With the expanding role of neural networks, the need for complete and sound verification of their property has become critical. In the recent years, it was established that Binary Neural Networks (BNNs) have an equivalent representation in Boolean logic and can be formally analyzed using logical reasoning tools such as SAT solvers. However, to date, only BNNs can be transformed into a SAT formula. In this work, we introduce Truth Table Deep Convolutional Neural Networks (TTnets), a new family of SAT-encodable models featuring for the first time real-valued weights. Furthermore, it admits, by construction, some valuable conversion features including post-tuning and tractability in the robustness verification setting. The latter property leads to a more compact SAT symbolic encoding than BNNs. This enables the use of general SAT solvers, making property verification easier. We demonstrate the value of TTnets regarding the formal robustness property: TTnets outperform the verified accuracy of all BNNs with a comparable computation time. More generally, they represent a relevant trade-off between all known complete verification methods: TTnets achieve high verified accuracy with fast verification time, being complete with no timeouts. We are exploring here a proof of concept of TTnets for a very important application (complete verification of robustness) and we believe this novel real-valued network constitutes a practical response to the rising need for functional formal verification. We postulate that TTnets can apply to various CNN-based architectures and be extended to other properties such as fairness, fault attack and exact rule extraction.
Cold temperatures during fall and spring have the potential to cause frost damage to grapevines and other fruit plants, which can significantly decrease harvest yields. To help prevent these losses, farmers deploy expensive frost mitigation measures, such as, sprinklers, heaters, and wind machines, when they judge that damage may occur. This judgment, however, is challenging because the cold hardiness of plants changes throughout the dormancy period and it is difficult to directly measure. This has led scientists to develop cold hardiness prediction models that can be tuned to different grape cultivars based on laborious field measurement data. In this paper, we study whether deep-learning models can improve cold hardiness prediction for grapes based on data that has been collected over a 30-year time period. A key challenge is that the amount of data per cultivar is highly variable, with some cultivars having only a small amount. For this purpose, we investigate the use of multi-task learning to leverage data across cultivars in order to improve prediction performance for individual cultivars. We evaluate a number of multi-task learning approaches and show that the highest performing approach is able to significantly improve over learning for single cultivars and outperforms the current state-of-the-art scientific model for most cultivars.
Most state-of-the-art localization algorithms rely on robust relative pose estimation and geometry verification to obtain moving object agnostic camera poses in complex indoor environments. However, this approach is prone to mistakes if a scene contains repetitive structures, e.g., desks, tables, boxes, or moving people. We show that the movable objects incorporate non-negligible localization error and present a new straightforward method to predict the six-degree-of-freedom (6DoF) pose more robustly. We equipped the localization pipeline InLoc with real-time instance segmentation network YOLACT++. The masks of dynamic objects are employed in the relative pose estimation step and in the final sorting of camera pose proposal. At first, we filter out the matches laying on masks of the dynamic objects. Second, we skip the comparison of query and synthetic images on the area related to the moving object. This procedure leads to a more robust localization. Lastly, we describe and improve the mistakes caused by gradient-based comparison between synthetic and query images and publish a new pipeline for simulation of environments with movable objects from the Matterport scans. All the codes are available on github.com/dubenma/D-InLocpp .
Automatic tomato disease recognition from leaf images is vital to avoid crop losses by applying control measures on time. Even though recent deep learning-based tomato disease recognition methods with classical training procedures showed promising recognition results, they demand large labelled data and involve expensive training. The traditional deep learning models proposed for tomato disease recognition also consume high memory and storage because of a high number of parameters. While lightweight networks overcome some of these issues to a certain extent, they continue to show low performance and struggle to handle imbalanced data. In this paper, a novel Siamese network-based lightweight framework is proposed for automatic tomato leaf disease recognition. This framework achieves the highest accuracy of 96.97% on the tomato subset obtained from the PlantVillage dataset and 95.48% on the Taiwan tomato leaf disease dataset. Experimental results further confirm that the proposed framework is effective with imbalanced and small data. The backbone deep network integrated with this framework is lightweight with approximately 2.9629 million trainable parameters, which is way lower than existing lightweight deep networks.
The graph structure of road networks is critical for downstream tasks of autonomous driving systems, such as global planning, motion prediction and control. In the past, the road network graph is usually manually annotated by human experts, which is time-consuming and labor-intensive. To obtain the road network graph with better effectiveness and efficiency, automatic approaches for road network graph detection are required. Previous works either post-process semantic segmentation maps or propose graph-based algorithms to directly predict the road network graph. However, previous works suffer from hard-coded heuristic processing algorithms and inferior final performance. To enhance the previous SOTA (State-of-the-Art) approach RNGDet, we add an instance segmentation head to better supervise the model training, and enable the model to leverage multi-scale features of the backbone network. Since the new proposed approach is improved from RNGDet, it is named RNGDet++. All approaches are evaluated on a large publicly available dataset. RNGDet++ outperforms baseline models on almost all metrics scores. It improves the topology correctness APLS (Average Path Length Similarity) by around 3\%. The demo video and supplementary materials are available on our project page \url{https://tonyxuqaq.github.io/projects/RNGDetPlusPlus/}.
