Neural architecture search (NAS) has gained significant traction in automating the design of neural networks. To reduce search time, differentiable architecture search (DAS) reframes the traditional paradigm of discrete candidate sampling and evaluation into a differentiable optimization over a super-net, followed by discretization. However, most existing DAS methods primarily focus on optimizing the coarse-grained operation-level topology, while neglecting finer-grained structures such as filter-level and weight-level patterns. This limits their ability to balance model performance with model size. In addition, many methods compromise search quality to save memory during the search process. To tackle these issues, we propose Multigranularity DAS (MG-DARTS), a unified framework that aims to discover both effective and efficient architectures from scratch by comprehensively yet memory-efficiently exploring a multigranularity search space. Specifically, we improve the existing DAS methods in two aspects. First, we adaptively adjust the retention ratios of searchable units across different granularity levels through adaptive pruning, which is achieved by learning granularity-specific discretization functions along with the evolving architecture. Second, we decompose the super-net optimization and discretization into multiple stages, each operating on a subnet, and introduce progressive re-evaluation to enable repruning and regrowth of previous units, thereby mitigating potential bias. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet demonstrate that MG-DARTS outperforms other state-of-the-art methods in achieving a better tradeoff between model accuracy and parameter efficiency. Codes are available at: https://github.com/lxy12357/MG_DARTS.

Generative adversarial networks (GAN) rely on the adversarial interplay between a generator (G) and a discriminator (D), where G aims to synthesize realistic data while D learns to distinguish real from generated samples. However, this dynamic often suffers from imbalance, as D quickly becomes proficient at classification, leading to overfitting and limiting G’s ability to generate diverse, high-quality outputs especially in data-limited scenarios. Existing methods primarily focus on improving D, inadvertently exacerbating this imbalance by making G’s task even more challenging. In this paper, we introduce DGA-GAN, a novel training paradigm that dynamically reshapes G’s strategy throughout training, allowing it to generate more diverse and high-fidelity samples while maintaining adversarial balance. Our approach introduces an expansion-consolidation mechanism, where G iteratively broadens its generative scope by learning from varied data subsets and consolidates this knowledge to refine its overall strategy. This dynamic adaptation forces D to continuously evolve rather than overfit, leading to a more stable and effective adversarial game. We demonstrate that DGA-GAN improves sample quality and diversity across multiple datasets (CIFAR-10, TinyImageNet, few-shot generation benchmarks) and GAN architectures (SNGAN, BigGAN, StyleGAN2, GAN-LTH, Re-GAN). Furthermore, our method seamlessly integrates with existing data augmentation techniques to enhance performance in data-limited settings. By introducing a generative model that adapts its learning trajectory in response to D’s evolution, DGA-GAN represents a paradigm shift in GAN training, paving the way for more adaptive and data-efficient generative models.

Recent advances in deep learning have witnessed many successful medical image translation models that learn correspondences between two visual domains. However, building robust mappings between domains is a significant challenge when handling misalignments caused by factors such as respiratory motion and anatomical changes. This issue is further exacerbated in scenarios with limited data availability, leading to a significant degradation in translation quality. In this paper, we introduce a novel data-efficient framework for aligning medical images via Reconfigurable Generative Network (Reconfig-MIT) for high-quality image translation. The key idea of Reconfig-MIT is to adaptively expand the generative network width within a Generative Adversarial Networks (GAN) architecture, initially expanding rapidly to capture low-level features and then slowing to refine high-level complexities. This dynamic network adaptation mechanism allows to adaptively learn at different rates, thus the model can better respond to deviations in the data caused by misalignments, while maintaining an effective equilibrium with the discriminator (D). We also introduce the Recursive Cycle-Consistency Loss (R-CCL), which extends the cycle consistency loss to effectively preserve key anatomical structures and their spatial relationships, improving translation quality. Extensive experiments show that Reconfig-MIT is a generic framework that enables easy integration with existing image translation methods, including those incorporating registration networks used for correcting misalignments, and provides robust and high-quality translation on paired and unpaired misaligned data in both data-rich and data-limited scenarios. https://github.com/IntellicentAI-Lab/Reconfig-MIT.

