Data augmentation (DA) is a crucial technique for enhancing the sample efficiency of visual reinforcement learning (RL) algorithms. Notably, employing simple observation transformations alone can yield outstanding performance without extra auxiliary representation tasks or pre-trained encoders. However, it remains unclear which attributes of DA account for its effectiveness in achieving sample-efficient visual RL. To investigate this issue and further explore the potential of DA, this work conducts comprehensive experiments to assess the impact of DA's attributes on its efficacy and provides the following insights and improvements: (1) For individual DA operations, we reveal that both ample spatial diversity and slight hardness are indispensable. Building on this finding, we introduce Random PadResize (Rand PR), a new DA operation that offers abundant spatial diversity with minimal hardness. (2) For multi-type DA fusion schemes, the increased DA hardness and unstable data distribution result in the current fusion schemes being unable to achieve higher sample efficiency than their corresponding individual operations. Taking the non-stationary nature of RL into account, we propose a RL-tailored multi-type DA fusion scheme called Cycling Augmentation (CycAug), which performs periodic cycles of different DA operations to increase type diversity while maintaining data distribution consistency. Extensive evaluations on the DeepMind Control suite and CARLA driving simulator demonstrate that our methods achieve superior sample efficiency compared with the prior state-of-the-art methods.
Living systems can use a single periphery to perform a variety of tasks and adapt to a dynamic environment. This multifunctionality is achieved through the use of neural circuitry that adaptively controls the reconfigurable musculature. Current robotic systems struggle to flexibly adapt to unstructured environments. Through mimicry of the neuromechanical coupling seen in living organisms, robotic systems could potentially achieve greater autonomy. The tractable neuromechanics of the sea slug $\textit{Aplysia californica's}$ feeding apparatus, or buccal mass, make it an ideal candidate for applying neuromechanical principles to the control of a soft robot. In this work, a robotic grasper was designed to mimic specific morphology of the $\textit{Aplysia}$ feeding apparatus. These include the use of soft actuators akin to biological muscle, a deformable grasping surface, and a similar muscular architecture. A previously developed Boolean neural controller was then adapted for the control of this soft robotic system. The robot was capable of qualitatively replicating swallowing behavior by cyclically ingesting a plastic tube. The robot's normalized translational and rotational kinematics of the odontophore followed profiles observed $\textit{in vivo}$ despite morphological differences. This brings $\textit{Aplysia}$-inspired control $\textit{in roboto}$ one step closer to multifunctional neural control schema $\textit{in vivo}$ and $\textit{in silico}$. Future additions may improve SLUGBOT's viability as a neuromechanical research platform.
In deep metric learning, the Triplet Loss has emerged as a popular method to learn many computer vision and natural language processing tasks such as facial recognition, object detection, and visual-semantic embeddings. One issue that plagues the Triplet Loss is network collapse, an undesirable phenomenon where the network projects the embeddings of all data onto a single point. Researchers predominately solve this problem by using triplet mining strategies. While hard negative mining is the most effective of these strategies, existing formulations lack strong theoretical justification for their empirical success. In this paper, we utilize the mathematical theory of isometric approximation to show an equivalence between the Triplet Loss sampled by hard negative mining and an optimization problem that minimizes a Hausdorff-like distance between the neural network and its ideal counterpart function. This provides the theoretical justifications for hard negative mining's empirical efficacy. In addition, our novel application of the isometric approximation theorem provides the groundwork for future forms of hard negative mining that avoid network collapse. Our theory can also be extended to analyze other Euclidean space-based metric learning methods like Ladder Loss or Contrastive Learning.
Synthesizing free-view photo-realistic images is an important task in multimedia. With the development of advanced driver assistance systems~(ADAS) and their applications in autonomous vehicles, experimenting with different scenarios becomes a challenge. Although the photo-realistic street scenes can be synthesized by image-to-image translation methods, which cannot produce coherent scenes due to the lack of 3D information. In this paper, a large-scale neural rendering method is proposed to synthesize the autonomous driving scene~(READ), which makes it possible to synthesize large-scale driving scenarios on a PC through a variety of sampling schemes. In order to represent driving scenarios, we propose an {\omega} rendering network to learn neural descriptors from sparse point clouds. Our model can not only synthesize realistic driving scenes but also stitch and edit driving scenes. Experiments show that our model performs well in large-scale driving scenarios.
Tactile sensing typically involves active exploration of unknown surfaces and objects, making it especially effective at processing the characteristics of materials and textures. A key property extracted by human tactile perception is surface roughness, which relies on measuring vibratory signals using the multi-layered fingertip structure. Existing robotic systems lack tactile sensors that are able to provide high dynamic sensing ranges, perceive material properties, and maintain a low hardware cost. In this work, we introduce the reference design and fabrication procedure of a miniature and low-cost tactile sensor consisting of a biomimetic cutaneous structure, including the artificial fingerprint, dermis, epidermis, and an embedded magnet-sensor structure which serves as a mechanoreceptor for converting mechanical information to digital signals. The presented sensor is capable of detecting high-resolution magnetic field data through the Hall effect and creating high-dimensional time-frequency domain features for material texture classification. Additionally, we investigate the effects of different superficial sensor fingerprint patterns for classifying materials through both simulation and physical experimentation. After extracting time series and frequency domain features, we assess a k-nearest neighbors classifier for distinguishing between different materials. The results from our experiments show that our biomimetic tactile sensors with fingerprint ridges can classify materials with more than 8% higher accuracy and lower variability than ridge-less sensors. These results, along with the low cost and customizability of our sensor, demonstrate high potential for lowering the barrier to entry for a wide array of robotic applications, including model-less tactile sensing for texture classification, material inspection, and object recognition.
