我们与压缩神经语音增强（SE）下降到最佳配置基于神经加速器的新一代低功耗微控制器的目的（microNPU的）探讨网络稀疏化战略。我们研究了三个独特的稀疏结构：重修剪，修剪块和单位修剪;当应用于SE讨论他们的优点和缺点。我们专注于计算吞吐量，内存占用和模型质量之间的相互作用。我们的方法支持所有三种结构之上，共同学习整数，稀疏一起量化权重。此外，我们证明整数离线幅度修剪基于量化模型的性能基准。虽然高效的语音增强是一个活跃的研究领域，我们的工作是第一个块修剪适用于SE和第一个地址SE车型压缩microNPU的背景下。使用重修剪，我们表明，我们能够通过的42X的一个因素，从3.7MB压缩本已紧凑型的内存占用87KB，而只有在性能上失去0.1分贝SDR。我们也表现出与6.7倍使用块修剪只0.59分贝SDR的相应的SDR下降的计算加速。

Graph Neural Networks (GNNs) have been widely applied to different tasks such as bioinformatics, drug design, and social networks. However, recent studies have shown that GNNs are vulnerable to adversarial attacks which aim to mislead the node or subgraph classification prediction by adding subtle perturbations. Detecting these attacks is challenging due to the small magnitude of perturbation and the discrete nature of graph data. In this paper, we propose a general adversarial edge detection pipeline EDoG without requiring knowledge of the attack strategies based on graph generation. Specifically, we propose a novel graph generation approach combined with link prediction to detect suspicious adversarial edges. To effectively train the graph generative model, we sample several sub-graphs from the given graph data. We show that since the number of adversarial edges is usually low in practice, with low probability the sampled sub-graphs will contain adversarial edges based on the union bound. In addition, considering the strong attacks which perturb a large number of edges, we propose a set of novel features to perform outlier detection as the preprocessing for our detection. Extensive experimental results on three real-world graph datasets including a private transaction rule dataset from a major company and two types of synthetic graphs with controlled properties show that EDoG can achieve above 0.8 AUC against four state-of-the-art unseen attack strategies without requiring any knowledge about the attack type; and around 0.85 with knowledge of the attack type. EDoG significantly outperforms traditional malicious edge detection baselines. We also show that an adaptive attack with full knowledge of our detection pipeline is difficult to bypass it.

Learning on Graphs (LoG) is widely used in multi-client systems when each client has insufficient local data, and multiple clients have to share their raw data to learn a model of good quality. One scenario is to recommend items to clients with limited historical data and sharing similar preferences with other clients in a social network. On the other hand, due to the increasing demands for the protection of clients' data privacy, Federated Learning (FL) has been widely adopted: FL requires models to be trained in a multi-client system and restricts sharing of raw data among clients. The underlying potential data-sharing conflict between LoG and FL is under-explored and how to benefit from both sides is a promising problem. In this work, we first formulate the Graph Federated Learning (GFL) problem that unifies LoG and FL in multi-client systems and then propose sharing hidden representation instead of the raw data of neighbors to protect data privacy as a solution. To overcome the biased gradient problem in GFL, we provide a gradient estimation method and its convergence analysis under the non-convex objective. In experiments, we evaluate our method in classification tasks on graphs. Our experiment shows a good match between our theory and the practice.

Multivariate time series forecasting constitutes important functionality in cyber-physical systems, whose prediction accuracy can be improved significantly by capturing temporal and multivariate correlations among multiple time series. State-of-the-art deep learning methods fail to construct models for full time series because model complexity grows exponentially with time series length. Rather, these methods construct local temporal and multivariate correlations within subsequences, but fail to capture correlations among subsequences, which significantly affect their forecasting accuracy. To capture the temporal and multivariate correlations among subsequences, we design a pattern discovery model, that constructs correlations via diverse pattern functions. While the traditional pattern discovery method uses shared and fixed pattern functions that ignore the diversity across time series. We propose a novel pattern discovery method that can automatically capture diverse and complex time series patterns. We also propose a learnable correlation matrix, that enables the model to capture distinct correlations among multiple time series. Extensive experiments show that our model achieves state-of-the-art prediction accuracy.

