随着深度神经网络（DNN）的出现，成为许多计算机视觉任务中的骨干，它们在现实世界中的消费应用程序中的采用不断扩大。鉴于智能设备的丰富性和无所不能，正在形成“智能生态系统”，同时进行感应而不是独立。这将处式推理范式转移到在边缘部署集中式神经加工单元（NPU），其中多个设备（例如，在智能家居或自动驾驶汽车中）可以通过动态速率流式传输数据以进行处理。尽管这为输入批处理提供了增强的潜力，但幼稚的解决方案可以导致表现不佳的性能和经验质量，尤其是在尖峰负载下。同时，动态DNN的部署，包括随机计算图（例如早期 - 外观（EE）模型），引入了此类系统中动态行为的新维度。在这项工作中，我们提出了一种新颖的早期感知的调度算法，该算法允许在运行时进行样本抢占，以说明到达和早期外来过程引入的动态性。同时，我们向NPU硬件体系结构的设计空间介绍了两个新颖的维度，即流体批处理和可堆叠的处理元素，这些元素可以使运行时适应性适应不同的批次尺寸，并显着改善了NPU利用率，即使在小批次尺寸下也是如此。我们的评估表明，我们的系统分别在平均延迟和尾部潜伏期SLO满意度方面，平均达到1.97倍和6.7倍的改善。

基于注意力的神经网络在许多AI任务中都普遍存在。尽管其出色的算法性能，但注意力机制和前馈网络（FFN）的使用仍需要过多的计算和内存资源，这通常会损害其硬件性能。尽管已经引入了各种稀疏变体，但大多数方法仅着重于缓解算法级别上的二次注意力缩放，而无需明确考虑将其方法映射到真实硬件设计上的效率。此外，大多数努力仅专注于注意机制或FFN，但没有共同优化这两个部分，导致当前的大多数设计在处理不同的输入长度时缺乏可扩展性。本文从硬件角度系统地考虑了不同变体中的稀疏模式。在算法级别上，我们提出了Fabnet，这是一种适合硬件的变体，它采用统一的蝴蝶稀疏模式来近似关注机制和FFN。在硬件级别上，提出了一种新颖的适应性蝴蝶加速器，可以在运行时通过专用硬件控件配置，以使用单个统一的硬件引擎加速不同的蝴蝶层。在远程 - ARENA数据集上，FabNet达到了与香草变压器相同的精度，同时将计算量减少10到66次，参数数量为2至22次。通过共同优化算法和硬件，我们的基于FPGA的蝴蝶加速器在归一化到同一计算预算的最新加速器上达到了14.2至23.2倍的速度。与Raspberry Pi 4和Jetson Nano上优化的CPU和GPU设计相比，我们的系统在相同的功率预算下的最大273.8和15.1倍。

语义细分是许多视觉系统的骨干，从自动驾驶汽车和机器人导航到增强现实和电信。在有限的资源信封内经常在严格的延迟约束下运行，对有效执行的优化变得很重要。同时，目标平台的异质功能以及不同应用程序的不同限制需要设计和培训多个针对特定目标的细分模型，从而导致过度维护成本。为此，我们提出了一个框架，用于将最新的分割CNN转换为多EXIT语义细分（MESS）网络：经过特殊训练的模型，这些模型沿其深度沿其深度进行参数化的早期出口到i）在推断过程中动态保存计算更容易的样本和ii）通过提供可定制的速度准确性权衡来节省培训和维护成本。设计和培训此类网络天真地损害了性能。因此，我们为多EXIT网络提出了新颖的两期培训方案。此外，Mess的参数化可以使附件分割头的数字，位置和体系结构以及退出策略通过详尽的搜索在<1GPUH中进行部署。这使得混乱能够快速适应每个目标用例的设备功能和应用要求，并提供火车一路上的部署解决方案。与原始的骨干网络相比，Mess变体具有相同精度的潜伏期增长率高达2.83倍，而相同的计算预算的潜伏期提高到同一计算预算的准确性高5.33 pp。最后，与最先进的技术相比，MESS提供了更快的架构选择订单。

Modern speech recognition systems exhibits rapid performance degradation under domain shift. This issue is especially prevalent in data-scarce settings, such as low-resource languages, where diversity of training data is limited. In this work we propose M2DS2, a simple and sample-efficient finetuning strategy for large pretrained speech models, based on mixed source and target domain self-supervision. We find that including source domain self-supervision stabilizes training and avoids mode collapse of the latent representations. For evaluation, we collect HParl, a $120$ hour speech corpus for Greek, consisting of plenary sessions in the Greek Parliament. We merge HParl with two popular Greek corpora to create GREC-MD, a test-bed for multi-domain evaluation of Greek ASR systems. In our experiments we find that, while other Unsupervised Domain Adaptation baselines fail in this resource-constrained environment, M2DS2 yields significant improvements for cross-domain adaptation, even when a only a few hours of in-domain audio are available. When we relax the problem in a weakly supervised setting, we find that independent adaptation for audio using M2DS2 and language using simple LM augmentation techniques is particularly effective, yielding word error rates comparable to the fully supervised baselines.

