edge devices with limited computing resources.
论文检索
输入标题、作者或关键词，从 11,272 篇学术成果中精准定位
Most reinforcement learning algorithms seek a single optimal strategy that solves a given task. However, it can often be valuable to learn a diverse set of solutions, for instance, to make an agent's interaction with users more engaging, or improve the robustness of a policy to an unexpected perturbance. We propose Diversity-Guided Policy Optimization (DGPO), an on-policy algorithm that discovers multiple strategies for solving a given task. Unlike prior work, it achieves this with a shared policy network trained over a single run. Specifically, we design an intrinsic reward based on an information-theoretic diversity objective. Our final objective alternately constraints on the diversity of the strategies and on the extrinsic reward. We solve the constrained optimization problem by casting it as a probabilistic inference task and use policy iteration to maximize the derived lower bound. Experimental results show that our method efficiently discovers diverse strategies in a wide variety of reinforcement learning tasks. Compared to baseline methods, DGPO achieves comparable rewards, while discovering more diverse strategies, and often with better sample efficiency.
PG-LBO: Enhancing High-Dimensional Bayesian Optimization with Pseudo-Label and Gaussian Process Guidance
PDF ↗Variational Autoencoder based Bayesian Optimization (VAE-BO) has demonstrated its excellent performance in addressing high-dimensional structured optimization problems. However, current mainstream methods overlook the potential of utilizing a pool of unlabeled data to construct the latent space, while only concentrating on designing sophisticated models to leverage the labeled data. Despite their effective usage of labeled data, these methods often require extra network structures, additional procedure, resulting in computational inefficiency. To address this issue, we propose a novel method to effectively utilize unlabeled data with the guidance of labeled data. Specifically, we tailor the pseudo-labeling technique from semi-supervised learning to explicitly reveal the relative magnitudes of optimization objective values hidden within the unlabeled data. Based on this technique, we assign appropriate training weights to unlabeled data to enhance the construction of a discriminative latent space. Furthermore, we treat the VAE encoder and the Gaussian Process (GP) in Bayesian optimization as a unified deep kernel learning process, allowing the direct utilization of labeled data, which we term as Gaussian Process guidance. This directly and effectively integrates the goal of improving GP accuracy into the VAE training, thereby guiding the construction of the latent space. The extensive experiments demonstrate that our proposed method outperforms existing VAE-BO algorithms in various optimization scenarios. Our code will be published at https://github.com/TaicaiChen/PG-LBO.
Traditional gradient descent (GD) has been fully investigated for convex or L-smoothness functions, and it is widely utilized in current neural network optimization. The classical descent lemma ensures that for a function with L-smoothness, the GD trajectory converges stably towards the minimum when the learning rate is below 2 / L. This convergence is marked by a consistent reduction in the loss function throughout the iterations. However, recent experimental studies have demonstrated that even when the L-smoothness condition is not met, or if the learning rate is increased leading to oscillations in the loss function during iterations, the GD trajectory still exhibits convergence over the long run. This phenomenon is referred to as the unstable convergence regime of GD. In this paper, we present a theoretical perspective to offer a qualitative analysis of this phenomenon. The unstable convergence is in fact an inherent property of GD for general twice differentiable functions. Specifically, the forwardinvariance of GD is established, i.e., it ensures that any point within a local region will always remain within this region under GD iteration. Then, based on the forward-invariance, for the initialization outside an open set containing the local minimum, the loss function will oscillate at the first several iterations and then become monotonely decreasing after the GD trajectory jumped into the open set. This work theoretically clarifies the unstable convergence phenomenon of GD discussed in previous experimental works. The unstable convergence of GD mainly depends on the selection of the initialization, and it is actually inevitable due to the complex nature of loss function.
Graph Convolutional Network (GCN) has exhibited remarkable potential in improving graph-based clustering. To handle the general clustering scenario without a prior graph, these models estimate an initial graph beforehand to apply GCN. Throughout the literature, we have witnessed that 1) most models focus on the initial graph while neglecting the original features. Therefore, the discriminability of the learned representation may be corrupted by a low-quality initial graph; 2) the training procedure lacks effective clustering guidance, which may lead to the incorporation of clustering-irrelevant information into the learned graph. To tackle these problems, the Deep Contrastive Graph Learning (DCGL) model is proposed for general data clustering. Specifically, we establish a pseudo-siamese network, which incorporates auto-encoder with GCN to emphasize both the graph structure and the original features. On this basis, feature-level contrastive learning is introduced to enhance the discriminative capacity, and the relationship between samples and centroids is employed as the clustering-oriented guidance. Afterward, a two-branch graph learning mechanism is designed to extract the local and global structural relationships, which are further embedded into a unified graph under the cluster-level contrastive guidance. Experimental results on several benchmark datasets demonstrate the superiority of DCGL against state-of-the-art algorithms.
