context stringlengths 121 4.86k | A stringlengths 101 2.98k | B stringlengths 105 4.3k | C stringlengths 101 4.65k | D stringlengths 104 1.69k | label stringclasses 4
values |
|---|---|---|---|---|---|
Similar work [26] considers the identical sampling model as ours, where H𝐻Hitalic_H consists of the product of two orthogonal basis, and the result shows that m≥Cμ2klog(n/δ)𝑚𝐶superscript𝜇2𝑘𝑛𝛿m\geq C{{\mu}^{2}}k\log(n/\delta)italic_m ≥ italic_C italic_μ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT italic_k rom... | From our result (Theorem 1), it can be observed that when H𝐻Hitalic_H is a matrix composed of the product of two orthogonal basis, the result is consistent with the work [26]. | The result of the experiment is shown in Fig. 6. It is clear from Fig. 6. that the proposed variable density sampling strategy can improve the recovery performance to some extent. | Similar work [26] considers the identical sampling model as ours, where H𝐻Hitalic_H consists of the product of two orthogonal basis, and the result shows that m≥Cμ2klog(n/δ)𝑚𝐶superscript𝜇2𝑘𝑛𝛿m\geq C{{\mu}^{2}}k\log(n/\delta)italic_m ≥ italic_C italic_μ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT italic_k rom... | where the signal is a linear combination of a few atoms in the dictionary. In addition, according to the definition of our sparsity condition number, our result reduces the lower bound of the number of samples compared to [30]. | A |
To further evaluate the adaptability of Diff4MMLiTS, we use three backbone models in the MS module, namely U-Net, AttentionUNet, and nnUNet, with results presented in Table III. The findings indicate that our framework adapts seamlessly to all backbones, achieving notable performance improvements. Compared to segmentat... | Figure 1: The architecture of Diff4MMLiTS. Normal CT Generator (NCG) module uses the extended PVP mask to inpaint multimodal images to acquire normal CTs. Multimodal CT Synthesizer (MCS) module uses normal CTs to synthesize multimodal CTs. Multimodal Segmenter (MS) module trains segmenter using real and synthetic data.... | Table IV: Results Using Only a Portion of Real Diseased CT and Corresponding Normal CT Training. Note that uni is Short for Unimodal and multi is Short for Multimodal. | As illustrated in Table IV, we further evaluate the performance of the synthesis strategy on multimodal and unimodal segmentation methods. In all quantity settings, we employ nnUNet as the segmentation model architecture. The performance of unimodal segmentation models typically rely heavily on the quantity and diversi... | We fine-tune the model using pre-trained parameters and followed the training strategy of Suvorov et al [28]. For the fine-tuning of NCG, we employ original internal multimodal data to adapt it to the inpainting of multimodal CT images. After training, the inpainter is utilized to remove tumor foreground regions from m... | B |
Most datasets (Plotz and Roth 2017; Xu et al. 2018; Nam et al. 2016) rely on multi-frame averaging, which is not only challenging to obtain but also fails to provide diverse noise types and cannot address structural noise. Some approaches (Foi et al. 2008; Foi 2009; Brooks et al. 2019) model noise as Gaussian white noi... | In this paper, we introduce Realistic Noise Synthesize Diffusor (RNSD), a novel method for synthesizing realistic rgb noise data based on the diffusion model. RNSD has the capability to generate a large amount of noise images that closely resemble the distribution of real-world noise by clean images from various public... | Visual Analysis of Denoising Performance. It is evident from Fig. 6 that the DnCNN (Zhang et al. 2017) trained on synthetic noise samples generated by RNSD significantly outperforms other methods in terms of denoising effectiveness, closely matching the performance of the DnCNN trained with realistic noise data. | Enhance Denoising Performance with Data Augmentation. In the actual training process of denoising models, the performance of the model often decreases due to insufficient noise samples and limited sampling scenarios in the dataset. Our RNSD method can improve model performance by increasing noise samples and augmenting... | In all experiments, we train RNSD on SIDD small. We augment the noise sampling and scene sampling of the denoising dataset by using RNSD to generate noisy samples from the clean samples of SIDD small and 1000 randomly selected high-quality samples from LSDIR, respectively. The SIDD validation set and DND benchmark are ... | A |
