Early language development shapes children’s later literacy and learning, yet many families have limited access to scalable, high-quality support at home. Recent advances in generative AI make it possible for social robots to move beyond scripted interactions and engage children in adaptive, conversational activities, but it remains unclear how to design such systems for pre-schoolers and how children engage with them over time in the home. We present ELLA (Early Language Learning Agent), an autonomous, LLM-powered social robot that supports early language development through interactive storytelling, parent-selected language targets, and scaffolded dialogue. Using a multi-phased, human-centered process, we interviewed parents (n=7) and educators (n=5) and iteratively refined ELLA through twelve in-home design workshops. We then deployed ELLA with ten children for eight days. We report design insights from in-home workshops, characterize children’s engagement and behaviors during deployment, and distill design implications for generative AI–powered social robots supporting early language learning at home.

Out-of-distribution (OOD) generalization in Graph Neural Networks (GNNs) has gained significant attention due to its critical importance in graph-based predictions in real-world scenarios. Existing methods primarily focus on extracting a single causal subgraph from the input graph to achieve generalizable predictions. However, relying on a single subgraph can lead to susceptibility to spurious correlations and is insufficient for learning invariant patterns behind graph data. Moreover, in many real-world applications, such as molecular property prediction, multiple critical subgraphs may influence the target label property. To address these challenges, we propose a novel framework, SubGraph Aggregation(SuGAr), designed to learn a diverse set of subgraphs that are crucial for OOD generalization on graphs. Specifically, SuGAr employs a tailored subgraph sampler and diversity regularizer to extract a diverse set of invariant subgraphs. These invariant subgraphs are then aggregated by averaging their representations, which enriches the subgraph signals and enhances coverage of the underlying causal structures, thereby improving OOD generalization. Extensive experiments on both synthetic and real-world datasets demonstrate that SuGAr outperforms state-of-the-art methods, achieving up to a 24% improvement in OOD generalization on graphs. To the best of our knowledge, this is the first work to study graph OOD generalization by learning multiple invariant subgraphs.

Wireless charging holds significant promise for IoT devices and transportation networks by facilitating convenient and autonomous power supply. Traditional wireless charging technologies have typically adhered to a singular approach, choosing between near-field coupling or far-field radiation. However, our investigations uncover that each method outperforms the other at specific distances. This insight leads us to integrating the advantages of both to enable rapid wireless charging across any distance within the charging range. For this vision, we poses an intriguing question: "Can we develop a system that supports both near-field and far-field charging simultaneously?"

Figure 1: UltraPoser enables ubiquitous full-body pose estimation by integrating ultrasound sensing and IMU using commodity wearable devices.In addition to measuring IMU data, a smartphone and smartwatch are used to transmit and receive ultrasound signals.The extracted ultrasound features capture motions from joints without any attached devices and offer drift-free range measurements to complement IMU data for more accurate pose estimation.

Silent Speech Interfaces (SSI) have been developed to convert silent articulatory gestures into speech, aiding communication in public spaces and assisting individuals with aphasia. Previous SSIs, which rely on wearable devices or cameras, often pose issues like prolonged contact or privacy risks. Recent advancements in acoustic sensing present new opportunities for gesture sensing, but they typically focus on content classification rather than reconstructing audible speech. This results in the loss of crucial speech characteristics such as rate, intonation, and emotion.In this paper, we propose UltraSR, a novel sensing system designed for accurate audible speech reconstruction by analyzing the disturbance of tiny articulatory gestures on reflected ultrasound signals. UltraSR employs a multi-scale feature extraction scheme to aggregate information from multiple views and introduces a new model that maps ultrasound to speech signals, enabling the reconstruction of audible speech from silent gestures.Instead of the laborious collection of massive training data, UltraSR constructs an inverse task to generate virtual gestures from widely available audio (e.g., phone calls) for efficient model training. Additionally, it incorporates a finetuning mechanism using unlabeled data for user adaptation.We implemented UltraSR on a portable smartphone and evaluated it in various environments. Results show that UltraSR can achieve a Character Error Rate (CER) as low as 5.22% and reduce the CER from 80.13% to 6.31% for new users with only 1 hour of ultrasound data, outperforming state-of-the-art acoustic-based approaches while preserving rich speech information.

