Existing analyses of neural network training often operate under the unrealistic assumption of an extremely small learning rate. This lies in stark contrast to practical wisdom and empirical studies, such as the work of J. Cohen et al. (ICLR 2021), which exhibit startling new phenomena (the "edge of stability" or "unstable convergence") and potential benefits for generalization in the large learning rate regime. Despite a flurry of recent works on this topic, however, the latter effect is still poorly understood. In this paper, we take a step towards understanding genuinely non-convex training dynamics with large learning rates by performing a detailed analysis of gradient descent for simplified models of two-layer neural networks. For these models, we provably establish the edge of stability phenomenon and discover a sharp phase transition for the step size below which the neural network fails to learn "threshold-like" neurons (i.e., neurons with a non-zero first-layer bias). This elucidates one possible mechanism by which the edge of stability can in fact lead to better generalization, as threshold neurons are basic building blocks with useful inductive bias for many tasks.

尽管深度神经网络能够在各个领域中实现最先进的性能，但他们的培训通常需要对数据集的许多通行证进行迭代。但是，由于计算和内存约束和潜在的隐私问题，在数据到达流中的许多现实情况下，存储和访问所有数据都是不切实际的。在本文中，我们研究了一通学习的问题，其中模型是在未重新验证之前对数据进行依次到达数据的培训。通过越来越多参数化模型的使用，我们开发了正交递归拟合（ORFIT），这是一种用于一通学习的算法，旨在完全适合每个新数据点，同时在更改参数的方向上，导致对先前预测的最小变化参数数据点。通过这样做，我们在自适应过滤和机器学习中桥接了两种看似不同的算法，即递归最小二乘（RLS）算法和正交梯度下降（OGD）。我们的算法通过通过增量主组件分析（IPCA）利用流数据的结构来有效地使用内存。此外，我们表明，对于过度参数的线性模型，我们算法获得的参数矢量是随机梯度下降（SGD）在标准的多通用设置中收敛到的。最后，我们将结果推广到高度参数化模型的非线性设置，这与深度学习有关。我们的实验显示了与基准相比，提出的方法的有效性。

（随机）梯度下降的大多数现有分析都取决于$ l $ smorth成本的条件，步骤尺寸小于$ 2/l $。但是，许多作品观察到，在机器学习中，阶梯尺寸通常无法满足这种情况，但（随机）梯度下降仍在收敛，尽管以不稳定的方式。我们从第一原则研究了这种不稳定的收敛现象，并讨论其背后的关键原因。我们还确定了其主要特征，以及它们如何基于理论和实验相互关联，为理解现象提供了有原则的观点。

We initiate a formal study of reproducibility in optimization. We define a quantitative measure of reproducibility of optimization procedures in the face of noisy or error-prone operations such as inexact or stochastic gradient computations or inexact initialization. We then analyze several convex optimization settings of interest such as smooth, non-smooth, and strongly-convex objective functions and establish tight bounds on the limits of reproducibility in each setting. Our analysis reveals a fundamental trade-off between computation and reproducibility: more computation is necessary (and sufficient) for better reproducibility.

Cartoonization is a task that renders natural photos into cartoon styles. Previous deep cartoonization methods only have focused on end-to-end translation, which may hinder editability. Instead, we propose a novel solution with editing features of texture and color based on the cartoon creation process. To do that, we design a model architecture to have separate decoders, texture and color, to decouple these attributes. In the texture decoder, we propose a texture controller, which enables a user to control stroke style and abstraction to generate diverse cartoon textures. We also introduce an HSV color augmentation to induce the networks to generate diverse and controllable color translation. To the best of our knowledge, our work is the first deep approach to control the cartoonization at inference while showing profound quality improvement over to baselines.

Prostate cancer (PCa) is one of the most prevalent cancers in men and many people around the world die from clinically significant PCa (csPCa). Early diagnosis of csPCa in bi-parametric MRI (bpMRI), which is non-invasive, cost-effective, and more efficient compared to multiparametric MRI (mpMRI), can contribute to precision care for PCa. The rapid rise in artificial intelligence (AI) algorithms are enabling unprecedented improvements in providing decision support systems that can aid in csPCa diagnosis and understanding. However, existing state of the art AI algorithms which are based on deep learning technology are often limited to 2D images that fails to capture inter-slice correlations in 3D volumetric images. The use of 3D convolutional neural networks (CNNs) partly overcomes this limitation, but it does not adapt to the anisotropy of images, resulting in sub-optimal semantic representation and poor generalization. Furthermore, due to the limitation of the amount of labelled data of bpMRI and the difficulty of labelling, existing CNNs are built on relatively small datasets, leading to a poor performance. To address the limitations identified above, we propose a new Zonal-aware Self-supervised Mesh Network (Z-SSMNet) that adaptatively fuses multiple 2D, 2.5D and 3D CNNs to effectively balance representation for sparse inter-slice information and dense intra-slice information in bpMRI. A self-supervised learning (SSL) technique is further introduced to pre-train our network using unlabelled data to learn the generalizable image features. Furthermore, we constrained our network to understand the zonal specific domain knowledge to improve the diagnosis precision of csPCa. Experiments on the PI-CAI Challenge dataset demonstrate our proposed method achieves better performance for csPCa detection and diagnosis in bpMRI.