We report the average running time of our method and recent proposed deep learning-based methods on LOL dataset in Fig. 9. Also, we list model parameters of all test methods in Fig. 9. Although LLFormer and MIRNet can achieve noticeable performance, they are heavy in model size. URetinex and Zero-DCE show compact model... | We propose a dual degradation model based on degradation specificity of low-light images on different spaces. It is unfolded to form dual degradation-inspired deep unfolding network for low-light image enhancement, which can jointly learn two degradation priors from luminance space and chrominance space. More important... | In this paper, we have proposed a dual degradation-inspired deep unfolding method (DASUNet) for low-light image enhancement. Specifically, we design a dual degradation model (DDM) based on the degradation specificity among luminance and chrominance spaces. An alternative optimization solution is proposed to solve it an... | To push the frontiers of deep unfolding-based image enhancement, we propose a Dual degrAdation-inSpired deep Unfolding network, termed DASUNet, for low-light image enhancement, which is shown in Fig. 2. The motivation originates from the degradation specificity of low-light images between luminance and chrominance spac... | In this section, we firstly introduce our designed dual degradation model (DDM) as the objective function. Then, an iterative optimization solution is designed as the solver of DDM. Moreover, we elaborate on dual degradation-inspired unfolding network (DASUNet) for low-light image enhancement, which is delineated on Fi... | B |
Throughout the training and testing phases, safety constraints were incorporated to ensure that the control actions remained within safe operating limits. The method successfully avoided unsafe states, thereby ensuring reliable and safe operation. | While the proposed PPO-based reinforcement learning (RL) approach for DC-DC boost converter control shows significant promise, | This section depicts the proposed control method for the DC-DC boost converter. This paper introduces an improved method for controlling dc-to-dc boost converters, which combines | The computational complexity of the proposed RL-based control method is slightly higher than that of traditional control methods. | The proposed PPO-based reinforcement learning (RL) method for DC-DC boost converter control does have slightly higher computational demands compared to traditional control methods. This is primarily due to the complexity of the PPO algorithm and the | D |
In order to assess the performance of the proposed scheme, the fault detection model is evaluated with semi-supervised learning using self-training as in [31]. | The 11520 faults and 2400 non-fault events are pre-trained by removing the target class labels from the data and then this architecture having the pre-trained model weights is trained again with 5%, 25% and 50% labeled datasets. η𝜂\etaitalic_η of 90.3%, 96.6% and 98% respectively show that fully supervised ICTT model ... | The number of fault events (11520) and non-fault events (2400) results in an imbalanced dataset, with faults and non-faults forming the majority and minority classes respectively. | The proposed scheme identified the faults from transients with 99.7% η𝜂\etaitalic_η on 2400 transients and 2520 faults simulated by varying v𝑣vitalic_v (8, 9, 11, and 22m/s), Rfsubscript𝑅𝑓R_{f}italic_R start_POSTSUBSCRIPT italic_f end_POSTSUBSCRIPT (0.01, 1, and 10ΩΩ\Omegaroman_Ω), 10 fault types, and 21 (FIT) at t... | The second stage incorporates the AR coefficient-based fuzzy inference system for fault detection which is supervised by the AR coefficient-based InceptionTime (ICTT) model. Third, after faults are detected, the faulty region is determined by differentiating faults among internal, forward, and backward faults. If the f... | A |