The training of Generative Adversarial Networks (GANs) for high-fidelity images has predominantly relied on large-scale datasets. Emerging research, particularly on GANs 'lottery tickets', suggests that dense GANs models have sparse sub-networks capable of superior performance with limited data. However, the conventional process to uncover these 'lottery tickets' involves a resource-intensive train-prune-retrain cycle. Addressing this, our paper introduces Re-GAN, a novel, data-efficient approach for GANs training that dynamically reconfigures the GANs architecture during training. This method focuses on iterative pruning of non-important connections and regrowing them, thereby preventing premature loss of important features and maintaining the model's representational strength. Re-GAN provides a more stable and efficient solution for GANs models with limited data, offering an alternative to existing progressive growing methods and GANs tickets. While Re-GAN has already demonstrated its potential in image generation across diverse datasets, domains, and resolutions, in this paper, we significantly expand our study. We incorporate new applications, notably Image-to-Image translation, include additional datasets, provide in-depth analyses, and explore compatibility with data augmentation techniques. This expansion not only broadens the scope of Re-GAN but also establishes it as a generic training methodology, demonstrating its effectiveness and adaptability in different GANs scenarios.

Distribution shifts from external events and new entities can significantly compromise spatial-temporal prediction accuracy, potentially leading to severe outcomes like traffic accidents. Existing methods often fail under these conditions due to two main limitations: they focus on invariant patterns, missing the diversity required to capture the evolving dynamics of distribution shifts; they rely on often inaccessible future knowledge, such as spatial information of new entities, limiting their generalizability. To address these limitations, we formally define the problem of inductive spatial-temporal prediction under continuous distribution shifts and introduce the Contrastive Learning Based Inductive Graph Neural Network (COIN-GNN) as a solution. We develop a novel metric, Relation Importance (RI), to effectively select stable entities and distinct spatial relationships, forming an informative subgraph. Additionally, we construct an informative temporal memory buffer to store and review influential timestamps identified using influence functions. COIN-GNN then generates pseudo-observations for unstable and uninformative entities during these influential timestamps, simulating potential distribution shifts. By applying contrastive learning, the network learns stable and informative representations that can effectively counter distribution shifts without relying on future knowledge. Our extensive experiments on several real-world datasets—from traffic to weather—demonstrate COIN-GNN’s superior performance across different domains without requiring future knowledge.

Managing heterogeneous datasets that vary in complexity, size, and similarity in continual learning presents a significant challenge. Task-agnostic continual learning is necessary to address this challenge, as datasets with varying similarity pose difficulties in distinguishing task boundaries. Conventional task-agnostic continual learning practices typically rely on rehearsal or regularization techniques. However, rehearsal methods may struggle with varying dataset sizes and regulating the importance of old and new data due to rigid buffer sizes. Meanwhile, regularization methods apply generic constraints to promote generalization but can hinder performance when dealing with dissimilar datasets lacking shared features, necessitating a more adaptive approach. In this article, we propose a novel adaptive continual learning (AdaptCL) method to tackle heterogeneity in sequential datasets. AdaptCL employs fine-grained data-driven pruning to adapt to variations in data complexity and dataset size. It also utilizes task-agnostic parameter isolation to mitigate the impact of varying degrees of catastrophic forgetting caused by differences in data similarity. Through a two-pronged case study approach, we evaluate AdaptCL on both datasets of MNIST variants and DomainNet, as well as datasets from different domains. The latter include both large-scale, diverse binary-class datasets and few-shot, multiclass datasets. Across all these scenarios, AdaptCL consistently exhibits robust performance, demonstrating its flexibility and general applicability in handling heterogeneous datasets.

Training Generative Adversarial Networks (GAN) to generate high-quality images typically requires large datasets. Network pruning during training has recently emerged as a significant advancement for data-efficient GAN. However, simple and straightforward pruning can lead to the risk of losing key information, resulting in suboptimal results due to GAN’s competitive dynamics between generator (G) and discriminator (D). Addressing this, we present RG-GAN, a novel approach that marks the first incorporation of dynamic weight regeneration and pruning in GAN training to improve the quality of the generated samples, even with limited data. Specifically, RG-GAN initiates layer-wise dynamic pruning by removing less important weights to the quality of the generated images. While pruning enhances efficiency, excessive sparsity within layers can pose a risk of model collapse. To mitigate this issue, RG-GAN applies a dynamic regeneration method to reintroduce specific weights when they become important, ensuring a balance between sparsity and image quality. Though effective, the sparse network achieved through this process might eliminate some weights important to the combined G and D performance, a crucial aspect for achieving stable and effective GAN training. RG-GAN addresses this loss of weights by integrating learned sparse network weights back into the dense network at the previous stage during a follow-up regeneration step. Our results consistently demonstrate RG-GAN’s robust performance across a variety of scenarios, including different GAN architectures, datasets, and degrees of data scarcity, reinforcing its value as a generic training methodology. Results also show that data augmentation exhibits improved performance in conjunction with RG-GAN. Furthermore, RG-GAN can achieve fewer parameters without compromising, and even enhancing, the quality of the generated samples. Code can be found at this link: https://github.com/IntellicentAI-Lab/RG-GAN

