We report on a significant discovery linking deep convolutional neural networks (CNN) to biological vision and fundamental particle physics. A model of information propagation in a CNN is proposed via an analogy to an optical system, where bosonic particles (i.e. photons) are concentrated as the 2D spatial resolution of the image collapses to a focal point $1\times 1=1$. A 3D space $(x,y,t)$ is defined by $(x,y)$ coordinates in the image plane and CNN layer $t$, where a principal ray $(0,0,t)$ runs in the direction of information propagation through both the optical axis and the image center pixel located at $(x,y)=(0,0)$, about which the sharpest possible spatial focus is limited to a circle of confusion in the image plane. Our novel insight is to model the principal optical ray $(0,0,t)$ as geometrically equivalent to the medial vector in the positive orthant $I(x,y) \in R^{N+}$ of a $N$-channel activation space, e.g. along the greyscale (or luminance) vector $(t,t,t)$ in $RGB$ colour space. Information is thus concentrated into an energy potential $E(x,y,t)=\|I(x,y,t)\|^2$, which, particularly for bottleneck layers $t$ of generic CNNs, is highly concentrated and symmetric about the spatial origin $(0,0,t)$ and exhibits the well-known "Sombrero" potential of the boson particle. This symmetry is broken in classification, where bottleneck layers of generic pre-trained CNN models exhibit a consistent class-specific bias towards an angle $\theta \in U(1)$ defined simultaneously in the image plane and in activation feature space. Initial observations validate our hypothesis from generic pre-trained CNN activation maps and a bare-bones memory-based classification scheme, with no training or tuning. Training from scratch using a random $U(1)$ class label the leads to improved classification in all cases.
相關內容
Deep neural networks perform well on prediction and classification tasks in the canonical setting where data streams are i.i.d., labeled data is abundant, and class labels are balanced. Challenges emerge with distribution shifts, including non-stationary or imbalanced data streams. One powerful approach that has addressed this challenge involves self-supervised pretraining of large encoders on volumes of unlabeled data, followed by task-specific tuning. Given a new task, updating the weights of these encoders is challenging as a large number of weights needs to be fine-tuned, and as a result, they forget information about the previous tasks. In the present work, we propose a model architecture to address this issue, building upon a discrete bottleneck containing pairs of separate and learnable (key, value) codes. In this setup, we follow the encode; process the representation via a discrete bottleneck; and decode paradigm, where the input is fed to the pretrained encoder, the output of the encoder is used to select the nearest keys, and the corresponding values are fed to the decoder to solve the current task. The model can only fetch and re-use a limited number of these (key, value) pairs during inference, enabling localized and context-dependent model updates. We theoretically investigate the ability of the proposed model to minimize the effect of the distribution shifts and show that such a discrete bottleneck with (key, value) pairs reduces the complexity of the hypothesis class. We empirically verified the proposed methods' benefits under challenging distribution shift scenarios across various benchmark datasets and show that the proposed model reduces the common vulnerability to non-i.i.d. and non-stationary training distributions compared to various other baselines.
Traffic speed prediction is the key to many valuable applications, and it is also a challenging task because of its various influencing factors. Recent work attempts to obtain more information through various hybrid models, thereby improving the prediction accuracy. However, the spatial information acquisition schemes of these methods have two-level differentiation problems. Either the modeling is simple but contains little spatial information, or the modeling is complete but lacks flexibility. In order to introduce more spatial information on the basis of ensuring flexibility, this paper proposes IRNet (Transferable Intersection Reconstruction Network). First, this paper reconstructs the intersection into a virtual intersection with the same structure, which simplifies the topology of the road network. Then, the spatial information is subdivided into intersection information and sequence information of traffic flow direction, and spatiotemporal features are obtained through various models. Third, a self-attention mechanism is used to fuse spatiotemporal features for prediction. In the comparison experiment with the baseline, not only the prediction effect, but also the transfer performance has obvious advantages.
Although deep-learning based methods for monocular pedestrian detection have made great progress, they are still vulnerable to heavy occlusions. Using multi-view information fusion is a potential solution but has limited applications, due to the lack of annotated training samples in existing multi-view datasets, which increases the risk of overfitting. To address this problem, a data augmentation method is proposed to randomly generate 3D cylinder occlusions, on the ground plane, which are of the average size of pedestrians and projected to multiple views, to relieve the impact of overfitting in the training. Moreover, the feature map of each view is projected to multiple parallel planes at different heights, by using homographies, which allows the CNNs to fully utilize the features across the height of each pedestrian to infer the locations of pedestrians on the ground plane. The proposed 3DROM method has a greatly improved performance in comparison with the state-of-the-art deep-learning based methods for multi-view pedestrian detection.
We introduce a neural implicit framework that exploits the differentiable properties of neural networks and the discrete geometry of point-sampled surfaces to approximate them as the level sets of neural implicit functions. To train a neural implicit function, we propose a loss functional that approximates a signed distance function, and allows terms with high-order derivatives, such as the alignment between the principal directions of curvature, to learn more geometric details. During training, we consider a non-uniform sampling strategy based on the curvatures of the point-sampled surface to prioritize points with more geometric details. This sampling implies faster learning while preserving geometric accuracy when compared with previous approaches. We also present the analytical differential geometry formulas for neural surfaces, such as normal vectors and curvatures.