Silent Speech Interface (SSI) has been proposed as a means of reconstructing audible speech from silent articulatory gestures for covert voice communication in public and voice assistance for the aphasic. Prior arts of SSI, either relying on wearable devices or cameras, may lead to extended contact requirements or privacy leakage risks. The recent advances in acoustic sensing have brought new opportunities for sensing gestures, but their original intention is to infer speech content for classification instead of audible speech reconstruction, resulting in the loss of some important speech information (e.g., speech rate, intonation, and emotion). In this paper, we propose, the first system that supports accurate audible speech reconstruction by analyzing the disturbance of tiny articulatory gestures on the reflected ultrasound signal. The design of introduces a new model that provides the unique mapping relationship between ultrasound and speech signals, so that the audible speech can be successfully reconstructed from the silent speech. However, establishing the mapping relationship depends on plenty of training data. Instead of the time-consuming collection of massive amounts of data for training, we construct an inverse task that constitutes a dual form with the original task to generate virtual gestures from widely available audio (e.g., phone calls) for facilitating model training. Furthermore, we introduce a fine-tuning mechanism using unlabeled data for user adaptation. We implement using a portable smartphone and evaluate it in various environments. The evaluation results show that can reconstruct speech with a (Character Error Rate) CER as low as 7.62%, and decrease the CER from 82.77% to 9.42% on new users with only 1 hour of ultrasound signals provided, which outperforms state-of-the-art acoustic-based approaches while preserving rich speech information.

The adaptive hinging hyperplane (AHH) model is a popular piecewise linear representation with a generalized tree structure and has been successfully applied in dynamic system identification. In this article, we aim to construct the deep AHH (DAHH) model to extend and generalize the networking of AHH model for high-dimensional problems. The network structure of DAHH is determined through a forward growth, in which the activity ratio is introduced to select effective neurons and no connecting weights are involved between the layers. Then, all neurons in the DAHH network can be flexibly connected to the output in a skip-layer format, and only the corresponding weights are the parameters to optimize. With such a network framework, the backpropagation algorithm can be implemented in DAHH to efficiently tackle large-scale problems and the gradient vanishing problem is not encountered in the training of DAHH. In fact, the optimization problem of DAHH can maintain convexity with convex loss in the output layer, which brings natural advantages in optimization. Different from the existing neural networks, DAHH is easier to interpret, where neurons are connected sparsely and analysis of variance (ANOVA) decomposition can be applied, facilitating to revealing the interactions between variables. A theoretical analysis toward universal approximation ability and explicit domain partitions are also derived. Numerical experiments verify the effectiveness of the proposed DAHH.

Deep Neural Networks (DNN) has gained great success in solving several challenging problems in recent years. It is well known that training a DNN model from scratch requires a lot of data and computational resources. However, using a pre-trained model directly or using it to initialize weights cost less time and often gets better results. Therefore, well pre-trained DNN models are valuable intellectual property that we should protect. In this work, we propose DeepTrace, a framework for model owners to secretly fingerprinting the target DNN model using a special trigger set and verifying from outputs. An embedded fingerprint can be extracted to uniquely identify the information of model owner and authorized users. Our framework benefits from both white-box and black-box verification, which makes it useful whether we know the model details or not. We evaluate the performance of DeepTrace on two different datasets, with different DNN architectures. Our experiment shows that, with the advantages of combining white-box and black-box verification, our framework has very little effect on model accuracy, and is robust against different model modifications. It also consumes very little computing resources when extracting fingerprint.

In this paper, we present an approach to nonlinear system approximation, called the lattice trajectory piecewise linear (LTPWL) model. The approach involves determining a lattice piecewise linear (PWL) approximation to the state trajectory of a nonlinear system. It has been shown in the literature that the lattice PWL expression can represent any PWL function in any dimension. After the LTPWL approximation has been obtained, the order of each model piece is reduced using a Krylov projection technique. Compared to existing trajectory piecewise linear (TPWL) models, which are quasi-PWL in the whole region, LTPWL models are virtually linear in each subregion. Besides, the single output LTPWL model can be seen as a special kind of TPWL model, in which only one weight is 1, and the other weights are 0. In general, for multiple output LTPWL model, the weights set to be 1 for each component are different, which makes the LTPWL model more flexible in approximation of nonlinear function. The proposed strategy is applied to simulate diode circuits, and the experimental results show that the performance of the LTPWL model is better than that of the traditional TPWL model in terms of approximation accuracy and generalization ability.

This brief proposes a truncated distance (TL1) kernel, which results in a classifier that is nonlinear in the global region but is linear in each subregion. With this kernel, the subregion structure can be trained using all the training data and local linear classifiers can be established simultaneously. The TL1 kernel has good adaptiveness to nonlinearity and is suitable for problems which require different nonlinearities in different areas. Though the TL1 kernel is not positive semidefinite, some classical kernel learning methods are still applicable which means that the TL1 kernel can be directly used in standard toolboxes by replacing the kernel evaluation. In numerical experiments, the TL1 kernel with a pregiven parameter achieves similar or better performance than the radial basis function kernel with the parameter tuned by cross validation, implying the TL1 kernel a promising nonlinear kernel for classification tasks.