In this paper, we present the development of a sensing system with the capability to compute multispectral point clouds in real-time. The proposed multi-eye sensor system effectively registers information from the visible, (long-wave) infrared, and ultraviolet spectrum to its depth sensing frame, thus enabling to measure a wider range of surface features that are otherwise hidden to the naked eye. For that, we designed a new cross-calibration apparatus that produces consistent features which can be sensed by each of the cameras, therefore, acting as a multispectral "chessboard". The performance of the sensor is evaluated with two different cases of studies, where we show that the proposed system can detect "hidden" features of a 3D environment.
The task of video-based commonsense captioning aims to generate event-wise captions and meanwhile provide multiple commonsense descriptions (e.g., attribute, effect and intention) about the underlying event in the video. Prior works explore the commonsense captions by using separate networks for different commonsense types, which is time-consuming and lacks mining the interaction of different commonsense. In this paper, we propose a Hybrid Reasoning Network (HybridNet) to endow the neural networks with the capability of semantic-level reasoning and word-level reasoning. Firstly, we develop multi-commonsense learning for semantic-level reasoning by jointly training different commonsense types in a unified network, which encourages the interaction between the clues of multiple commonsense descriptions, event-wise captions and videos. Then, there are two steps to achieve the word-level reasoning: (1) a memory module records the history predicted sequence from the previous generation processes; (2) a memory-routed multi-head attention (MMHA) module updates the word-level attention maps by incorporating the history information from the memory module into the transformer decoder for word-level reasoning. Moreover, the multimodal features are used to make full use of diverse knowledge for commonsense reasoning. Experiments and abundant analysis on the large-scale Video-to-Commonsense benchmark show that our HybridNet achieves state-of-the-art performance compared with other methods.
Recently, plenty of work has tried to introduce transformers into computer vision tasks, with good results. Unlike classic convolution networks, which extract features within a local receptive field, transformers can adaptively aggregate similar features from a global view using self-attention mechanism. For object detection, Feature Pyramid Network (FPN) proposes feature interaction across layers and proves its extremely importance. However, its interaction is still in a local manner, which leaves a lot of room for improvement. Since transformer was originally designed for NLP tasks, adapting processing subject directly from text to image will cause unaffordable computation and space overhead. In this paper, we utilize a linearized attention function to overcome above problems and build a novel architecture, named Content-Augmented Feature Pyramid Network (CA-FPN), which proposes a global content extraction module and deeply combines with FPN through light linear transformers. What's more, light transformers can further make the application of multi-head attention mechanism easier. Most importantly, our CA-FPN can be readily plugged into existing FPN-based models. Extensive experiments on the challenging COCO object detection dataset demonstrated that our CA-FPN significantly outperforms competitive baselines without bells and whistles. Code will be made publicly available.
Cycle representatives of persistent homology classes can be used to provide descriptions of topological features in data. However, the non-uniqueness of these representatives creates ambiguity and can lead to many different interpretations of the same set of classes. One approach to solving this problem is to optimize the choice of representative against some measure that is meaningful in the context of the data. In this work, we provide a study of the effectiveness and computational cost of several $\ell_1$-minimization optimization procedures for constructing homological cycle bases for persistent homology with rational coefficients in dimension one, including uniform-weighted and length-weighted edge-loss algorithms as well as uniform-weighted and area-weighted triangle-loss algorithms. We conduct these optimizations via standard linear programming methods, applying general-purpose solvers to optimize over column bases of simplicial boundary matrices. Our key findings are: (i) optimization is effective in reducing the size of cycle representatives, (ii) the computational cost of optimizing a basis of cycle representatives exceeds the cost of computing such a basis in most data sets we consider, (iii) the choice of linear solvers matters a lot to the computation time of optimizing cycles, (iv) the computation time of solving an integer program is not significantly longer than the computation time of solving a linear program for most of the cycle representatives, using the Gurobi linear solver, (v) strikingly, whether requiring integer solutions or not, we almost always obtain a solution with the same cost and almost all solutions found have entries in {-1, 0, 1} and therefore, are also solutions to a restricted $\ell_0$ optimization problem, and (vi) we obtain qualitatively different results for generators in Erd\H{o}s-R\'enyi random clique complexes.
We investigate how the addition of quantum resources changes the statistical complexity of quantum circuits by utilizing the framework of quantum resource theories. Measures of statistical complexity that we consider include the Rademacher complexity and the Gaussian complexity, which are well-known measures in computational learning theory that quantify the richness of classes of real-valued functions. We derive bounds for the statistical complexities of quantum circuits that have limited access to certain resources and apply our results to two special cases: (1) stabilizer circuits that are supplemented with a limited number of T gates and (2) instantaneous quantum polynomial-time Clifford circuits that are supplemented with a limited number of CCZ gates. We show that the increase in the statistical complexity of a quantum circuit when an additional quantum channel is added to it is upper bounded by the free robustness of the added channel. Finally, we derive bounds for the generalization error associated with learning from training data arising from quantum circuits.