We offer a self-supervised deep-learning-based mathematical framework for concurrently estimating motion correction and signal decay model parameters. | We introduce an innovative registration loss function, guaranteeing physically sound deformation fields that align with the signal decay model. | Figure 3: Signal decay along the relaxation axis (b-value) after registration with SyN-Reg to b0, Iterative SyN-TRF, and IVIM-Morph. Although IVIM-Morph yielded a lower Dice score in this case compared to other methods, it better preserved the expected signal decay behavior, reflecting a more physically plausible regis... | In this study, we tackle this challenge with the introduction of a self-supervised Deep Neural Network (DNN) framework named “IVIM-Morph.” This approach addresses simultaneous motion compensation and bi-exponential IVIM model parameter estimation. Our model comprises two key sub-networks: the first focuses on estimatin... | We introduce an innovative loss function comprising a weighted combination of the following three terms: | A |
In the literature, drone detection, recognition, localization, and tracking has been intensively studied, e.g., via computer vision (CV), acoustic arrays, radio frequency (RF) fingerprinting, and millimeter wave (mmWave) radar systems [1]. For instance, by equipping high-definition cameras on the roof of the buildings,... | In the paper, we extend the study from indoor scenario to outdoor scenario, and introduce RIS for performance enhancement of drone detection. We study the passive drone detection with the assistance of current cellular multiple-input multiple-output (MIMO) BS along with a passive RIS via the generalized likelihood rati... | A signaling link is established from the BS to the RIS, either via wireless or wired connection, for the purpose of coordination of RIS training beam design during the pilot transmission phase. The BS, the RIS, and the UE have an uniform planar/rectangular array (UPA/URA) structure. To be specific, the BS antenna array... | During the sounding stage, the BS sends beamformed pilot signals illuminating the RIS and a certain part of the sky simultaneously, meanwhile the RIS steers the beams to cover a certain range of the sky, where the drone is potentially located. By following this, we can collaboratively cover a larger space of the sky co... | Recently, there are ongoing activities under the IEEE 802.11bf on wireless sensing, which include absence/presence detection of human being, gesture recognition, object tracking, based on the received WiFi signals [2]. However, the activity is mainly constrained to the indoor scenario. The sensing capability of the cel... | D |
All the papers presented so far have focused on improving ID and OOD calibration for FNNs, and are thus limited in the capacity to provide reliable decisions that fully account for epistemic uncertainty. | This paper proposes an extension of variational inference (VI)-based Bayesian learning that integrates calibration regularization for improved ID performance [4], confidence minimization for enhanced OOD detection [11], and selective calibration to ensure a synergistic use of calibration regularization and confidence m... | So far, we have seen that there is generally a trade-off between ID and OOD performance. In particular, from Fig. 9, it was concluded that, for a given fixed level of OOD detection probability, calibration-regularized learning improved ID calibration at the cost of the ID accuracy for BNN-OCM. In this subsection, we as... | In this paper, we have proposed SCBNN-OCM, a general framework that generalizes variational inference-based Bayesian learning to target both ID and OOD calibration. To improve ID calibration, we have introduced a regularizer based on the calibration error, while OOD calibration is enhanced by means of data augmentation... | This paper proposes an extension of VI-based BNN learning that integrates calibration regularization for improved ID performance, OCM for OOD detection, and selective calibration to ensure a synergistic use of calibration regularization and OCM. As illustrated in Fig. 1, the scheme is constructed successively by first ... | D |
We examine the impact of training overhead on performance. Various levels of training overhead, i.e., K𝐾Kitalic_K values, are considered. The simulation results are illustrated in Fig. 4. As expected, higher training overhead leads to improved drone localization performance. | In this section, we evaluate the performance of the RIS-assisted drone localization systems using the following parameter configuration: MB=8×8,MU=4×4formulae-sequencesubscript𝑀B88subscript𝑀U44M_{\text{B}}=8\times 8,M_{\text{U}}=4\times 4italic_M start_POSTSUBSCRIPT B end_POSTSUBSCRIPT = 8 × 8 , italic_M start_POSTSU... | In this work, we only consider the angular parameter estimates for passive 3D drone