This paper presents a new approach for supervised power disaggregation by using deep recurrent long short term memory network. It is useful to extract the signal of one dominant appliance or any subcircuit from aggregate signal. To train network, measurement target in addition total during same time period required. The method supervised, but less restrictive practice since submetering an important feasible. main advantages this are: a) also applicable variable load and not restricted on-off...

Nowadays, renewable energies play an important role to cover the increasing power demand in accordance with environment protection. Solar energy, produced by large solar farms, is a fast growing technology offering environmental friendly supply. However, its efficiency suffers from cell defects occurring during operation life or caused incidents. These can be made visible using electroluminescence (EL) imaging. A manual classification of these EL images very time and cost demanding prone...

Efficient deep neural network (DNN) inference on mobile or embedded devices typically involves quantization of the parameters and activations. In particular, mixed precision networks achieve better performance than with homogeneous bitwidth for same size constraint. Since choosing optimal bitwidths is not straight forward, training methods, which can learn them, are desirable. Differentiable straight-through gradients allows to quantizer's using gradient methods. We show that a suited...

This paper presents a new supervised approach to extract the power trace of individual loads from single channel aggregate signals in non-intrusive load monitoring (NILM) systems. Recent approaches this source separation problem are based on factorial hidden Markov models (FHMM). Drawbacks needed knowledge HMM for all loads, what is infeasible large buildings, and combinatorial complexity. Our trains with two emission probabilities, one be extracted other signal. A Gaussian distribution used...

We systematically investigate multi-token prediction (MTP) capabilities within LLMs pre-trained for next-token (NTP). first show that such models inherently possess MTP via numerical marginalization over intermediate token probabilities, though performance is data-dependent and improves with model scale. Furthermore, we explore the challenges of integrating heads into frozen find their hidden layers are strongly specialized NTP, making adaptation non-trivial. Finally, while joint training...

The problem of identifying end-use electrical appliances from their individual consumption profiles, known as the appliance identification problem, is a primary stage in both Non-Intrusive Load Monitoring (NILM) and automated plug-wise metering. Therefore, has received dedicated studies with various electric signatures, classification models, evaluation datasets. In this paper, we propose neural network ensembles approach to address using high resolution measurements. models are trained on...

Operating deep neural networks (DNNs) on devices with limited resources requires the reduction of their memory as well computational footprint. Popular methods are network quantization or pruning, which either reduce word length parameters remove weights from if they not needed. In this article we discuss a general framework for call `Look-Up Table Quantization` (LUT-Q). For each layer, learn value dictionary and an assignment matrix to represent weights. We propose special solver combines...

Deep neural networks (DNN) are powerful models for many pattern recognition tasks, yet they tend to have layers and neurons resulting in a high computational complexity. This limits their application high-performance computing platforms. In order evaluate trained DNN on lower-performance platform like mobile or embedded device, model reduction techniques which shrink the network size reduce number of parameters without considerable performance degradation highly desirable. this paper, we...

In recent years, the rapid evolution of computer vision has seen emergence various foundation models, each tailored to specific data types and tasks. this study, we explore adaptation these models for few-shot semantic segmentation. Specifically, conduct a comprehensive comparative analysis four prominent models: DINO V2, Segment Anything, CLIP, Masked AutoEncoders, straightforward ResNet50 pre-trained on COCO dataset. We also include 5 methods, ranging from linear probing fine tuning. Our...

In this paper we examine the use of deep convolutional neural networks for semantic image segmentation, which separates an input into multiple regions corresponding to predefined object classes. We encoder-decoder structure and aim improve it in convergence speed segmentation accuracy by adding shortcuts between network layers. Besides, investigate how extend already trained model other new propose a strategy class extension with only little training data labels. experiments two street scene...

Deep neural networks (DNN) achieve very good performance in many machine learning tasks, but are computationally demanding. Hence, there is a growing interest on model reduction methods for DNN. Model allows to reduce the number of computations needed evaluate trained DNN without significant degradation. In this paper, we study layerwise that each layer independently. We consider pruning and low-rank approximation method reduction. Up now, often constant factor used all layers. show...

The estimation of the generalization error classifiers often relies on a validation set. Such set is hardly available in few-shot learning scenarios, highly disregarded shortcoming field. In these it common to rely features extracted from pre-trained neural networks combined with distance-based such as nearest class mean. this work, we introduce Gaussian model feature distribution. By estimating parameters model, are able predict new classification tasks few samples. We observe that accurate...

Magnetic resonance (MR) plays an important role in medical imaging. It can be flexibly tuned towards different applications for deriving a meaningful diagnosis. However, its long acquisition times and flexible parametrization make it on the other hand prone to artifacts which obscure underlying image content or misinterpreted as anatomy. Patient-induced motion are still one of major extrinsic factors degrade quality. In this work, automatic reference-free artifact detection, including...

The deep convolutional neural network (CNN) has recently shown state-of-the-art performance in many image processing tasks. We examine the use of CNN for semantic segmentation, which separates an input into multiple regions corresponding to predefined object classes. follow most successful CNN-based segmentation recent years and focus on study contextual aspects. To context-awareness, we manually modify context images effects results. experiments through systematic changes show that model is...

In this paper, we propose a new layerwise pruning method to reduce the number of computations needed evaluate convolutional neural networks (CNN) after training. This least-squares (LS) based improves state-of-the-art methods as it solves both problems, how select feature maps be pruned and adapt remaining parameters in kernel tensor compensate introduced errors, jointly. Therefore, our utilizes correlations between input structure tensor. experiments, show that high reduction rates with...

In clinical diagnostic, magnetic resonance imaging (MRI) is a valuable and versatile tool. The acquisition process is, however, susceptible to image distortions (artifacts) which may lead degradation of quality. Automated reference-free localization quantification artifacts by employing convolutional neural networks (CNNs) promising way for early detection artifacts. Training relies on high amount expert labeled data time-demanding process. Previous studies were based global labels, i.e....

Machine learning models are advancing circuit design, particularly in analog circuits. They typically generate netlists that lack human interpretability. This is a problem as designers heavily rely on the interpretability of diagrams or schematics to intuitively understand, troubleshoot, and develop designs. Hence, integrate domain knowledge effectively, it crucial translate ML-generated into interpretable quickly accurately. We propose Schemato, large language model (LLM) for...

We consider the problem of zero-shot one-class visual classification. In this setting, only label target class is available, and goal to discriminate between positive negative query samples without requiring any validation example from task. propose a two-step solution that first queries large language models for visually confusing objects then relies on vision-language pre-trained (e.g., CLIP) perform By adapting large-scale vision benchmarks, we demonstrate ability proposed method...

Handling distribution shifts from training data, known as out-of-distribution (OOD) generalization, poses a significant challenge in the field of machine learning. While pre-trained vision-language model like CLIP has demonstrated remarkable zero-shot performance, further adaptation to downstream tasks leads undesirable degradation for OOD data. In this work, we introduce Sparse Adaptation Fine-Tuning (SAFT), method that prevents fine-tuning forgetting general knowledge model. SAFT only...