大多数自治车辆都配备了LIDAR传感器和立体声相机。前者非常准确，但产生稀疏数据，而后者是密集的，具有丰富的纹理和颜色信息，但难以提取来自的强大的3D表示。在本文中，我们提出了一种新的数据融合算法，将准确的点云与致密的，但不太精确的点云组合在立体对。我们开发一个框架，将该算法集成到各种3D对象检测方法中。我们的框架从两个RGB图像中的2D检测开始，计算截肢和它们的交叉点，从立体声图像创建伪激光雷达数据，并填补了LIDAR数据缺少密集伪激光器的交叉区域的部分要点。我们训练多个3D对象检测方法，并表明我们的融合策略一致地提高了探测器的性能。

Wind power forecasting helps with the planning for the power systems by contributing to having a higher level of certainty in decision-making. Due to the randomness inherent to meteorological events (e.g., wind speeds), making highly accurate long-term predictions for wind power can be extremely difficult. One approach to remedy this challenge is to utilize weather information from multiple points across a geographical grid to obtain a holistic view of the wind patterns, along with temporal information from the previous power outputs of the wind farms. Our proposed CNN-RNN architecture combines convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to extract spatial and temporal information from multi-dimensional input data to make day-ahead predictions. In this regard, our method incorporates an ultra-wide learning view, combining data from multiple numerical weather prediction models, wind farms, and geographical locations. Additionally, we experiment with global forecasting approaches to understand the impact of training the same model over the datasets obtained from multiple different wind farms, and we employ a method where spatial information extracted from convolutional layers is passed to a tree ensemble (e.g., Light Gradient Boosting Machine (LGBM)) instead of fully connected layers. The results show that our proposed CNN-RNN architecture outperforms other models such as LGBM, Extra Tree regressor and linear regression when trained globally, but fails to replicate such performance when trained individually on each farm. We also observe that passing the spatial information from CNN to LGBM improves its performance, providing further evidence of CNN's spatial feature extraction capabilities.

Continuous long-term monitoring of motor health is crucial for the early detection of abnormalities such as bearing faults (up to 51% of motor failures are attributed to bearing faults). Despite numerous methodologies proposed for bearing fault detection, most of them require normal (healthy) and abnormal (faulty) data for training. Even with the recent deep learning (DL) methodologies trained on the labeled data from the same machine, the classification accuracy significantly deteriorates when one or few conditions are altered. Furthermore, their performance suffers significantly or may entirely fail when they are tested on another machine with entirely different healthy and faulty signal patterns. To address this need, in this pilot study, we propose a zero-shot bearing fault detection method that can detect any fault on a new (target) machine regardless of the working conditions, sensor parameters, or fault characteristics. To accomplish this objective, a 1D Operational Generative Adversarial Network (Op-GAN) first characterizes the transition between normal and fault vibration signals of (a) source machine(s) under various conditions, sensor parameters, and fault types. Then for a target machine, the potential faulty signals can be generated, and over its actual healthy and synthesized faulty signals, a compact, and lightweight 1D Self-ONN fault detector can then be trained to detect the real faulty condition in real time whenever it occurs. To validate the proposed approach, a new benchmark dataset is created using two different motors working under different conditions and sensor locations. Experimental results demonstrate that this novel approach can accurately detect any bearing fault achieving an average recall rate of around 89% and 95% on two target machines regardless of its type, severity, and location.

Basecalling is an essential step in nanopore sequencing analysis where the raw signals of nanopore sequencers are converted into nucleotide sequences, i.e., reads. State-of-the-art basecallers employ complex deep learning models to achieve high basecalling accuracy. This makes basecalling computationally-inefficient and memory-hungry; bottlenecking the entire genome analysis pipeline. However, for many applications, the majority of reads do no match the reference genome of interest (i.e., target reference) and thus are discarded in later steps in the genomics pipeline, wasting the basecalling computation. To overcome this issue, we propose TargetCall, the first fast and widely-applicable pre-basecalling filter to eliminate the wasted computation in basecalling. TargetCall's key idea is to discard reads that will not match the target reference (i.e., off-target reads) prior to basecalling. TargetCall consists of two main components: (1) LightCall, a lightweight neural network basecaller that produces noisy reads; and (2) Similarity Check, which labels each of these noisy reads as on-target or off-target by matching them to the target reference. TargetCall filters out all off-target reads before basecalling; and the highly-accurate but slow basecalling is performed only on the raw signals whose noisy reads are labeled as on-target. Our thorough experimental evaluations using both real and simulated data show that TargetCall 1) improves the end-to-end basecalling performance of the state-of-the-art basecaller by 3.31x while maintaining high (98.88%) sensitivity in keeping on-target reads, 2) maintains high accuracy in downstream analysis, 3) precisely filters out up to 94.71% of off-target reads, and 4) achieves better performance, sensitivity, and generality compared to prior works. We freely open-source TargetCall at https://github.com/CMU-SAFARI/TargetCall.

