A recent empirical observation (Li et al., 2022b) of activation sparsity in MLP blocks offers an opportunity to drastically reduce computation costs for free. Although having attributed it to training dynamics, existing theoretical explanations of activation sparsity are restricted to shallow networks, small training steps and special training, despite its emergence in deep models standardly trained for a large number of steps. To fill these gaps, we propose the notion of gradient sparsity as one source of activation sparsity and a theoretical explanation based on it that sees sparsity a necessary step to adversarial robustness w.r.t. hidden features and parameters, which is approximately the flatness of minima for well-learned models. The theory applies to standardly trained LayerNorm-ed MLPs, and further to Transformers or other architectures trained with weight noises. Eliminating other sources of flatness except for sparsity, we discover the phenomenon that the ratio between the largest and smallest non-zero singular values of weight matrices is small. When discussing the emergence of this spectral concentration, we use random matrix theory (RMT) as a powerful tool to analyze stochastic gradient noises. Validational experiments are conducted to verify our gradient-sparsity-based explanation. We propose two plug-and-play modules for both training and finetuning for sparsity. Experiments on ImageNet-1k and C4 demonstrate their 50% sparsity improvements, indicating further potential cost reduction in both training and inference.
相關內容
Source Routing, currently facilitated by Segment Routing (SR), enables precise control of forwarding paths by specifying detours (or segments) to deviate IP packets along routes with advanced properties beyond typical shortest IGP paths. Computing the desired optimal segment lists, known as encoding, leads to interesting challenges as the number of detours is tightly constrained for hardware performance. Existing solutions either lack generality, correctness, optimality, or practical computing efficiency-in particular for sparse realistic networks. In this paper, we address all such challenges with GOFOR-SR. Our framework extends usual path computation algorithms to inherently look at optimal and feasible segment lists, streamlining the deployment of TE-compliant paths. By integrating encoding within the path computation itself and modifying the distance comparison method, GOFOR allows algorithms with various optimization objectives to efficiently compute optimal segment lists. Despite the loss of substructure optimality induced by SR, GOFOR proves particularly efficient, inducing only a linear overhead at worst. It also offers different strategies and path diversity options for intricate TE-aware loadbalancing. We formally prove the correctness and optimality of GOFOR, implement our framework for various practical usecases, and demonstrate its performance and benefits on both real and challenging topologies.
Although the expenses associated with DNA sequencing have been rapidly decreasing, the current cost of sequencing information stands at roughly $120/GB, which is dramatically more expensive than reading from existing archival storage solutions today. In this work, we aim to reduce not only the cost but also the latency of DNA storage by initiating the study of the DNA coverage depth problem, which aims to reduce the required number of reads to retrieve information from the storage system. Under this framework, our main goal is to understand the effect of error-correcting codes and retrieval algorithms on the required sequencing coverage depth. We establish that the expected number of reads that are required for information retrieval is minimized when the channel follows a uniform distribution. We also derive upper and lower bounds on the probability distribution of this number of required reads and provide a comprehensive upper and lower bound on its expected value. We further prove that for a noiseless channel and uniform distribution, MDS codes are optimal in terms of minimizing the expected number of reads. Additionally, we study the DNA coverage depth problem under the random-access setup, in which the user aims to retrieve just a specific information unit from the entire DNA storage system. We prove that the expected retrieval time is at least k for [n,k] MDS codes as well as for other families of codes. Furthermore, we present explicit code constructions that achieve expected retrieval times below k and evaluate their performance through analytical methods and simulations. Lastly, we provide lower bounds on the maximum expected retrieval time. Our findings offer valuable insights for reducing the cost and latency of DNA storage.
Extraterrestrial autonomous lander missions increasingly demand adaptive capabilities to handle the unpredictable and diverse nature of the terrain. This paper discusses the deployment of a Deep Meta-Learning with Controlled Deployment Gaps (CoDeGa) trained model for terrain scooping tasks in Ocean Worlds Lander Autonomy Testbed (OWLAT) at NASA Jet Propulsion Laboratory. The CoDeGa-powered scooping strategy is designed to adapt to novel terrains, selecting scooping actions based on the available RGB-D image data and limited experience. The paper presents our experiences with transferring the scooping framework with CoDeGa-trained model from a low-fidelity testbed to the high-fidelity OWLAT testbed. Additionally, it validates the method's performance in novel, realistic environments, and shares the lessons learned from deploying learning-based autonomy algorithms for space exploration. Experimental results from OWLAT substantiate the efficacy of CoDeGa in rapidly adapting to unfamiliar terrains and effectively making autonomous decisions under considerable domain shifts, thereby endorsing its potential utility in future extraterrestrial missions.
Deep reinforcement learning (DRL) provides a promising way for intelligent agents (e.g., autonomous vehicles) to learn to navigate complex scenarios. However, DRL with neural networks as function approximators is typically considered a black box with little explainability and often suffers from suboptimal performance, especially for autonomous navigation in highly interactive multi-agent environments. To address these issues, we propose three auxiliary tasks with spatio-temporal relational reasoning and integrate them into the standard DRL framework, which improves the decision making performance and provides explainable intermediate indicators. We propose to explicitly infer the internal states (i.e., traits and intentions) of surrounding agents (e.g., human drivers) as well as to predict their future trajectories in the situations with and without the ego agent through counterfactual reasoning. These auxiliary tasks provide additional supervision signals to infer the behavior patterns of other interactive agents. Multiple variants of framework integration strategies are compared. We also employ a spatio-temporal graph neural network to encode relations between dynamic entities, which enhances both internal state inference and decision making of the ego agent. Moreover, we propose an interactivity estimation mechanism based on the difference between predicted trajectories in these two situations, which indicates the degree of influence of the ego agent on other agents. To validate the proposed method, we design an intersection driving simulator based on the Intelligent Intersection Driver Model (IIDM) that simulates vehicles and pedestrians. Our approach achieves robust and state-of-the-art performance in terms of standard evaluation metrics and provides explainable intermediate indicators (i.e., internal states, and interactivity scores) for decision making.