With the rapid growth of audio data, there's a pressing need for automatic audio classification. As a type of time-series data, audio exhibits waveform fluctuations in both the time and frequency domains that evolve over time, with similar instances sharing consistent patterns. This study introduces the Audio Scanning Network (ASNet), designed to leverage abundant information for achieving stable and effective audio classification. ASNet captures real-time changes in audio waveforms across both time and frequency domains through reservoir computing, supported by Reservoir Kernel Canonical Correlation Analysis (RKCCA) to explore correlations between time-domain and frequency-domain waveform fluctuations. This innovative approach empowers ASNet to comprehensively capture the changes and inherent correlations within the audio waveform, and without the need for time-consuming iterative training. Instead of converting audio into spectrograms, ASNet directly utilizes audio feature sequences to uncover associations between time and frequency fluctuations. Experiments on environmental sound and music genre classification tasks demonstrate ASNet's comparable performance to state-of-the-art methods.
In this work, we focus on high-dimensional single index models with non-Gaussian sensing vectors and generative priors. More specifically, our goal is to estimate the underlying signal from i.i.d. realizations of the semi-parameterized single index model, where the underlying signal is contained in (up to a constant scaling) the range of a Lipschitz continuous generative model with bounded low-dimensional inputs, the sensing vector follows a non-Gaussian distribution, the noise is a random variable that is independent of the sensing vector, and the unknown non-linear link function is differentiable. Using the first- and second-order Stein's identity, we introduce efficient algorithms to obtain estimated vectors that achieve the near-optimal statistical rate. Experimental results on image datasets are provided to support our theory.
Improving the diversity of Artificial Intelligence Generated Content (AIGC) is one of the fundamental problems in the theory of generative models such as generative adversarial networks (GANs). Previous studies have demonstrated that the discriminator in GANs should have high capacity and robustness to achieve the diversity of generated data. However, a discriminator with high capacity tends to overfit and guide the generator toward collapsed equilibrium. In this study, we propose a novel discriminative forest GAN, named Forest-GAN, that replaces the discriminator to improve the capacity and robustness for modeling statistics in real-world data distribution. A discriminative forest is composed of multiple independent discriminators built on bootstrapped data. We prove that a discriminative forest has a generalization error bound, which is determined by the strength of individual discriminators and the correlations among them. Hence, a discriminative forest can provide very large capacity without any risk of overfitting, which subsequently improves the generative diversity. With the discriminative forest framework, we significantly improved the performance of AutoGAN with a new record FID of 19.27 from 30.71 on STL10 and improved the performance of StyleGAN2-ADA with a new record FID of 6.87 from 9.22 on LSUN-cat.
Federated learning encounters substantial challenges with heterogeneous data, leading to performance degradation and convergence issues. While considerable progress has been achieved in mitigating such an impact, the reliability aspect of federated models has been largely disregarded. In this study, we conduct extensive experiments to investigate the reliability of both generic and personalized federated models. Our exploration uncovers a significant finding: federated models exhibit unreliability when faced with heterogeneous data, demonstrating poor calibration on in-distribution test data and low uncertainty levels on out-of-distribution data. This unreliability is primarily attributed to the presence of biased projection heads, which introduce miscalibration into the federated models. Inspired by this observation, we propose the "Assembled Projection Heads" (APH) method for enhancing the reliability of federated models. By treating the existing projection head parameters as priors, APH randomly samples multiple initialized parameters of projection heads from the prior and further performs targeted fine-tuning on locally available data under varying learning rates. Such a head ensemble introduces parameter diversity into the deterministic model, eliminating the bias and producing reliable predictions via head averaging. We evaluate the effectiveness of the proposed APH method across three prominent federated benchmarks. Experimental results validate the efficacy of APH in model calibration and uncertainty estimation. Notably, APH can be seamlessly integrated into various federated approaches but only requires less than 30% additional computation cost for 100x inferences within large models.
