We propose Tailored ViT Slimming (TVS), a budget-aware multi-dimensional pruning framework for Vision Transformers. TVS injects learnable masks into MHSA and MLP modules and applies adaptive non-convex sparsity regularization to achieve maximal utilization of parameters under strict module-wise budgets. In addition, by retaining scaled masks after pruning, TVS avoids abrupt accuracy drops and provides stable initialization for fine-tuning. On ImageNet-1k with DeiT-S and DeiT-B, TVS consistently outperforms prior ViT compression methods. This result empirically shows that the non-convex sparsity regularizer is effective not only in CNNs but also in ViTs.

Abstract Several pruning methods prune a neural network at initialization. These methods carefully determine the importance of each weight, retaining only the important ones and pruning the others. However, subsequent studies have shown that random pruning can perform similarly, provided that the number of remaining weights within each layer is the same as the number of important weights within the corresponding layer. How can this random pruning, which disregards weight importance, especially within layers, still achieve comparable performance? In this work, we shed light on this question. Specifically, we demonstrate that by simply setting the number of weights to retain within layers in this manner, a large number of important weights tend to remain after pruning, and the importance of the remaining weights tends to be high. These statistical benefits are shown by comparing this random pruning method, which retains weights in a layer equal to the number of important weights in that layer, to another random pruning method that uses a uniform pruning ratio across all layers. With randomness applied, where important weights cannot be selectively distinguished from unimportant ones, the superiority of the former over the latter should be clarified. Theoretical proofs are provided, as well as empirical results from various architectures and datasets.

Rotation-equivariance is an essential yet challenging property in oriented object detection. While general object detectors naturally leverage robustness to spatial shifts due to the translation-equivariance of the conventional CNNs, achieving rotation-equivariance remains an elusive goal. Current detectors deploy various alignment techniques to derive rotation-invariant features, but still rely on high capacity models and heavy data augmentation with all possible rotations. In this paper, we introduce a Fully Rotation-Equivariant Oriented Object Detector (FRED), whose entire process from the image to the bounding box prediction is strictly equivariant. Specifically, we decouple the invariant task (object classification) and the equivariant task (object localization) to achieve end-to-end equivariance. We represent the bounding box as a set of rotation-equivariant vectors to implement rotation-equivariant localization. Moreover, we utilized these rotation-equivariant vectors as offsets in the deformable convolution, thereby enhancing the existing advantages of spatial adaptation. Leveraging full rotation-equivariance, our FRED demonstrates higher robustness to image-level rotation compared to existing methods. Furthermore, we show that FRED is one step closer to non-axis aligned learning through our experiments. Compared to state-of-the-art methods, our proposed method delivers comparable performance on DOTA-v1.0 and outperforms by 1.5 mAP on DOTA-v1.5, all while significantly reducing the model parameters to 16%.

Masked generative models (MGMs) have shown impressive generative ability while providing an order of magnitude efficient sampling steps compared to continuous diffusion models. However, MGMs still underperform in image synthesis compared to recent well-developed continuous diffusion models with similar size in terms of quality and diversity of generated samples. A key factor in the performance of continuous diffusion models stems from the guidance methods, which enhance the sample quality at the expense of diversity. In this paper, we extend these guidance methods to generalized guidance formulation for MGMs and propose a self-guidance sampling method, which leads to better generation quality. The proposed approach leverages an auxiliary task for semantic smoothing in vector-quantized token space, analogous to the Gaussian blur in continuous pixel space. Equipped with the parameter-efficient fine-tuning method and high-temperature sampling, MGMs with the proposed self-guidance achieve a superior quality-diversity trade-off, outperforming existing sampling methods in MGMs with more efficient training and sampling costs. Extensive experiments with the various sampling hyperparameters confirm the effectiveness of the proposed self-guidance.