The evolution of wireless communications into 6G and beyond is expected to rely on new machine learning (ML)-based capabilities. These can enable proactive decisions and actions from wireless-network components to sustain quality-of-service (QoS) and user experience. Moreover, new use cases in the area of vehicular and industrial communications will emerge. Specifically in the area of vehicle communication, vehicle-to-everything (V2X) schemes will benefit strongly from such advances. With this in mind, we have conducted a detailed measurement campaign with the purpose of enabling a plethora of diverse ML-based studies. The resulting datasets offer GPS-located wireless measurements across diverse urban environments for both cellular (with two different operators) and sidelink radio access technologies, thus enabling a variety of different studies towards V2X. The datasets are labeled and sampled with a high time resolution. Furthermore, we make the data publicly available with all the necessary information to support the on-boarding of new researchers. We provide an initial analysis of the data showing some of the challenges that ML needs to overcome and the features that ML can leverage, as well as some hints at potential research studies.

Designing powerful adversarial attacks is of paramount importance for the evaluation of $\ell_p$-bounded adversarial defenses. Projected Gradient Descent (PGD) is one of the most effective and conceptually simple algorithms to generate such adversaries. The search space of PGD is dictated by the steepest ascent directions of an objective. Despite the plethora of objective function choices, there is no universally superior option and robustness overestimation may arise from ill-suited objective selection. Driven by this observation, we postulate that the combination of different objectives through a simple loss alternating scheme renders PGD more robust towards design choices. We experimentally verify this assertion on a synthetic-data example and by evaluating our proposed method across 25 different $\ell_{\infty}$-robust models and 3 datasets. The performance improvement is consistent, when compared to the single loss counterparts. In the CIFAR-10 dataset, our strongest adversarial attack outperforms all of the white-box components of AutoAttack (AA) ensemble, as well as the most powerful attacks existing on the literature, achieving state-of-the-art results in the computational budget of our study ($T=100$, no restarts).

We present RAVEn, a self-supervised multi-modal approach to jointly learn visual and auditory speech representations. Our pre-training objective involves encoding masked inputs, and then predicting contextualised targets generated by slowly-evolving momentum encoders. Driven by the inherent differences between video and audio, our design is asymmetric w.r.t. the two modalities' pretext tasks: Whereas the auditory stream predicts both the visual and auditory targets, the visual one predicts only the auditory targets. We observe strong results in low- and high-resource labelled data settings when fine-tuning the visual and auditory encoders resulting from a single pre-training stage, in which the encoders are jointly trained. Notably, RAVEn surpasses all self-supervised methods on visual speech recognition (VSR) on LRS3, and combining RAVEn with self-training using only 30 hours of labelled data even outperforms a recent semi-supervised method trained on 90,000 hours of non-public data. At the same time, we achieve state-of-the-art results in the LRS3 low-resource setting for auditory speech recognition (as well as for VSR). Our findings point to the viability of learning powerful speech representations entirely from raw video and audio, i.e., without relying on handcrafted features. Code and models will be made public.

Multimodal learning pipelines have benefited from the success of pretrained language models. However, this comes at the cost of increased model parameters. In this work, we propose Adapted Multimodal BERT (AMB), a BERT-based architecture for multimodal tasks that uses a combination of adapter modules and intermediate fusion layers. The adapter adjusts the pretrained language model for the task at hand, while the fusion layers perform task-specific, layer-wise fusion of audio-visual information with textual BERT representations. During the adaptation process the pre-trained language model parameters remain frozen, allowing for fast, parameter-efficient training. In our ablations we see that this approach leads to efficient models, that can outperform their fine-tuned counterparts and are robust to input noise. Our experiments on sentiment analysis with CMU-MOSEI show that AMB outperforms the current state-of-the-art across metrics, with 3.4% relative reduction in the resulting error and 2.1% relative improvement in 7-class classification accuracy.

We consider the problem of modelling high-dimensional distributions and generating new examples of data with complex relational feature structure coherent with a graph skeleton. The model we propose tackles the problem of generating the data features constrained by the specific graph structure of each data point by splitting the task into two phases. In the first it models the distribution of features associated with the nodes of the given graph, in the second it complements the edge features conditionally on the node features. We follow the strategy of implicit distribution modelling via generative adversarial network (GAN) combined with permutation equivariant message passing architecture operating over the sets of nodes and edges. This enables generating the feature vectors of all the graph objects in one go (in 2 phases) as opposed to a much slower one-by-one generations of sequential models, prevents the need for expensive graph matching procedures usually needed for likelihood-based generative models, and uses efficiently the network capacity by being insensitive to the particular node ordering in the graph representation. To the best of our knowledge, this is the first method that models the feature distribution along the graph skeleton allowing for generations of annotated graphs with user specified structures. Our experiments demonstrate the ability of our model to learn complex structured distributions through quantitative evaluation over three annotated graph datasets.

We extend Textual Inversion to learn pseudo-words that represent a concept at different resolutions. This allows us to generate images that use the concept with different levels of detail and also to manipulate different resolutions using language. Once learned, the user can generate images at different levels of agreement to the original concept; "A photo of $S^*(0)$" produces the exact object while the prompt "A photo of $S^*(0.8)$" only matches the rough outlines and colors. Our framework allows us to generate images that use different resolutions of an image (e.g. details, textures, styles) as separate pseudo-words that can be composed in various ways. We open-soure our code in the following URL: https://github.com/giannisdaras/multires_textual_inversion