Accelerate Multi-Agent Reinforcement Learning in Zero-Sum Games with Subgame Curriculum Learning
PDF ↗Learning Nash equilibrium (NE) in complex zero-sum games with multi-agent reinforcement learning (MARL) can be extremely computationally expensive. Curriculum learning is an effective way to accelerate learning, but an under-explored dimension for generating a curriculum is the difficulty-to-learn of the subgames –games induced by starting from a specific state. In this work, we present a novel subgame curriculum learning framework for zero-sum games. It adopts an adaptive initial state distribution by resetting agents to some previously visited states where they can quickly learn to improve performance. Building upon this framework, we derive a subgame selection metric that approximates the squared distance to NE values and further adopt a particle-based state sampler for subgame generation. Integrating these techniques leads to our new algorithm, Subgame Automatic Curriculum Learning (SACL), which is a realization of the subgame curriculum learning framework. SACL can be combined with any MARL algorithm such as MAPPO. Experiments in the particle-world environment and Google Research Football environment show SACL produces much stronger policies than baselines. In the challenging hide-and-seek quadrant environment, SACL produces all four emergent stages and uses only half the samples of MAPPO with self-play. The project website is at https://sites.google.com/view/sacl-neurips.
On Disentanglement of Asymmetrical Knowledge Transfer for Modality-Task Agnostic Federated Learning
PDF ↗There has been growing concern regarding data privacy during the development and deployment of Multimodal Foundation Models for Artificial General Intelligence (AGI), while Federated Learning (FL) allows multiple clients to collaboratively train models in a privacy-preserving manner. This paper formulates and studies Modality-task Agnostic Federated Learning (AFL) to pave the way toward privacy-preserving AGI. A unique property of AFL is the asymmetrical knowledge relationships among clients due to modality gaps, task gaps, and domain shifts between clients. This raises a challenge in learning an optimal inter-client information-sharing scheme that maximizes positive transfer and minimizes negative transfer for AFL. However, prior FL methods, mostly focusing on symmetrical knowledge transfer, tend to exhibit insufficient positive transfer and fail to fully avoid negative transfer during inter-client collaboration. To address this issue, we propose DisentAFL, which leverages a two-stage Knowledge Disentanglement and Gating mechanism to explicitly decompose the original asymmetrical inter-client information-sharing scheme into several independent symmetrical inter-client information-sharing schemes, each of which corresponds to certain semantic knowledge type learned from the local tasks. Experimental results demonstrate the superiority of our method on AFL than baselines.
How our brain encodes complex concepts has been a longstanding mystery in neuroscience. The answer to this problem can lead to new understandings about how the brain retrieves information in large-scale data with high efficiency and robustness. Neuroscience studies suggest the brain represents concepts in a locality-sensitive hashing (LSH) strategy, i.e., similar concepts will be represented by similar responses. This finding has inspired the design of similarity-based algorithms, especially in contrastive learning. Here, we hypothesize that the brain and large neural network models, both using similarity-based learning rules, could contain a similar semantic embedding space. To verify that, this paper proposes a functional Magnetic Resonance Imaging (fMRI) semantic learning network named BrainSem, aimed at seeking a joint semantic latent space that bridges the brain and a Contrastive Language-Image Pre-training (CLIP) model. Given that our perception is inherently cross-modal, we introduce a fuzzy (one-to-many) matching loss function to encourage the models to extract high-level semantic components from neural signals. Our results claimed that using only a small set of fMRI recordings for semantic space alignment, we could obtain shared embedding valid for unseen categories out of the training set, which provided potential evidence for the semantic representation similarity between the brain and large neural networks. In a zero-shot classification task, our BrainSem achieves an 11.6% improvement over the state-of-the-art.
Accurately detecting multiple change-points is critical for various applications, but determining the optimal number of change-points remains a challenge. Existing approaches based on information criteria attempt to balance goodness-of-fit and model complexity, but their performance varies depending on the model. Recently, data-driven selection criteria based on cross-validation has been proposed, but these methods can be prone to slight overfitting in finite samples. In this paper, we introduce a method that controls the probability of overestimation and provides uncertainty quantification for learning multiple change-points via cross-validation. We frame this problem as a sequence of model comparison problems and leverage high-dimensional inferential procedures. We demonstrate the effectiveness of our approach through experiments on finite-sample data, showing superior uncertainty quantification for overestimation compared to existing methods. Our approach has broad applicability and can be used in diverse change-point models.
