Vision language models (VLMs) have seen growing adoption in recent years, but many still struggle with basic spatial reasoning errors. We hypothesize that this is due to VLMs adopting pre-trained vision backbones, specifically vision transformers (ViTs) trained with image-level supervision and minimal inductive biases. Such models may fail to encode the class contents at each position in the image, and our goal is to resolve this with a vision backbone that effectively captures both local and global image semantics. Our main insight is that we do not require new supervision to learn this capability – pre-trained models contain significant knowledge of local semantics that we can extract and use for scalable self-supervision. We propose a new efficient post-training stage for ViTs called locality alignment and a novel fine-tuning procedure called MaskEmbed that uses a masked reconstruction loss to learn semantic contributions for each image patch. We first evaluate locality alignment with a vision-only benchmark, finding that it improves a model’s performance at patch-level semantic segmentation, especially for strong backbones trained with image-caption pairs (e.g., CLIP and SigLIP). We then train a series of VLMs with and without locality alignment, and show that locality-aligned backbones improve performance across a range of benchmarks, particularly ones that involve spatial understanding (e.g., RefCOCO, OCID-Ref, TallyQA, VSR, AI2D). Overall, we demonstrate that we can efficiently learn local semantic extraction via a locality alignment stage, and that this procedure benefits VLM training recipes that use off-the-shelf vision backbones.

Neural text summarization has shown great potential in recent years. However, current state-of-the-art summarization models are limited by their maximum input length, posing a challenge to summarizing longer texts comprehensively. As part of a layered summarization architecture, we introduce PureText, a simple yet effective pre-processing layer that removes low- quality sentences in articles to improve existing summarization models. When evaluated on popular datasets like WikiHow and Reddit TIFU, we show up to 3.84 and 8.57 point ROUGE-1 absolute improvement on the full test set and the long article subset, respectively, for state-of-the-art summarization models such as BertSum and BART. Our approach provides downstream models with higher-quality sentences for summarization, improving overall model performance, especially on long text articles.

Andrew Gaut, Tony Sun, Shirlyn Tang, Yuxin Huang, Jing Qian, Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-Wei Chang, William Yang Wang. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020.

Tony Sun, Andrew Gaut, Shirlyn Tang, Yuxin Huang, Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-Wei Chang, William Yang Wang. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019.

Opportunistic network is a type of delay tolerant networks (DTN) where network communication opportunities appear opportunistic. In this study, we investigate opportunistic network scenarios based on public network traces, and our contributions are the following: First, we identify the censorship issue in network traces that usually leads to strongly skewed distribution of the measurements. Based on this knowledge, we then apply the Kaplan-Meier Estimator to calculate the survivorship of network measurements, which is used in designing our proposed censorship removal algorithm (CRA) that is used to recover censored data. Second, we perform a rich set of analysis illustrating that UCSD and Dartmouth network traces show strong self-similarity, and can be modeled as such. Third, we pointed out the importance of these newly revealed characteristics in future development and evaluation of opportunistic networks.

We demonstrate a seamless vertical handoff solution, called Universal Seamless Handoff Architecture (USHA). USHA is simple and requires minimal modification to the current Internet infrastructure. Therefore, it is instantly ready for realworld deployment. In this demonstration, we present USHA in two scenarios. Using video streaming applications, we demonstrate that USHA is able to successfully maintain the application connectivity and achieve almost zero delay during a vertical handoff. Moreover, we present a vertical handoff detection technique based on the end-to-end link capacity monitoring. The ongoing work of this study is to improve the accuracy of the handoff detection and to enhance application QoS support for vertical handoffs given the accurate handoff detection is provided.

In this demonstration, we will showcase SenProbe, a lightweight capacity estimation technique specifically designed for the CSMA based wireless sensor networks. SenProbe is a packet train technique based on the concept pioneer by CapProbe, but designed specially for the wireless environment. We will use wireless devices with limited radio ranges to display the properties of multi-hop wireless networks (a table top experiment). More specifically, we will illustrate wireless properties with IEEE 802.15.4 enabled devices, and offer insights into how the capacity of a wireless path changes in real wireless environments can be deployed, measured, and used. This tool will be useful for network users as well as network designers to gain better understanding of their network, and plan their activities accordingly. This demonstration aims to provide results that can be of assistance in various facets of IEEE 802.15.4 enabled intelligent home/industrial networking environments.