An important challenge in vision-based action recognition is the embedding of spatiotemporal features with two or more heterogeneous modalities into a single feature. In this study, we propose a new 3D deformable transformer for action recognition with adaptive spatiotemporal receptive fields and a cross-modal learning scheme. The 3D deformable transformer consists of three attention modules: 3D deformability, local joint stride, and temporal stride attention. The two cross-modal tokens are input into the 3D deformable attention module to create a cross-attention token with a reflected spatiotemporal correlation. Local joint stride attention is applied to spatially combine attention and pose tokens. Temporal stride attention temporally reduces the number of input tokens in the attention module and supports temporal expression learning without the simultaneous use of all tokens. The deformable transformer iterates L times and combines the last cross-modal token for classification. The proposed 3D deformable transformer was tested on the NTU60, NTU120, FineGYM, and Penn Action datasets, and showed results better than or similar to pre-trained state-of-the-art methods even without a pre-training process. In addition, by visualizing important joints and correlations during action recognition through spatial joint and temporal stride attention, the possibility of achieving an explainable potential for action recognition is presented.

Deep learning-based weather prediction models have advanced significantly in recent years. However, data-driven models based on deep learning are difficult to apply to real-world applications because they are vulnerable to spatial-temporal shifts. A weather prediction task is especially susceptible to spatial-temporal shifts when the model is overfitted to locality and seasonality. In this paper, we propose a training strategy to make the weather prediction model robust to spatial-temporal shifts. We first analyze the effect of hyperparameters and augmentations of the existing training strategy on the spatial-temporal shift robustness of the model. Next, we propose an optimal combination of hyperparameters and augmentation based on the analysis results and a test-time augmentation. We performed all experiments on the W4C22 Transfer dataset and achieved the 1st performance.

Traditional weather forecasting relies on domain expertise and computationally intensive numerical simulation systems. Recently, with the development of a data-driven approach, weather forecasting based on deep learning has been receiving attention. Deep learning-based weather forecasting has made stunning progress, from various backbone studies using CNN, RNN, and Transformer to training strategies using weather observations datasets with auxiliary inputs. All of this progress has contributed to the field of weather forecasting; however, many elements and complex structures of deep learning models prevent us from reaching physical interpretations. This paper proposes a SImple baseline with a spatiotemporal context Aggregation Network (SIANet) that achieved state-of-the-art in 4 parts of 5 benchmarks of W4C22. This simple but efficient structure uses only satellite images and CNNs in an end-to-end fashion without using a multi-model ensemble or fine-tuning. This simplicity of SIANet can be used as a solid baseline that can be easily applied in weather forecasting using deep learning.

This paper proposes a graph-based approach to representing spatio-temporal trajectory data that allows an effective visualization and characterization of city-wide traffic dynamics. With the advance of sensor, mobile, and Internet of Things (IoT) technologies, vehicle and passenger trajectories are being increasingly collected on a massive scale and are becoming a critical source of insight into traffic pattern and traveller behaviour. To leverage such trajectory data to better understand traffic dynamics in a large-scale urban network, this study develops a trajectory-based network traffic analysis method that converts individual trajectory data into a sequence of graphs that evolve over time (known as dynamic graphs or time-evolving graphs) and analyses network-wide traffic patterns in terms of a compact and informative graph-representation of aggregated traffic flows. First, we partition the entire network into a set of cells based on the spatial distribution of data points in individual trajectories, where the cells represent spatial regions between which aggregated traffic flows can be measured. Next, dynamic flows of moving objects are represented as a time-evolving graph, where regions are graph vertices and flows between them are treated as weighted directed edges. Given a fixed set of vertices, edges can be inserted or removed at every time step depending on the presence of traffic flows between two regions at a given time window. Once a dynamic graph is built, we apply graph mining algorithms to detect change-points in time, which represent time points where the graph exhibits significant changes in its overall structure and, thus, correspond to change-points in city-wide mobility pattern throughout the day (e.g., global transition points between peak and off-peak periods).