FedDAT: An Approach for Foundation Model Finetuning in Multi-Modal Heterogeneous Federated Learning
PDF ↗Recently, foundation models have exhibited remarkable advancements in multi-modal learning. These models, equipped with millions (or billions) of parameters, typically require a substantial amount of data for finetuning. However, collecting and centralizing training data from diverse sectors becomes challenging due to distinct privacy regulations. Federated Learning (FL) emerges as a promising solution, enabling multiple clients to collaboratively train neural networks without centralizing their local data. To alleviate client computation burdens and communication overheads, previous works have adapted Parameter-efficient Finetuning (PEFT) methods for FL. Hereby, only a small fraction of the model parameters are optimized and communicated during federated communications. Nevertheless, most previous works have focused on a single modality and neglected one common phenomenon, i.e., the presence of data heterogeneity across the clients. Therefore, in this work, we propose a finetuning framework tailored to heterogeneous multi-modal FL, called Federated Dual-Aadapter Teacher (FedDAT). Specifically, our approach leverages a Dual-Adapter Teacher (DAT) to address data heterogeneity by regularizing the client local updates and applying Mutual Knowledge Distillation (MKD) for an efficient knowledge transfer. FedDAT is the first approach that enables an efficient distributed finetuning of foundation models for a variety of heterogeneous Vision-Language tasks. To demonstrate its effectiveness, we conduct extensive experiments on four multi-modality FL benchmarks with different types of data heterogeneity, where FedDAT substantially outperforms the existing centralized PEFT methods adapted for FL.
One important desideratum of lifelong learning aims to discover novel classes from unlabelled data in a continuous manner. The central challenge is twofold: discovering and learning novel classes while mitigating the issue of catastrophic forgetting of established knowledge. To this end, we introduce a new paradigm called Adaptive Discovering and Merging (ADM) to discover novel categories adaptively in the incremental stage and integrate novel knowledge into the model without affecting the original knowledge. To discover novel classes adaptively, we decouple representation learning and novel class discovery, and use Triple Comparison (TC) and Probability Regularization (PR) to constrain the probability discrepancy and diversity for adaptive category assignment. To merge the learned novel knowledge adaptively, we propose a hybrid structure with base and novel branches named Adaptive Model Merging (AMM), which reduces the interference of the novel branch on the old classes to preserve the previous knowledge, and merges the novel branch to the base model without performance loss and parameter growth. Extensive experiments on several datasets show that ADM significantly outperforms existing class-incremental Novel Class Discovery (class-iNCD) approaches. Moreover, our AMM also benefits the class-incremental Learning (class-IL) task by alleviating the catastrophic forgetting problem. The source code is included in the supplementary materials.
A Generalized Shuffle Framework for Privacy Amplification: Strengthening Privacy Guarantees and Enhancing Utility
PDF ↗The shuffle model of local differential privacy is an advanced method of privacy amplification designed to enhance privacy protection with high utility. It achieves this by randomly shuffling sensitive data, making linking individual data points to specific individuals more challenging. However, most existing studies have focused on the shuffle model based on (ε0,0)-Locally Differentially Private (LDP) randomizers, with limited consideration for complex scenarios such as (ε0,δ0)-LDP or personalized LDP (PLDP). This hinders a comprehensive understanding of the shuffle model's potential and limits its application in various settings. To bridge this research gap, we propose a generalized shuffle framework that can be applied to PLDP setting. This generalization allows for a broader exploration of the privacy-utility trade-off and facilitates the design of privacy-preserving analyses in diverse contexts. We prove that the shuffled PLDP process approximately preserves μ-Gaussian Differential Privacy with μ = O(1/√n). This approach allows us to avoid the limitations and potential inaccuracies associated with inequality estimations. To strengthen the privacy guarantee, we improve the lower bound by utilizing hypothesis testing instead of relying on rough estimations like the Chernoff bound or Hoeffding's inequality. Furthermore, extensive comparative evaluations clearly show that our approach outperforms existing methods in achieving strong central privacy guarantees while preserving the utility of the global model. We have also carefully designed corresponding algorithms for average function, frequency estimation, and stochastic gradient descent.