Differential privacy is known to protect against threats to validity incurred due to adaptive, or exploratory, data analysis -- even when the analyst adversarially searches for a statistical estimate that diverges from the true value of the quantity of interest on the underlying population. The cost of this protection is the accuracy loss incurred by differential privacy. In this work, inspired by standard models in the genomics literature, we consider data models in which individuals are represented by a sequence of attributes with the property that where distant attributes are only weakly correlated. We show that, under this assumption, it is possible to "re-use" privacy budget on different portions of the data, significantly improving accuracy without increasing the risk of overfitting.
Calculus of Variation combined with Differential Geometry as tools of modelling and solving problems in image processing and computer vision were introduced in the late 80's and the 90s of the 20th century. The beginning of an extensive work in these directions was marked by works such as Geodesic Active Contours (GAC), the Beltrami framework, level set method of Osher and Sethian the works of Charpiat et al. and the works by Chan and Vese to name just a few. In many cases the optimization of these functional are done by the gradient descent method via the calculation of the Euler-Lagrange equations. Straightforward use of the resulted EL equations in the gradient descent scheme leads to non-geometric and in some cases non sensical equations. It is costumary to modify these EL equations or even the functional itself in order to obtain geometric and/or sensical equations. The aim of this note is to point to the correct way to derive the EL and the gradient descent equations such that the resulted gradient descent equation is geometric and makes sense.
As the importance of intrusion detection and prevention systems (IDPSs) increases, great costs are incurred to manage the signatures that are generated by malicious communication pattern files. Experts in network security need to classify signatures by importance for an IDPS to work. We propose and evaluate a machine learning signature classification model with a reject option (RO) to reduce the cost of setting up an IDPS. To train the proposed model, it is essential to design features that are effective for signature classification. Experts classify signatures with predefined if-then rules. An if-then rule returns a label of low, medium, high, or unknown importance based on keyword matching of the elements in the signature. Therefore, we first design two types of features, symbolic features (SFs) and keyword features (KFs), which are used in keyword matching for the if-then rules. Next, we design web information and message features (WMFs) to capture the properties of signatures that do not match the if-then rules. The WMFs are extracted as term frequency-inverse document frequency (TF-IDF) features of the message text in the signatures. The features are obtained by web scraping from the referenced external attack identification systems described in the signature. Because failure needs to be minimized in the classification of IDPS signatures, as in the medical field, we consider introducing a RO in our proposed model. The effectiveness of the proposed classification model is evaluated in experiments with two real datasets composed of signatures labeled by experts: a dataset that can be classified with if-then rules and a dataset with elements that do not match an if-then rule. In the experiment, the proposed model is evaluated. In both cases, the combined SFs and WMFs performed better than the combined SFs and KFs. In addition, we also performed feature analysis.
Convolutional neural networks (CNN) are the dominant deep neural network (DNN) architecture for computer vision. Recently, Transformer and multi-layer perceptron (MLP)-based models, such as Vision Transformer and MLP-Mixer, started to lead new trends as they showed promising results in the ImageNet classification task. In this paper, we conduct empirical studies on these DNN structures and try to understand their respective pros and cons. To ensure a fair comparison, we first develop a unified framework called SPACH which adopts separate modules for spatial and channel processing. Our experiments under the SPACH framework reveal that all structures can achieve competitive performance at a moderate scale. However, they demonstrate distinctive behaviors when the network size scales up. Based on our findings, we propose two hybrid models using convolution and Transformer modules. The resulting Hybrid-MS-S+ model achieves 83.9% top-1 accuracy with 63M parameters and 12.3G FLOPS. It is already on par with the SOTA models with sophisticated designs. The code and models will be made publicly available.
Humans perceive the world by concurrently processing and fusing high-dimensional inputs from multiple modalities such as vision and audio. Machine perception models, in stark contrast, are typically modality-specific and optimised for unimodal benchmarks, and hence late-stage fusion of final representations or predictions from each modality (`late-fusion') is still a dominant paradigm for multimodal video classification. Instead, we introduce a novel transformer based architecture that uses `fusion bottlenecks' for modality fusion at multiple layers. Compared to traditional pairwise self-attention, our model forces information between different modalities to pass through a small number of bottleneck latents, requiring the model to collate and condense the most relevant information in each modality and only share what is necessary. We find that such a strategy improves fusion performance, at the same time reducing computational cost. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple audio-visual classification benchmarks including Audioset, Epic-Kitchens and VGGSound. All code and models will be released.
This paper proposes a generic method to learn interpretable convolutional filters in a deep convolutional neural network (CNN) for object classification, where each interpretable filter encodes features of a specific object part. Our method does not require additional annotations of object parts or textures for supervision. Instead, we use the same training data as traditional CNNs. Our method automatically assigns each interpretable filter in a high conv-layer with an object part of a certain category during the learning process. Such explicit knowledge representations in conv-layers of CNN help people clarify the logic encoded in the CNN, i.e., answering what patterns the CNN extracts from an input image and uses for prediction. We have tested our method using different benchmark CNNs with various structures to demonstrate the broad applicability of our method. Experiments have shown that our interpretable filters are much more semantically meaningful than traditional filters.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191