With the increasingly deployed Wireless Personal Area Network (WPAN) devices, channel conflict has become very frequent and severe when one WPAN technology coexists with other WPAN technologies in the same interfering range. In this paper, we study the coexistence issue between various IEEE 802.15 based WPAN technologies. We present analytical models on the non-conflicting channel allocation probabilities, focusing on the coexistence scenarios of one WPAN technology coexisting with another. The results show that channel allocation conflicts occurs frequently in all cases, and is especially severe between IEEE 802.15.3 and IEEE 802.15.4 networks. On the other hand, the probability of non-conflict channel allocation is less dramatic between a single IEEE 802.15.1 and coexisting IEEE 802.15.4 networks. In addition, the proposed models in this paper are also applicable to other wireless technologies, as long as the channel allocation mechanisms are known.

Mobile ad-hoc networks (MANETs) are most useful in unprepared emergencies where critical applications must be launched quickly. However, they often operate in an adverse environment where end-to-end connectivity is highly susceptible to disruption. Adjusting the motion of existing nodes or deploying additional nodes can improve the connectivity under some circumstances, but for scenarios where connectivity cannot be immediately improved, disruption must be coped with properly. In this paper we propose the ad-hoc storage overlay system (ASOS). ASOS is a self-organized overlay of storage-abundant nodes to jointly provide distributed and reliable storage to data flows under disruption. ASOS is a delay-tolerant networking (DTN) approach that significantly improves the applicability of MANETs in practice

Based on the IEEE 802.15.4 LR-WPAN standard, the ZigBee standard has been proposed to interconnect simple, low rate, and battery powered wireless devices. The deployment of ZigBee networks is expected to facilitate numerous applications, such as home-appliance networks, home healthcare, medical monitoring, consumer electronics, and environmental sensors. An effective routing scheme in a ZigBee network is particularly important in that it is the key to achieve resource (e.g., bandwidth and energy) efficiency in ZigBee networks. Routing in a ZigBee network is not exactly the same as in a MANET. In particular, while full function devices (FFD) can serve as network coordinators or network routers, reduced function devices (RFD) can only associate and communicate with FFDs in a ZigBee network. Therefore, different from traditional MANET routing algorithms, which only take into account node mobility to figure out a best route to a given destination, node heterogeneity plays an important role in ZigBee network routing. In this paper, we perform extensive evaluation, using NS-2 simulator, to study the impact of node heterogeneity on ZigBee mesh network routing. The results show that the ZigBee mesh routing algorithm exhibits significant performance difference when the network is highly heterogenous. We also reveal that the node type and the role of the node plays a critical role in deciding routing performances.

IEEE 802.15.4 is an emerging wireless standard addressing the needs of low-rate wireless personal area networks with a focus on enabling various pervasive and ubiquitous applications that require interactions with our surrounding environments. In view of the application potential of IEEE 802.15.4, knowing the fundamental network properties soon becomes essential in fasten the interactivity between these devices. Among all, knowing effective capacity of a path in wireless networks is of particular importance in routing and traffic management. In this paper, we implement SenProbe, a recently proposed path capacity estimation tool specially designed for the multi-hop ad hoc wireless environment. We present an implementation of SenProbe in sensor operating system (SOS), and evaluate the behavior/effectiveness of SenProbe in various testbed setups; including an interfered setting that cannot be simulated. Experiment results validate the workings of SenProbe and offer insights into how the capacity of a wireless path changes in real wireless environments. Our efforts provide a basis for realistic results that can be of assistance in activities such as capacity planning, protocol design, performance analysis, and etc

In this article we study a smooth and efficient transport protocol for real-time video over wireless networks. The proposed scheme, named the video transport protocol (VTP), has a new and unique end-to-end rate control mechanism that aims to avoid drastic rate fluctuations while maintaining friendliness to legacy protocols. VTP is also equipped with an achieved rate estimation scheme and a loss discrimination algorithm, both end-to-end, to cope with random errors in wireless networks efficiently. We show by analysis that VTP preserves most of the convergence properties of AIMD and converges to its fair share fast. VTP is compared to two recent TCP friendly rate control (TFRC) extensions, namely TFRC Wireless and MULTFRC, in wired-cum-wireless scenarios in Ns-2. Results show that VTP excels in all tested scenarios in terms of smoothness, fairness, and opportunistic friendliness. VTP is also implemented to work with a video camera and an H.263 video codec as part of our hybrid testbed, where its good performance as a transport layer protocol is confirmed by measurement results.