localization. The transform matrix, a.k.a. Jacobian matrix, is calculated based on the geometrical relationship among the network nodes [10]. In its first six columns, the transform matrix 𝐓∈ℝ3×10𝐓superscriptℝ310{\mathbf{T}}\in\mathbb... | We examine the impact of training overhead on performance. Various levels of training overhead, i.e., K𝐾Kitalic_K values, are considered. The simulation results are illustrated in Fig. 4. As expected, higher training overhead leads to improved drone localization performance. | The RCS values play a pivotal role in affecting the performance. Hence, we evaluate the following three cases: ζ∈{0.5,1,2}𝜁0.512\zeta\in\{0.5,1,2\}italic_ζ ∈ { 0.5 , 1 , 2 }. The simulation results are depicted in Fig. 5. Notably, RCS values exhibit a substantial impact on the accuracy of angular parameter estimation ... | D |
In this work, we present the framework that consists a set of algorithms for efficiently solving the problem via exploiting the unique features of Spatial AirFusion. | Figure 1: (a) An ISEA system for environment perception in the context of autonomous driving. (b) Spatial AirFusion protocol. | First, the performance of Spatial AirFusion and naive AirComp is evaluated on the synthetic dataset. We test Spatial AirFusion controlled by Greedy VoCa-PPA in Algorithm 1 and Optimal VoCa-PPA in Algorithm 2, termed “AirFusion-Greedy” and “AirFusion-Optimal”, respectively. The performance is measured by AirComp error, ... | AirFusion Protocol. A communication protocol is presented to realize spatial AirFusion in a multi-agent system, comprising the following three phases. | The proposed Spatial AirFusion framework aims at efficiently aggregating multi-agent voxel features over a broadband channel, where the feature vectors on different agents but attributed to the same voxel are aggregated over a particular subcarrier. Targeting environment perception, Spatial AirFusion is differentiated ... | C |
)}))^{2}+\lambda\;\Psi(\Theta).roman_min start_POSTSUBSCRIPT roman_Θ end_POSTSUBSCRIPT ∑ start_POSTSUBSCRIPT italic_i = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_n end_POSTSUPERSCRIPT ( bold_O start_POSTSUBSCRIPT ( italic_i , italic_N + 1 ) end_POSTSUBSCRIPT - italic_f start_POSTSUBSCRIPT roman_Θ end_POSTSUBSCRI... | We remark that the difference-based NeurTV has stronger requirements to be defined, i.e., it needs a meshgrid point set to define the difference-based regularization. Comparatively, the NeurTV based on derivatives can be easily applied to both meshgrid and non-meshgrid data. Hence, in the main experiments, we have cons... | In this new continuous representation model (4), the classical discrete difference-based TV can no longer be directly applied. Hence, it needs to develop a new regularization that can capture local correlations of both meshgrid and non-meshgrid data under model (4)111We remark that the data recovery models (2) & (4) ca... | In this work, we suggest a new TV regularization by using a deep neural network (DNN) to continuously represent data. Here, the continuous representation refers to using a DNN to represent data by feeding the coordinate of data into the DNN and outputting the corresponding value, allowing the network to continuously re... | We have proposed the NeurTV regularization to capture local correlations of data based on continuous representation. As compared with classical discrete meshgrid-based TV, our NeurTV is free of discretization error induced by the difference operator, and is suitable for both meshgrid and non-meshgrid data. By virtue of... | B |
As shown in Table 1, the model demonstrated strong performance, with R² values exceeding 0.96 and MAP values above 88%. These results highlight the robustness of LSTM-FINs in learning closed-form feature representations. | In this paper, we explored the application of Feature Imitation Networks (FINs) to represent four temporal sEMG signal features. Our LSTM-based FIN accurately simulated these features with R² >>> 0.96 and MAP >>> 88%. The CNN classifier with feature augmentation outperformed traditional SVM classifiers, and when combin... | The target variable was a one-hot vector representing the 17 hand movements. Following the dataset’s split method [17], repetitions 2 and 5 were designated as the test set, while 1, 3, 4, and 6 were for training. We further divided this into “within-subjects” (1-25) and “cross-subjects” (26-40) sets to explore transfer... | We combined the LSTM-FIN and CNN-II models in four steps: (1) Pre-train the LSTM-FIN and CNN classifier separately with original data and ground-truth features. (2) Initialize the encoder (LSTM-FIN) and decoder (CNN) with pre-trained weights from corresponding subjects. (3) Integrate encoder and decoder models, enablin... | We evaluated the performance of both Feature Engineering and Deep Learning, to be specific: (1) SVM: A traditional feature engineering approach using an SVM model with a Gaussian kernel, trained on ground-truth features; (2) CNN: CNN models trained on ground-truth features, implemented in two variations, (a) CNN-I with... | D |