Multi-robot reinforcement learning (MRRL) is a promising approach to solving cooperation problems and has been widely adopted in many applications. In the past decades, researchers have proposed various approaches to improve the efficiency of MRRL. However, most of them are trained and evaluated only in simulated environments with simple interaction scenarios. The problem of how these methods perform in the real-world environment with complex interaction scenarios remains unsolved. To meet this emergent need, we introduce a scalable multi-robot reinforcement learning platform (SMART) for training and evaluation. Specifically, SMART consists of two components: 1) a simulation environment with an uncertainty-aware social agent model that provides a variety of complex interaction scenarios for training and 2) a real-world multi-robot system for realistic performance evaluation. To evaluate the generalizability of MRRL baselines, we introduce a novel generalization metric that takes into account their performance across changes in the environment as well as the policies of other agents. Furthermore, we conduct a case study on the multi-vehicle cooperative lane change and summarize the unique challenges of MRRL, which are rarely considered previously. Finally, we open-source the simulation environments, associated benchmark tasks, and state-of-the-art baselines to encourage and empower MRRL research. Our code is available at https://github.com/Blackmamba-xuan/MRST.

Despite advancements in Text-to-Video (T2V) generation, producing videos with realistic motion remains challenging. Current models often yield static or minimally dynamic outputs, failing to capture complex motions described by text. This issue stems from the internal biases in text encoding which overlooks motions, and inadequate conditioning mechanisms in T2V generation models. To address this, we propose a novel framework called DEcomposed MOtion (DEMO), which enhances motion synthesis in T2V generation by decomposing both text encoding and conditioning into content and motion components. Our method includes a content encoder for static elements and a motion encoder for temporal dynamics, alongside separate content and motion conditioning mechanisms. Crucially, we introduce text-motion and video-motion supervision to improve the model's understanding and generation of motion. Evaluations on benchmarks such as MSR-VTT, UCF-101, WebVid-10M, EvalCrafter, and VBench demonstrate DEMO's superior ability to produce videos with enhanced motion dynamics while maintaining high visual quality. Our approach significantly advances T2V generation by integrating comprehensive motion understanding directly from textual descriptions. Project page: https://PR-Ryan.github.io/DEMO-project/

Training Generative Adversarial Networks (GANs) on high-fidelity images usually requires a vast number of training images. Recent research on GAN tickets reveals that dense GANs models contain sparse sub-networks or “lottery tickets” that, when trained separately, yield better results under limited data. However, finding GANs tickets requires an expensive process of train-prune-retrain. In this paper, we propose Re-GAN, a data-efficient GANs training that dynamically reconfigures GANs architecture during training to explore different sub-network structures in training time. Our method repeatedly prunes unimportant connections to regularize GANs network and regrows them to reduce the risk of prematurely pruning important connections. Re-GAN stabilizes the GANs models with less data and offers an alternative to the existing GANs tickets and progressive growing methods. We demonstrate that Re-GAN is a generic training methodology which achieves stability on datasets of varying sizes, domains, and resolutions (CIFAR-10, Tiny-ImageNet, and multiple few-shot generation datasets) as well as different GANs architectures (SNGAN, ProGAN, StyleGAN2 and AutoGAN). Re-GAN also improves performance when combined with the recent augmentation approaches. Moreover, Re-GAN requires fewer floating-point operations (FLOPs) and less training time by removing the unimportant connections during GANs training while maintaining comparable or even generating higher-quality samples. When compared to state-of-the-art StyleGAN2, our method outperforms without requiring any additional fine-tuning step. Code can be found at this link: https://github.com/IntellicentAI-Lab/Re-GAN