Real-time streaming over wireless links is challenging. The streaming protocol must be efficient and robust to random wireless loss, fair to itself, and friendly to legacy TCP. Various solutions have been proposed in the literature, among which the video transport protocol (VTP), TFRC wireless, and MULTFRC are end-to-end representatives. In this paper we provide an in-depth comparison on the performance of VTP, TFRC wireless, and MULTFRC in various wireless scenarios. The results show that VTP and TFRC wireless both perform well and deliver similar performance, with VTP exhibiting greater efficiency and smoothness in presence of heavy errors. In contrast, MULTFRC performs less satisfactorily, as it experiences large rate fluctuation and slow convergence caused by the frequent changes in the number of simultaneous connections.

Vertical handoffs between different wireless technologies usually lead to dramatic changes in the link capacity. A successful QoS solution for vertical handoffs must be able to fast track the capacity changes and agilely adapt the delivery rates and qualities of the ongoing applications. Though traditional AIMD-based source adaptation schemes (as found in TCP, TFRC, etc.) have been well designed for mild, gradual rate adjustments required by load fluctuations and network congestion, their response time is inadequate when the rate must be adjusted to the drastic network capacity changes that are typical in vertical handoff scenarios. To expedite the response to such changes, we propose in this paper two adaptive algorithms, named the fast rate adaptation (FRA) and early rate reduction (ERR), that are launched when the handoff is from low to high capacity (LOW-to-HIGH) or from high to low capacity (HIGH-to-LOW), respectively. We also propose two vertical handoff notification mechanisms to work with FRA and/or ERR, i.e. the implicit handoff notification (IHN) and explicit handoff notification (EHN). We show by simulation that our proposed schemes are able to provide better QoS support than the traditional AIMD based schemes during vertical handoffs.

Wireless communication is inherently vulnerable to errors from the dynamic wireless environment. Link layer packets discarded due to these errors impose a serious limitation on the maximum achievable throughput in the wireless channel. To enhance the overall throughput of wireless communication, it is necessary to have a link layer transmission scheme that is robust to the errors intrinsic to the wireless channel. To this end, we present interleaved-forward error correction (I-FEC), a clever link layer coding scheme that protects link layer data against random and busty errors. We examine the level of data protection provided by I-FEC against other popular schemes. We also simulate I-FEC in Bluetooth, and compare the TCP throughput result with Bluetooth's integrated FEC coding feature. We show that I-FEC consistently and significantly outperforms other link layer coding schemes by providing an impressive amount of protection against heavy channel burst errors.

Video streaming has become a popular form of transferring video over the Internet. With the emergence of mobile computing needs, a successful video streaming solution demands 1) uninterrupted services even with the presence of mobility and 2) adaptive video delivery according to current link properties. We study the need and evaluate the performance of adaptive video streaming in vertical handoff scenarios. We use universal seamless handoff architecture (USHA) to create a seamless handoff environment, and use the video transfer protocol (VTP) to adapt video streaming rates according to "eligible rate estimates". Using testbed measurements experiments, we verify the importance of service adaptation, as well as show the improvement of user-perceived video quality, via adapting video streaming in the vertical handoffs.

We explain that decomposition methods, in particular, SMO-type algorithms, are not suitable for linear SVMs with more data than attributes. To remedy this difficulty, we consider a recent result by S.S. Keerthi and C.-J. Lin (see http://www.csie.ntu.edu.tw//spl sim/cjlin/papers/limit.ps.gz, 2002) that for an SVM which is not linearly separable, after C is large enough, the dual solutions are at similar faces. Motivated by this property, we show that alpha seeding is extremely useful for solving a sequence of linear SVMs. It largely reduces the number of decomposition iterations to the point that solving many linear SVMs requires less time than the original decomposition method for one single SVM. We also conduct comparisons with other methods which are efficient for linear SVMs, and demonstrate the effectiveness of the proposed approach for helping the model selection.