Riemannian geometry provides powerful tools to explore the latent space of generative models while preserving the inherent structure of the data manifold. Lengths, energies and volume measures can be derived from a pullback metric, defined through the immersion that maps the latent space to the data space. With this in mind, most generative models are stochastic, and so is the pullback metric. Manipulating stochastic objects is strenuous in practice. In order to perform operations such as interpolations, or measuring the distance between data points, we need a deterministic approximation of the pullback metric. In this work, we are defining a new metric as the expected length derived from the stochastic pullback metric. We show this metric is Finslerian, and we compare it with the expected pullback metric. In high dimensions, we show that the metrics converge to each other at a rate of $\mathcal{O}\left(\frac{1}{D}\right)$.

Pretrained large-scale vision-language models like CLIP have exhibited strong generalization over unseen tasks. Yet imperceptible adversarial perturbations can significantly reduce CLIP's performance on new tasks. In this work, we identify and explore the problem of \emph{adapting large-scale models for zero-shot adversarial robustness}. We first identify two key factors during model adaption -- training losses and adaptation methods -- that affect the model's zero-shot adversarial robustness. We then propose a text-guided contrastive adversarial training loss, which aligns the text embeddings and the adversarial visual features with contrastive learning on a small set of training data. We apply this training loss to two adaption methods, model finetuning and visual prompt tuning. We find that visual prompt tuning is more effective in the absence of texts, while finetuning wins in the existence of text guidance. Overall, our approach significantly improves the zero-shot adversarial robustness over CLIP, seeing an average improvement of over 31 points over ImageNet and 15 zero-shot datasets. We hope this work can shed light on understanding the zero-shot adversarial robustness of large-scale models.

Many visual recognition models are evaluated only on their classification accuracy, a metric for which they obtain strong performance. In this paper, we investigate whether computer vision models can also provide correct rationales for their predictions. We propose a ``doubly right'' object recognition benchmark, where the metric requires the model to simultaneously produce both the right labels as well as the right rationales. We find that state-of-the-art visual models, such as CLIP, often provide incorrect rationales for their categorical predictions. However, by transferring the rationales from language models into visual representations through a tailored dataset, we show that we can learn a ``why prompt,'' which adapts large visual representations to produce correct rationales. Visualizations and empirical experiments show that our prompts significantly improve performance on doubly right object recognition, in addition to zero-shot transfer to unseen tasks and datasets.

Deep networks for computer vision are not reliable when they encounter adversarial examples. In this paper, we introduce a framework that uses the dense intrinsic constraints in natural images to robustify inference. By introducing constraints at inference time, we can shift the burden of robustness from training to the inference algorithm, thereby allowing the model to adjust dynamically to each individual image's unique and potentially novel characteristics at inference time. Among different constraints, we find that equivariance-based constraints are most effective, because they allow dense constraints in the feature space without overly constraining the representation at a fine-grained level. Our theoretical results validate the importance of having such dense constraints at inference time. Our empirical experiments show that restoring feature equivariance at inference time defends against worst-case adversarial perturbations. The method obtains improved adversarial robustness on four datasets (ImageNet, Cityscapes, PASCAL VOC, and MS-COCO) on image recognition, semantic segmentation, and instance segmentation tasks. Project page is available at equi4robust.cs.columbia.edu.

Incidental supervision from language has become a popular approach for learning generic visual representations that can be prompted to perform many recognition tasks in computer vision. We conduct an in-depth exploration of the CLIP model and show that its visual representation is often strongly biased towards solving some tasks more than others. Moreover, which task the representation will be biased towards is unpredictable, with little consistency across images. To resolve this task bias, we show how to learn a visual prompt that guides the representation towards features relevant to their task of interest. Our results show that these visual prompts can be independent of the input image and still effectively provide a conditioning mechanism to steer visual representations towards the desired task.

Small differences in a person's motion can engage drastically different muscles. While most visual representations of human activity are trained from video, people learn from multimodal experiences, including from the proprioception of their own muscles. We present a new visual perception task and dataset to model muscle activation in human activities from monocular video. Our Muscles in Action (MIA) dataset consists of 2 hours of synchronized video and surface electromyography (sEMG) data of subjects performing various exercises. Using this dataset, we learn visual representations that are predictive of muscle activation from monocular video. We present several models, including a transformer model, and measure their ability to generalize to new exercises and subjects. Putting muscles into computer vision systems will enable richer models of virtual humans, with applications in sports, fitness, and AR/VR.