DQN is a robust DRL method that integrates the traditional Q-learning algorithm with deep neural networks. DQN is particularly effective in environments with discrete action spaces, making it ideal for applications such as discrete bandwidth resource block allocation to the users. It is worth noting that the available ... | Here, we introduce a joint DRL-based algorithm devised for optimizing resource allocation to users and maximizing UAV utility. In this algorithm, the DQN model is employed to allocate optimal bandwidth to each user, considering its allocated power value, channel condition, and path loss effects. Subsequently, based on ... | The objective of the DQN model is to allocate optimal bandwidth resource blocks to users for various scenarios in the shortest possible time. | In the state updation, the following steps are performed: First, generate a continuous change in the power levels for all users based on the current policy. Second, apply this change to the current state to update the power levels. Third, use the trained DQN model to determine the optimal bandwidth allocation for each ... | Recent advances in DRL provide promising solutions for resource allocation in dynamic environments. These algorithms learn optimal policies through environmental interactions, making them ideal for complex scenarios. The work in [10] seeks to maximize the UAV’s service time and downlink throughput by utilizing the DDPG... | B |
Nguyen et al. [52] propose a multi-accent TTS framework utilizing a weight factorization approach. They decompose each weight matrix of the letter-to-sound component into shared and accent-dependent factors. | Zhong et al. [56] introduce a two-stage training pipeline for zero-shot accent generation. They first train a speaker-independent accent encoder and then build an accented TTS system conditioned on the pre-trained accent encoder. | Zhang et al. [53] develop a multi-accent TTS system that controls accents in the encoder, which is trained on an auxiliary accent classification task to generate multi-accent phoneme representations. | To model speaker identity, we use the widely adopted technique of speaker embedding [14, 15, 16], which learns speaker-discriminative representations from large-scale speaker datasets trained on a speaker classification task [17]. The speaker embedding, when integrated into multi-speaker TTS systems, effectively contro... | To meet these demands, it is crucial to develop multi-speaker multi-accent TTS systems that can generate voices of multiple speakers, each with various accents. | B |
The impact of the data window is examined using window sizes of half, one, and two cycles. η¯¯𝜂\bar{\eta}over¯ start_ARG italic_η end_ARGs of 98.0%, 98.0%, and 97.2% are obtained correspondingly, showing the scheme’s resilience to change in window size. | The technique recognized the faults and switching transients with η¯¯𝜂\bar{\eta}over¯ start_ARG italic_η end_ARG of 96.0% and 96.4% on 300 and 500 units respectively using 2400 no-fault transients and 2160 faults simulated by changing priority (P and Q), fault resistances (0.01, 1, and 10 ohms), fault types (10), and ... | The reliability of ground distance relays is threatened by the mutual coupling of double circuit t-lines, demanding additional consideration [38]. So, between buses 3PVsubscript3𝑃𝑉3_{PV}3 start_POSTSUBSCRIPT italic_P italic_V end_POSTSUBSCRIPT and 9, a 100 km long double-circuit t-line working at 230 kV and 60 Hz is... | The 39-bus system is also served by the CLT-based fault detection, which has a η¯¯𝜂\bar{\eta}over¯ start_ARG italic_η end_ARG of 98.0% on 1080 faults and 2400 non-fault cases. The priority (P and Q), fault locations (f2, f5, and f6), fault resistances (0.01, 1, and 10 ohms), fault types (10), and fault inception angle... | The performance of traditional distance relays may be impacted by TCSCs which are used to improve system stability, increase power transfer capability, and regulate voltage levels in presence of IBR [22]. In the 9-bus test system, a TCSC with a capacitor, an inductor, and a metal oxide varistor which handles overvoltag... | D |