Fault detection and diagnosis of electrical motors are of utmost importance in ensuring the safe and reliable operation of several industrial systems. Detection and diagnosis of faults at the incipient stage allows corrective actions to be taken in order to reduce the severity of faults. The existing data-driven deep learning approaches for machine fault diagnosis rely extensively on huge amounts of labeled samples, where annotations are expensive and time-consuming. However, a major portion of unlabeled condition monitoring data is not exploited in the training process. To overcome this limitation, we propose a foundational model-based Active Learning framework that utilizes less amount of labeled samples, which are most informative and harnesses a large amount of available unlabeled data by effectively combining Active Learning and Contrastive Self-Supervised Learning techniques. It consists of a transformer network-based backbone model trained using an advanced nearest-neighbor contrastive self-supervised learning method. This approach empowers the backbone to learn improved representations of samples derived from raw, unlabeled vibration data. Subsequently, the backbone can undergo fine-tuning to address a range of downstream tasks, both within the same machines and across different machines. The effectiveness of the proposed methodology has been assessed through the fine-tuning of the backbone for multiple target tasks using three distinct machine-bearing fault datasets. The experimental evaluation demonstrates a superior performance as compared to existing state-of-the-art fault diagnosis methods with less amount of labeled data.
The convergence of deterministic policy gradient under the Hadamard parameterization is studied in the tabular setting and the linear convergence of the algorithm is established. To this end, we first show that the error decreases at an $O(\frac{1}{k})$ rate for all the iterations. Based on this result, we further show that the algorithm has a faster local linear convergence rate after $k_0$ iterations, where $k_0$ is a constant that only depends on the MDP problem and the initialization. To show the local linear convergence of the algorithm, we have indeed established the contraction of the sub-optimal probability $b_s^k$ (i.e., the probability of the output policy $\pi^k$ on non-optimal actions) when $k\ge k_0$.
Language Models have ushered a new age of AI gaining traction within the NLP community as well as amongst the general population. AI's ability to make predictions, generations and its applications in sensitive decision-making scenarios, makes it even more important to study these models for possible biases that may exist and that can be exaggerated. We conduct a quality comparative study and establish a framework to evaluate language models under the premise of two kinds of biases: gender and race, in a professional setting. We find out that while gender bias has reduced immensely in newer models, as compared to older ones, racial bias still exists.
This study employs machine learning models to predict the failure of Peer-to-Peer (P2P) lending platforms, specifically in China. By employing the filter method and wrapper method with forward selection and backward elimination, we establish a rigorous and practical procedure that ensures the robustness and importance of variables in predicting platform failures. The research identifies a set of robust variables that consistently appear in the feature subsets across different selection methods and models, suggesting their reliability and relevance in predicting platform failures. The study highlights that reducing the number of variables in the feature subset leads to an increase in the false acceptance rate while the performance metrics remain stable, with an AUC value of approximately 0.96 and an F1 score of around 0.88. The findings of this research provide significant practical implications for regulatory authorities and investors operating in the Chinese P2P lending industry.
The allure of aesthetic appeal in images captivates our senses, yet the underlying intricacies of aesthetic preferences remain elusive. In this study, we pioneer a novel perspective by utilizing machine learning models that focus on aesthetic attributes known to influence preferences. Through a data mining approach, our models process these attributes as inputs to predict the aesthetic scores of images. Moreover, to delve deeper and obtain interpretable explanations regarding the factors driving aesthetic preferences, we utilize the popular Explainable AI (XAI) technique known as SHapley Additive exPlanations (SHAP). Our methodology involves employing various machine learning models, including Random Forest, XGBoost, Support Vector Regression, and Multilayer Perceptron, to compare their performances in accurately predicting aesthetic scores, and consistently observing results in conjunction with SHAP. We conduct experiments on three image aesthetic benchmarks, providing insights into the roles of attributes and their interactions. Ultimately, our study aims to shed light on the complex nature of aesthetic preferences in images through machine learning and provides a deeper understanding of the attributes that influence aesthetic judgements.
The emerging mission-critical Internet of Things (IoT) play a vital role in remote healthcare, haptic interaction, and industrial automation, where timely delivery of status updates is crucial. The Age of Information (AoI) is an effective metric to capture and evaluate information freshness at the destination. A system design based solely on the optimization of the average AoI might not be adequate to capture the requirements of mission-critical applications, since averaging eliminates the effects of extreme events. In this paper, we introduce a Deep Reinforcement Learning (DRL)-based algorithm to improve AoI in mission-critical IoT applications. The objective is to minimize an AoI-based metric consisting of the weighted sum of the average AoI and the probability of exceeding an AoI threshold. We utilize the actor-critic method to train the algorithm to achieve optimized scheduling policy to solve the formulated problem. The performance of our proposed method is evaluated in a simulated setup and the results show a significant improvement in terms of the average AoI and the AoI violation probability compared to the related-work.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191