To achieve autonomy in a priori unknown real-world scenarios, agents should be able to: i) act from high-dimensional sensory observations (e.g., images), ii) learn from past experience to adapt and improve, and iii) be capable of long horizon planning. Classical planning algorithms (e.g. PRM, RRT) are proficient at handling long-horizon planning. Deep learning based methods in turn can provide the necessary representations to address the others, by modeling statistical contingencies between observations. In this direction, we introduce a general-purpose planning algorithm called PALMER that combines classical sampling-based planning algorithms with learning-based perceptual representations. For training these perceptual representations, we combine Q-learning with contrastive representation learning to create a latent space where the distance between the embeddings of two states captures how easily an optimal policy can traverse between them. For planning with these perceptual representations, we re-purpose classical sampling-based planning algorithms to retrieve previously observed trajectory segments from a replay buffer and restitch them into approximately optimal paths that connect any given pair of start and goal states. This creates a tight feedback loop between representation learning, memory, reinforcement learning, and sampling-based planning. The end result is an experiential framework for long-horizon planning that is significantly more robust and sample efficient compared to existing methods.

Gathering properly labelled, adequately rich, and case-specific data for successfully training a data-driven or hybrid model for structural health monitoring (SHM) applications is a challenging task. We posit that a Transfer Learning (TL) method that utilizes available data in any relevant source domain and directly applies to the target domain through domain adaptation can provide substantial remedies to address this issue. Accordingly, we present a novel TL method that differentiates between the source's no-damage and damage cases and utilizes a domain adaptation (DA) technique. The DA module transfers the accumulated knowledge in contrasting no-damage and damage cases in the source domain to the target domain, given only the target's no-damage case. High-dimensional features allow employing signal processing domain knowledge to devise a generalizable DA approach. The Generative Adversarial Network (GAN) architecture is adopted for learning since its optimization process accommodates high-dimensional inputs in a zero-shot setting. At the same time, its training objective conforms seamlessly with the case of no-damage and damage data in SHM since its discriminator network differentiates between real (no damage) and fake (possibly unseen damage) data. An extensive set of experimental results demonstrates the method's success in transferring knowledge on differences between no-damage and damage cases across three strongly heterogeneous independent target structures. The area under the Receiver Operating Characteristics curves (Area Under the Curve - AUC) is used to evaluate the differentiation between no-damage and damage cases in the target domain, reaching values as high as 0.95. With no-damage and damage cases discerned from each other, zero-shot structural damage detection is carried out. The mean F1 scores for all damages in the three independent datasets are 0.978, 0.992, and 0.975.

In this paper, we aim to address the large domain gap between high-resolution face images, e.g., from professional portrait photography, and low-quality surveillance images, e.g., from security cameras. Establishing an identity match between disparate sources like this is a classical surveillance face identification scenario, which continues to be a challenging problem for modern face recognition techniques. To that end, we propose a method that combines face super-resolution, resolution matching, and multi-scale template accumulation to reliably recognize faces from long-range surveillance footage, including from low quality sources. The proposed approach does not require training or fine-tuning on the target dataset of real surveillance images. Extensive experiments show that our proposed method is able to outperform even existing methods fine-tuned to the SCFace dataset.

Resistive Random-Access Memory (RRAM) is well-suited to accelerate neural network (NN) workloads as RRAM-based Processing-in-Memory (PIM) architectures natively support highly-parallel multiply-accumulate (MAC) operations that form the backbone of most NN workloads. Unfortunately, NN workloads such as transformers require support for non-MAC operations (e.g., softmax) that RRAM cannot provide natively. Consequently, state-of-the-art works either integrate additional digital logic circuits to support the non-MAC operations or offload the non-MAC operations to CPU/GPU, resulting in significant performance and energy efficiency overheads due to data movement. In this work, we propose NEON, a novel compiler optimization to enable the end-to-end execution of the NN workload in RRAM. The key idea of NEON is to transform each non-MAC operation into a lightweight yet highly-accurate neural network. Utilizing neural networks to approximate the non-MAC operations provides two advantages: 1) We can exploit the key strength of RRAM, i.e., highly-parallel MAC operation, to flexibly and efficiently execute non-MAC operations in memory. 2) We can simplify RRAM's microarchitecture by eliminating the additional digital logic circuits while reducing the data movement overheads. Acceleration of the non-MAC operations in memory enables NEON to achieve a 2.28x speedup compared to an idealized digital logic-based RRAM. We analyze the trade-offs associated with the transformation and demonstrate feasible use cases for NEON across different substrates.

A new development in NLP is the construction of hyperbolic word embeddings. As opposed to their Euclidean counterparts, hyperbolic embeddings are represented not by vectors, but by points in hyperbolic space. This makes the most common basic scheme for constructing document representations, namely the averaging of word vectors, meaningless in the hyperbolic setting. We reinterpret the vector mean as the centroid of the points represented by the vectors, and investigate various hyperbolic centroid schemes and their effectiveness at text classification.

Facial recognition is fundamental for a wide variety of security systems operating in real-time applications. In video surveillance based face recognition, face images are typically captured over multiple frames in uncontrolled conditions; where head pose, illumination, shadowing, motion blur and focus change over the sequence. We can generalize that the three fundamental operations involved in the facial recognition tasks: face detection, face alignment and face recognition. This study presents comparative benchmark tables for the state-of-art face recognition methods by testing them with same backbone architecture in order to focus only on the face recognition solution instead of network architecture. For this purpose, we constructed a video surveillance dataset of face IDs that has high age variance, intra-class variance (face make-up, beard, etc.) with native surveillance facial imagery data for evaluation. On the other hand, this work discovers the best recognition methods for different conditions like non-masked faces, masked faces, and faces with glasses.