When a person solves the multi-choice problem, she considers not only what is the answer but also what is not the answer. Knowing what choice is not the answer and utilizing the relationships between choices, she can improve the prediction accuracy. Inspired by this human reasoning process, we propose a new training strategy to fully utilize inter-class relationships, namely LogitMix. Our strategy is combined with recent data augmentation techniques, e.g., Mixup, Manifold Mixup, CutMix, and PuzzleMix. Then, we suggest using a mixed logit, i.e., a mixture of two logits, as an auxiliary training objective. Since the logit can preserve both positive and negative inter-class relationships, it can impose a network to learn the probability of wrong answers correctly. Our extensive experimental results on the image- and language-based tasks demonstrate that LogitMix achieves state-of-the-art performance among recent data augmentation techniques regarding calibration error and prediction accuracy.

Convolutional networks have achieved great success in various vision tasks. This is mainly due to a considerable amount of research on network structure. In this study, instead of focusing on architectures, we focused on the convolution unit itself. The existing convolution unit has a fixed shape and is limited to observing restricted receptive fields. In earlier work, we proposed the active convolution unit (ACU), which can freely define its shape and learn by itself. In this paper, we provide a detailed analysis of the previously proposed unit and show that it is an efficient representation of a sparse weight convolution. Furthermore, we extend an ACU to a grouped ACU, which can observe multiple receptive fields in one layer. We found that the performance of a naive grouped convolution is degraded by increasing the number of groups; however, the proposed unit retains the accuracy even though the number of parameters decreases. Based on this result, we suggest a depthwise ACU (DACU), and various experiments have shown that our unit is efficient and can replace the existing convolutions.

Convolutional neural networks have achieved great success in various vision tasks; however, they incur heavy resource costs. By using deeper and wider networks, network accuracy can be improved rapidly. However, in an environment with limited resources (e.g., mobile applications), heavy networks may not be usable. This study shows that naive convolution can be deconstructed into a shift operation and pointwise convolution. To cope with various convolutions, we propose a new shift operation called active shift layer (ASL) that formulates the amount of shift as a learnable function with shift parameters. This new layer can be optimized end-to-end through backpropagation and it can provide optimal shift values. Finally, we apply this layer to a light and fast network that surpasses existing state-of-the-art networks.

In recent years, deep learning has achieved great success in many computer vision applications. Convolutional neural networks (CNNs) have lately emerged as a major approach to image classification. Most research on CNNs thus far has focused on developing architectures such as the Inception and residual networks. The convolution layer is the core of the CNN, but few studies have addressed the convolution unit itself. In this paper, we introduce a convolution unit called the active convolution unit (ACU). A new convolution has no fixed shape, because of which we can define any form of convolution. Its shape can be learned through backpropagation during training. Our proposed unit has a few advantages. First, the ACU is a generalization of convolution, it can define not only all conventional convolutions, but also convolutions with fractional pixel coordinates. We can freely change the shape of the convolution, which provides greater freedom to form CNN structures. Second, the shape of the convolution is learned while training and there is no need to tune it by hand. Third, the ACU can learn better than a conventional unit, where we obtained the improvement simply by changing the conventional convolution to an ACU. We tested our proposed method on plain and residual networks, and the results showed significant improvement using our method on various datasets and architectures in comparison with the baseline. Code is available at https://github.com/jyh2986/Active-Convolution.

It is shown that a multilayer perceptron can learn highly nonlinear continuous mappings more easily if the target values are thermometer-coded. Because of the similarity between sigmoidal functions and thermometer coding, a network does not need more hidden nodes to produce thermometer-coded target values when more output nodes are added to the original structure. Furthermore, the weights for the hidden layers of a network, which are trained by thermometer-coded target values, can be used in the initialization of a network which is then trained by the original target values. The reason why such a two-staged learning is possible is discussed. Experiments on synthetic data sets show that using thermometer-coded target values improves the learning performance of a network and that conversion to a single output network is more efficient and gives better results.

An object-oriented approach to behavior-based control architecture for autonomous mobile robots as presented. Instead of traditional implementations of the architecture with proprietary programming languages which resulted in slow execution speed and poor hardware compatibility, the C++ language is used and an optimal scheduling algorithm is proposed for fast and guaranteed response time. Each behavior module is compiled separately and loaded at run-time using dynamic binding to improve modularity and expandability. A vector valued representation of output signals is also proposed to arbitrate conflicting signals properly without losing any information. The proposed method is implemented and applied to control a laboratory developed autonomous mobile robot and tested in real-time.