In this paper, we propose a novel adaptive piecewise linear model for dynamic system identification. It has four unique features. First, the model designs a new kind of basis function for function approximation. It maintains the uniform shape for each basis function, so as to achieve a satisfactory tradeoff between generalization ability and model complexity. Second, the model takes the structure of basis functions as decision variables to optimize the formulated identification problems instead of taking expansion coefficients as decision variables as proposed by many existing approaches. Third, we establish an incremental design strategy to solve the system identification problems. In each step of the identification, the selection of optimal basis function is a Lipschitz continuous optimization problem that is likely to be easily handled with some mature toolboxes. This incremental design strategy greatly reduces the estimation cost. Fourth, we introduce a smoothing mechanism to avoid overfitting, when the output of dynamic systems is disturbed by noise. Tests on several benchmark dynamic systems demonstrate the potential of the proposed model.

This paper describes an optimization strategy based on the model of adaptive hinging hyperplanes (AHH) and genetic algorithm (GA). The sample points of physical model are approximated by the AHH model, and the resulting model is minimized using a modified GA. In the modified GA, each chromosome corresponds to a local optimum. A criterion based on γ-valid cut is used to judge whether the global optimum is reached. Simulation results show that if the parameters are carefully chosen, the global optimum of AHH minimization is close to the optimum of the original function.

Centrifugal chiller plants (CCP) are widely used in air conditioning systems, its operation optimization can save lots of energy and has great significance in environmental protection. The optimization is a large-scale nonlinear problem and there is no practical algorithm until now. This paper proposes a new method to do this operation optimization using continuous piecewise linear programming (CPWLP). The main idea is transforming the nonlinear problem into a series of linear programmings by approximating the original system using piecewise linear representation. For CPWLP, some properties are discussed and an algorithm is given. Using CPWLP, CCP system is optimized and its energy performance is improved significantly.

Smooth hinging hyperplane (SHH) has been proposed as an improvement over the well-known hinging hyperplane (HH) by the fact that it retains the useful features of HH while overcoming HH's drawback of nondifferentiability. This paper introduces a formal characterization of smooth hinge function (SHF), which can be used to generate SHH as a neural network. A method for the general construction of SHF is also given. Furthermore, the work proves that SHH is better than HH in functional approximation, i.e., the optimal error of SHH approximating a general function is always smaller or equal to that of HH. Particularly, in the case that the SHF is generated via the integration of a class of sigmoidal functions, it is further proven that the corresponding SHH of the 2m SHFs would outperform a neural network with m of the sigmoidal function from which the SHF is derived. Any upper bound established on the approximation error of a neural network of m sigmoidal activation functions can hence be translated to the SHH of m SHFs by replacing m with [m/2]. The work also includes an algorithm for the identification of SHH making use of its differentiability property. Simulation experiments are presented to validate the theoretical conclusions to possible extent.

The problem of constructing a general continuous piecewise-linear neural network is considered in this paper. It is shown that every projection domain of an arbitrary continuous piecewise-linear function can be partitioned into convex polyhedra by using difference functions of its local linear functions. Based on these convex polyhedra, a group of continuous piecewise-linear basis functions are formulated. It is proven that a linear combination of these basis functions plus a constant, which we call a standard continuous piecewise-linear neural network, can represent all continuous piecewise-linear functions. In addition, the proposed standard continuous piecewise-linear neural network is applied to solve some function approximation problems. A number of numerical experiments are presented to illustrate that the standard continuous piecewise-linear neural network can be a promising tool for function approximation.

The model of hinging hyperplanes (HH) can approximate a large class of nonlinear functions to arbitrary precision, but represent only a small part of continuous piecewise-linear (CPWL) functions in two or more dimensions. In this correspondence, the influence of this drawback for black-box modeling is first illustrated by a simple example. Then it is shown that the above shortcoming can be amended by adding a sufficient number of linear functions to current hinges. It is proven that any CPWL function of n variables can be represented by a sum of hinges containing at most n+1 linear functions. Hence the model of a sum of such expanded hinges is a general representation for all CPWL functions. The structure of the novel general representation is much simpler than the existing generalized canonical representation that consists of nested absolute-value functions. This characteristic is very useful for black-box modeling. Based on the new general representation, an upper bound on the number of nestings of nested absolute-value functions of a generalized canonical representation is established, which is much smaller than the known result.