Neural network compression techniques, such as knowledge distillation (KD) and network pruning, have received increasing attention. Recent work `Prune, then Distill' reveals that a pruned student-friendly teacher network can benefit the performance of KD. However, the conventional teacher-student pipeline, which entails cumbersome pre-training of the teacher and complicated compression steps, makes pruning with KD less efficient. In addition to compressing models, recent compression techniques also emphasize the aspect of efficiency. Early pruning demands significantly less computational cost in comparison to the conventional pruning methods as it does not require a large pre-trained model. Likewise, a special case of KD, known as self-distillation (SD), is more efficient since it requires no pre-training or student-teacher pair selection. This inspires us to collaborate early pruning with SD for efficient model compression. In this work, we propose the framework named Early Pruning with Self-Distillation (EPSD), which identifies and preserves distillable weights in early pruning for a given SD task. EPSD efficiently combines early pruning and self-distillation in a two-step process, maintaining the pruned network's trainability for compression. Instead of a simple combination of pruning and SD, EPSD enables the pruned network to favor SD by keeping more distillable weights before training to ensure better distillation of the pruned network. We demonstrated that EPSD improves the training of pruned networks, supported by visual and quantitative analyses. Our evaluation covered diverse benchmarks (CIFAR-10/100, Tiny-ImageNet, full ImageNet, CUB-200-2011, and Pascal VOC), with EPSD outperforming advanced pruning and SD techniques.
Pretrained large models, particularly large language models, have garnered increasing attention, as they have demonstrated remarkable abilities through contextual learning. Pretrained large models are increasingly recognized as fundamental tools for solving various tasks. However, the substantial computational demands of large models have dissuaded most product teams and individuals from running them. In such scenarios, to leverage the exceptional performance of large models, one must solely depend on costly APIs, further burdening product teams and individuals. On the other hand, despite the overall inferior performance of small models compared to large models, there are certain distributions where small models can achieve comparable or even superior results. For instance, during training, small models may become trapped in a local optimum that is unique to certain distributions, leading to superior performance. Hence, we propose Data Shunt (DS), a general paradigm for collaboration of small and large models. DS not only substantially reduces the cost associated with deploying large models but also effectively enhances overall performance. Specifically, DS determines the shunting direction by evaluating the confidence level of small models. When the confidence level falls below a specific threshold, the input data is forwarded to large models. To further leverage the advantages of the small and large models, we introduce Prompt Pruning (PP) and 2-Stage Confidence Distillation (2CD), which facilitate mutual collaboration, leading to better results and less cost. The remarkable performance across diverse modalities and tasks demonstrates the superiority of the proposed DS over large models. For instance, ChatGPT achieves an accuracy of 94.43% on Amazon Product sentiment analysis, and DS achieves an accuracy of 95.64%, while the cost has been reduced to only 31.18%. The code for the proposed method are provided for research purposes https://github.com/Anfeather/Data-Shunt.
Visual Reinforcement Learning (RL) is a promising approach to achieve human-like intelligence. However, it currently faces challenges in learning efficiently within noisy environments. In contrast, humans can quickly identify task-relevant objects in distraction-filled surroundings by applying previously acquired common knowledge. Recently, foundational models in natural language processing and computer vision have achieved remarkable successes, and the common knowledge within these models can significantly benefit downstream task training. Inspired by these achievements, we aim to incorporate common knowledge from foundational models into visual RL. We propose a novel Focus-Then-Decide (FTD) framework, allowing the agent to make decisions based solely on task-relevant objects. To achieve this, we introduce an attention mechanism to select task-relevant objects from the object set returned by a foundational segmentation model, and only use the task-relevant objects for the subsequent training of the decision module. Additionally, we specifically employed two generic self-supervised objectives to facilitate the rapid learning of this attention mechanism. Experimental results on challenging tasks based on DeepMind Control Suite and Franka Emika Robotics demonstrate that our method can quickly and accurately pinpoint objects of interest in noisy environments. Consequently, it achieves a significant performance improvement over current state-of-the-art algorithms. Project Page: https://www.lamda.nju.edu.cn/chenc/FTD.html Code: https://github.com/LAMDA-RL/FTD
Offline optimization is an emerging problem in many experimental engineering domains including protein, drug or aircraft design, where online experimentation to collect evaluation data is too expensive or dangerous. To avoid that, one has to optimize an unknown function given only its offline evaluation at a fixed set of inputs. A naive solution to this problem is to learn a surrogate model of the unknown function and optimize this surrogate instead. However, such a naive optimizer is prone to erroneous overestimation of the surrogate (possibly due to over-fitting on a biased sample of function evaluation) on inputs outside the offline dataset. Prior approaches addressing this challenge have primarily focused on learning robust surrogate models. However, their search strategies are derived from the surrogate model rather than the actual offline data. To fill this important gap, we introduce a new learning-to-search perspective for offline optimization by reformulating it as an offline reinforcement learning problem. Our proposed policy-guided gradient search approach explicitly learns the best policy for a given surrogate model created from the offline data. Our empirical results on multiple benchmarks demonstrate that the learned optimization policy can be combined with existing offline surrogates to significantly improve the optimization performance.