Many real-world applications can be formulated as multi-agent cooperation problems, such as network packet routing and coordination of autonomous vehicles. The emergence of deep reinforcement learning (DRL) provides a promising approach for multi-agent cooperation through the interaction of the agents and environments. However, traditional DRL solutions suffer from the high dimensions of multiple agents with continuous action space during policy search. Besides, the dynamicity of agents’ policies makes the training non-stationary. To tackle the issues, we propose a hierarchical reinforcement learning approach with high-level decision-making and low-level individual control for efficient policy search. In particular, the cooperation of multiple agents can be learned in high-level discrete action space efficiently. At the same time, the low-level individual control can be reduced to single-agent reinforcement learning. In addition to hierarchical reinforcement learning, we propose an opponent modeling network to model other agents’ policies during the learning process. In contrast to end-to-end DRL approaches, our approach reduces the learning complexity by decomposing the overall task into sub-tasks in a hierarchical way. To evaluate the efficiency of our approach, we conduct a real-world case study in the cooperative lane change scenario. Both simulation and real-world experiments show the superiority of our approach in the collision rate and convergence speed.

In Internet of Things (IoT) scenarios such as smart homes, autonomous vehicles, and wearable devices, data pattern changes over time due to changing environments and user requirements, known as domain shifts. When encountering domain shifts, deep neural network models in IoT suffers from performance degradation and need to retrain from scratch to adapt to domain shifts incrementally. Therefore, incremental learning is needed to adapt a model to domain shifts without retraining. Existing methods using the parameter isolation technique perform well in incremental learning of new domains without performance degradation. However, they cannot be directly adopted in IoT applications as they store masks and require users to label the task to indicate task-specific parameters during inference, which is memory inefficient and cumbersome. In this paper, we propose a memory-efficient method for IoT to incrementally adapt to domain shifts in a fixed neural network, named E-DomainIL. Our method freezes learned parameters and allows reusing them later in training to avoid interference between different domains. E-DomainIL does not require task labels or storing masks as it uses all parameters during inference. We use data-driven pruning to adjust the parameter ratio according to the dataset, thus maintaining the balance between accuracy and parameter efficiency. Experimental results on image classification benchmarks demonstrate our method's efficiency and accuracy.

Unsupervised image-to-image translation aims to learn the mapping from an input image in a source domain to an output image in a target domain without paired training dataset. Recently, remarkable progress has been made in translation due to the development of generative adversarial networks (GANs). However, existing methods suffer from the training instability as gradients passing from discriminator to generator become less informative when the source and target domains exhibit sufficiently large discrepancies in appearance or shape. To handle this challenging problem, in this paper, we propose a novel multi-constraint adversarial model (MCGAN) for image translation in which multiple adversarial constraints are applied at generator's multi-scale outputs by a single discriminator to pass gradients to all the scales simultaneously and assist generator training for capturing large discrepancies in appearance between two domains. We further notice that the solution to regularize generator is helpful in stabilizing adversarial training, but results may have unreasonable structure or blurriness due to less context information flow from discriminator to generator. Therefore, we adopt dense combinations of the dilated convolutions at discriminator for supporting more information flow to generator. With extensive experiments on three public datasets, cat-to-dog, horse-to-zebra, and apple-to-orange, our method significantly improves state-of-the-arts on all datasets.

Spatio-temporal (ST) data is a collection of multiple time series data with different spatial locations and is inherently stochastic and unpredictable. An accurate prediction over such data is an important building block for several urban applications, such as taxi demand prediction, traffic flow prediction, and so on. Existing deep learning based approaches assume that outcome is deterministic and there is only one plausible future; therefore, cannot capture the multimodal nature of future contents and dynamics. In addition, existing approaches learn spatial and temporal data separately as they assume weak correlation between them. To handle these issues, in this article, we propose a stochastic spatio-temporal generative model (named D-GAN) which adopts Generative Adversarial Networks (GANs)-based structure for more accurate ST prediction in multiple time steps. D-GAN consists of two components: (1) spatio-temporal correlation network which models spatio-temporal joint distribution of pixels and supports a stochastic sampling of latent variables for multiple plausible futures; (2) a stochastic adversarial network to jointly learn generation and variational inference of data through implicit distribution modeling. D-GAN also supports fusion of external factors through explicit objective to improve the model learning. Extensive experiments performed on two real-world datasets show that D-GAN achieves significant improvements and outperforms baseline models.