To fuse effectively features from each modality, we introduce a novel Dual-attention Feature Calibrating (DFC) module as depicted in Fig. 2(b) which employs distinct fusion strategies for the encoder and decoder features of the two mono-modality branches. For encoder features, motivated by Vaswani et al. [33], DFC util... | Intuitionally, the uncertainty output from different models should be positively correlated to data perturbation (i.e., the output uncertainty should increase with the incresement of perturbation levels of input data). The conservation of such a positive correlation is a measure of the confidential level of the output ... | Unlike conventional segmentation approaches, our segmentation backbone replaces Softmax in the traditional neural network classifier with Softplus to facilitate the quantification of uncertainty. Thus, the outcomes derived from the segmentation backbone should not be interpreted directly as the predictive segmentation ... | Defining CT image as XCTsubscript𝑋𝐶𝑇X_{CT}italic_X start_POSTSUBSCRIPT italic_C italic_T end_POSTSUBSCRIPT and PET image as XPETsubscript𝑋𝑃𝐸𝑇X_{PET}italic_X start_POSTSUBSCRIPT italic_P italic_E italic_T end_POSTSUBSCRIPT, the two inputs (XCTsubscript𝑋𝐶𝑇X_{CT}italic_X start_POSTSUBSCRIPT italic_C italic_T... | where pXcsuperscriptsubscript𝑝𝑋𝑐p_{X}^{c}italic_p start_POSTSUBSCRIPT italic_X end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_c end_POSTSUPERSCRIPT and yXcsuperscriptsubscript𝑦𝑋𝑐y_{X}^{c}italic_y start_POSTSUBSCRIPT italic_X end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_c end_POSTSUPERSCRIPT are the predicted pr... | B |
Generative models have recently achieved substantial success for text-to-audio generation. In particular, the development of language models [1, 2] and diffusion models [3, 4] have enabled the creation of powerful systems [5, 6] on generating high-fidelity audio clips. | We present Sound-VECaps, a large-scale dataset comprising 1.661.661.661.66M audio clips with captions augmented by video data, to address the challenge of prompt following in audio generation systems. Experiments show that systems trained on Sound-VECaps achieve SoTA performance and outperform baseline models. In addit... | To evaluate the effectiveness of the visual information in the captions, we compare the performance of different AudioLDM-T5-L systems trained and evaluated on various datasets that include and exclude visual-only content. Notably, all three versions of the testing dataset share the same group of audio clips (same targ... | Despite their success in generating audio with simple captions, current models struggle with complex prompts containing detailed information, which referred to the challenge as “prompt following” [3]. A potential reason for this limitation is that existing audio-caption datasets often lack in quantity and quality (deta... | In addition to our experiments on audio generation, we evaluated the effectiveness of Sound-VECaps for improving audio-language retrieval systems. Specifically, we employed the framework in WavCaps [7], which uses RoBERTa as the text encoder and HTSAT as the audio encoder, to train and evaluate CLAP-based models in aud... | C |
Typically, knowledge is distilled from a group of teacher networks in FL, each trained on data from a different location [52, 53]. | Feature-based KD, on the other hand, aligns the intermediate feature representations of both models, requiring compatible architectures to effectively match features at specific layers [26, 27]. | In both models the last layer is a 1×1111\times 11 × 1 convolution that only reweighs the feature maps from the previous layer. | For segmenting the coronary arteries, we restrict ourselves to only finetune the last output layer of both models trained with KD, the UNet and SWIN-UNETR, to assess the extent of feature extraction already achieved by the backbone of the federated model. | Other methods include distilling knowledge by matching attention maps between client models or aligning the feature maps of both models [23, 54, 55]. | D |