We introduce a mixture of heterogeneous experts framework called \texttt{MECATS}, which simultaneously forecasts the values of a set of time series that are related through an aggregation hierarchy. Different types of forecasting models can be employed as individual experts so that the form of each model can be tailored to the nature of the corresponding time series. \texttt{MECATS} learns hierarchical relationships during the training stage to help generalize better across all the time series being modeled and also mitigates coherency issues that arise due to constraints imposed by the hierarchy. We further build multiple quantile estimators on top of the point forecasts. The resulting probabilistic forecasts are nearly coherent, distribution-free, and independent of the choice of forecasting models. We conduct a comprehensive evaluation on both point and probabilistic forecasts and also formulate an extension for situations where change points exist in sequential data. In general, our method is robust, adaptive to datasets with different properties, and highly configurable and efficient for large-scale forecasting pipelines.
As an emerging secure learning paradigm in leveraging cross-silo private data, vertical federated learning (VFL) is expected to improve advertising models by enabling the joint learning of complementary user attributes privately owned by the advertiser and the publisher. However, the 1) restricted applicable scope to overlapped samples and 2) high system challenge of real-time federated serving have limited its application to advertising systems. In this paper, we advocate new learning setting Semi-VFL (Vertical Semi-Federated Learning) as a lightweight solution to utilize all available data (both the overlapped and non-overlapped data) that is free from federated serving. Semi-VFL is expected to perform better than single-party models and maintain a low inference cost. It's notably important to i) alleviate the absence of the passive party's feature and ii) adapt to the whole sample space to implement a good solution for Semi-VFL. Thus, we propose a carefully designed joint privileged learning framework (JPL) as an efficient implementation of Semi-VFL. Specifically, we build an inference-efficient single-party student model applicable to the whole sample space and meanwhile maintain the advantage of the federated feature extension. Novel feature imitation and ranking consistency restriction methods are proposed to extract cross-party feature correlations and maintain cross-sample-space consistency for both the overlapped and non-overlapped data. We conducted extensive experiments on real-world advertising datasets. The results show that our method achieves the best performance over baseline methods and validate its effectiveness in maintaining cross-view feature correlation.
Self-supervised learning makes great progress in large model pre-training but suffers in training small models. Previous solutions to this problem mainly rely on knowledge distillation and indeed have a two-stage learning procedure: first train a large teacher model, then distill it to improve the generalization ability of small ones. In this work, we present a new one-stage solution to obtain pre-trained small models without extra teachers: slimmable networks for contrastive self-supervised learning (\emph{SlimCLR}). A slimmable network contains a full network and several weight-sharing sub-networks. We can pre-train for only one time and obtain various networks including small ones with low computation costs. However, in self-supervised cases, the interference between weight-sharing networks leads to severe performance degradation. One evidence of the interference is \emph{gradient imbalance}: a small proportion of parameters produces dominant gradients during backpropagation, and the main parameters may not be fully optimized. The divergence in gradient directions of various networks may also cause interference between networks. To overcome these problems, we make the main parameters produce dominant gradients and provide consistent guidance for sub-networks via three techniques: slow start training of sub-networks, online distillation, and loss re-weighting according to model sizes. Besides, a switchable linear probe layer is applied during linear evaluation to avoid the interference of weight-sharing linear layers. We instantiate SlimCLR with typical contrastive learning frameworks and achieve better performance than previous arts with fewer parameters and FLOPs.
In this letter, we present a robust, real-time, inertial navigation system (INS)-Centric GNSS-Visual-Inertial navigation system (IC-GVINS) for wheeled robot, in which the precise INS is fully utilized in both the state estimation and visual process. To improve the system robustness, the INS information is employed during the whole keyframe-based visual process, with strict outlier-culling strategy. GNSS is adopted to perform an accurate and convenient initialization of the IC-GVINS, and is further employed to achieve absolute positioning in large-scale environments. The IMU, visual, and GNSS measurements are tightly fused within the framework of factor graph optimization. Dedicated experiments were conducted to evaluate the robustness and accuracy of the IC-GVINS on a wheeled robot. The IC-GVINS demonstrates superior robustness in various visual-degenerated scenes with moving objects. Compared to the state-of-the-art visual-inertial navigation systems, the proposed method yields improved robustness and accuracy in various environments. We open source our codes combined with the dataset on GitHub
State estimation is an essential part of autonomous systems. Integrating the Ultra-Wideband(UWB) technique has been shown to correct the long-term estimation drift and bypass the complexity of loop closure detection. However, few works on robotics adopt UWB as a stand-alone state estimation solution. The primary purpose of this work is to investigate planar pose estimation using only UWB range measurements and study the estimator's statistical efficiency. We prove the excellent property of a two-step scheme, which says that we can refine a consistent estimator to be asymptotically efficient by one step of Gauss-Newton iteration. Grounded on this result, we design the GN-ULS estimator and evaluate it through simulations and collected datasets. GN-ULS attains millimeter and sub-degree level accuracy on our static datasets and attains centimeter and degree level accuracy on our dynamic datasets, presenting the possibility of using only UWB for real-time state estimation.