Named Data Networking (NDN) is a future Internet paradigm which allows user to retrieve and distribute content using their application names. Each NDN router maintains the state of each request packet in the Pending Interest Table (PIT) until corresponding data packet returns. The use of application name, i.e., variable-length key of unbounded length for communication instead of IP address increases memory consumption and lookup cost at the router. Therefore, the PIT should be able to store millions/billions of entries into on-chip memory. However, traditional hash and trie based methods cannot meet these requirements separately. In this paper, we present a scalable and memory-efficient name encoding based lookup scheme (CRT-PIT) leveraging the benefits of both hash and trie data structures for implementing the PIT at NDN forwarding daemon. In CRT-PIT, we calculate the fixed-length encoded names of the content name and then, encoded names are stored in the concurrent path-compressed trie to reduce the storage and lookup latency requirement by not maintaining the redundant information. Extensive experiments show that CRTPIT consumes only 4.84 MB memory for one million names which is an order of magnitude improvement over the baseline solutions.

Human identification and monitoring are critical in many applications, such as surveillance, evacuation planning. Human identification and monitoring are not an easy task in the case of a large and densely populated crowd. However, none of the existing solutions consider seamless localization, identification, and tracking of the crowd for surveillance in both indoor and outdoor environments with significant accuracy. In this paper, we propose a novel and real-time surveillance system (named, SmartISS) which identifies, tracks and monitors individuals' wireless equipment(s) using their MAC ids. Our trackers/sensing units (PSUs) are the portable entities comprising of Smartphone/Jetson-TK1/PC which are enough to capture users' devices probe requests and locations. PSUs upload collected traces on the cloud server periodically where cloud server keeps finding the suspicious person(s). To retrieve the updated information, we propose an algorithm (named, LLTR) to select the optimal number of PSUs for finding the latest location(s) of the suspicious person(s). To validate and to show the usability of SmartISS, we develop a real prototype testbed and evaluate it extensively on a real-world dataset of 117,121 traces collected during the technical festival held at IIT Roorkee, India. SmartISS selects PSUs with an average selection accuracy of 95.3 percent.

Shared Sensor Network (SSN) refers to a scenario where the same sensing and communication resources are shared and queried by multiple Internet applications. Due to the burgeoning growth in Internet applications, multiple application queries can exhibit overlapping in their functional requirements, such as the region of interest, sensing attributes, and sensing time duration. This overlapping results in redundant sensing tasks generation leading to the increased overall network traffic and energy consumption. Existing approaches operate on data sharing among various tasks to minimize the upstream traffic. However, no existing work attempts to prevent the redundant task generation to reduce the downstream traffic. Moreover, the allocation of suitable sensor nodes to meet the Quality of Service (QoS) requirements of the queries is still an open issue. This article proposes an end-to-end query processing framework (named, QueryPM) that first, calculates the functional requirements similarity among queries to prevent the redundant task generation. Then, it takes the QoS and functional requirements into account while allocating the tasks on the sensor nodes. Extensive simulations on the proposed approach show that downstream traffic, upstream traffic, and energy consumption reduced to 60%, 20--40%, and 40%, respectively, as compared to state-of-the-art mechanisms.

Internet of Things (IoT) is an emerging networking paradigm where smart devices generate, aggregate, and seamlessly exchange data over the predominantly wireless medium. The Internet, so far, has played a significant role in connecting the world, but still, IoT-based solutions are suffering from two primary challenges: 1) how to secure the sensors data and 2) how to provide efficient local and global communication among various heterogeneous devices. Recently, named data networking (NDN), a future Internet paradigm is proposed to improve and simplify such IoT communication issues. NDN allowed users to fetch data by names irrespective of the actual hosting entity connected through a host-specific IP address. NDN well suits the contentcentric pattern of machine-to-machine (M2M) communications predominantly used in IoT. In this paper, we leverage the basic feats of NDN architecture for designing and verification of an NDN-based smart health IoT (NHealthIoT) system. NHealthIoT uses pure-NDN-based M2M communication for capturing and transmission of raw sensor data to the home server which can detect emergency healthcare events using Hidden Markov Model. Emergency events are notified to the cloud server using a novel context-aware adaptive forwarding (Cdf) strategy. Post emergency notifications, and user health information is periodically pulled by the cloud server and by other interested parties using NDN-based publish/subscribe paradigm. The cloud server carries out long-term decision making using probabilistic modeling for detecting the possibility of chronic diseases at the early stage. We extend the workflows intuitive formal approach model for verifying the correctness of NHealthIoT during the emergency. We evaluate the cdf strategy using ndnSIM. Moreover, to validate and to show the usability of NHealthIoT, we develop a proofof-concept prototype testbed and evaluate it extensively. We also identify some research challenges of the NDN-IoT for researchers.