After obtaining the solutions of three sub-problems, the proposed AO algorithm iteratively solves the three sub-problems until the increase of RSSRsubscript𝑅SSRR_{\mathrm{SSR}}italic_R start_POSTSUBSCRIPT roman_SSR end_POSTSUBSCRIPT is less than the threshold εAOsubscript𝜀AO\varepsilon_{{\mathrm{AO}}}italic_ε start_P... | The inequality marked by (α5)subscript𝛼5\left({{\alpha_{5}}}\right)( italic_α start_POSTSUBSCRIPT 5 end_POSTSUBSCRIPT ) holds because 𝐛kU(c)superscriptsubscript𝐛subscript𝑘U𝑐{\mathbf{b}_{{k_{\mathrm{U}}}}^{\left(c\right)}}bold_b start_POSTSUBSCRIPT italic_k start_POSTSUBSCRIPT roman_U end_POSTSUBSCRIPT end_POSTSUBS... | In Fig. 3a, the convergence of the proposed MVPSO is evaluated with different moving region sizes and path numbers. As can be observed, the three SSRs increase with the number of iterations and tend towards stable values within 100 iterations, validating the convergence performance. Additionally, to verify the effectiv... | The convergence and computational complexity of the overall AO algorithm are analyzed as follows. Define the iteration index and maximum number of iterations for AO as c𝑐citalic_c and C𝐶Citalic_C, respectively, where 1≤c≤C1𝑐𝐶1\leq c\leq C1 ≤ italic_c ≤ italic_C. The convergence is ensured by the following inequalit... | Besides, the convergence evaluation of the overall AO algorithm is shown in Fig. 3b. With different moving region sizes, numbers of paths, and numbers of MAs, the SSRs increase with the number of AO iterations and converge within 20 iterations, substantiating the previous discussions on the convergence of the AO algori... | C |
Table 2: Cumulative return over 1000100010001000 episodes with both shielding and learning in either of the state spaces. | Our motivation for applying a state-space transformation is to better align with a grid, and ultimately to make the synthesis more scalable. | We believe that they are strictly necessary when applying shield synthesis to many practical systems due to state-space explosion. | In this section, we show how a transformation of the state space can be used for grid-based shield synthesis, and demonstrate that it can be instrumental. | The only motivation for applying a state-space transformation was to be able to compute a cheaper shield. | D |
In addition, we can compute other partial derivatives, including ∂μ~n∂pnsubscript~𝜇𝑛subscript𝑝𝑛\frac{\partial\tilde{\mu}_{n}}{\partial p_{n}}divide start_ARG ∂ over~ start_ARG italic_μ end_ARG start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT end_ARG start_ARG ∂ italic_p start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT ... | The notion of the dominating point is useful for quantifying reserve requirements for rare and extreme events in power systems. Although multiple events could push the system out of its safe operating region, we only need to focus on the most critical scenario — the one with the highest probability of driving the syste... | Here, Greek letters in parentheses on the left denote dual multipliers of the corresponding constraints. The additional decision variable λ+∗superscript𝜆+\lambda^{\texttt{+}*}italic_λ start_POSTSUPERSCRIPT + ∗ end_POSTSUPERSCRIPT is the optimal value of the dual multiplier associated with the generator limit constrai... | Fig. 1 compares the risk hedging strategies of a regular CC and an LDT-CC under uncertainty 𝛀𝛀\boldsymbol{\Omega}bold_Ω. The red area represents deviations occurring during regular scenarios, where the cumulative probability (1−ϵ1italic-ϵ1-\epsilon1 - italic_ϵ) is hedged by the regular CC. In contrast, the yellow are... | Although the cutting-plane algorithm typically does not guarantee polynomial-time convergence [24], the convexity of the proposed models, combined with the compactness and non-emptiness of the feasible region, ensures convergence to the optimal solution [19]. | D |
In (2), we observe that the total electricity procured by a mining facility (EtMsubscriptsuperscript𝐸𝑀𝑡E^{M}_{t}italic_E start_POSTSUPERSCRIPT italic_M end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT) is equal to the sum of electricity purchased through long-term power purchase agreements (EtPsubs... | The response of other industrial facilities to electricity prices has already been thoroughly discussed in the literature (see [Golmohamadi, 2022] for a recent review article). For example, in the aluminium smelting industry, [Depree et al., 2022] have discussed ‘arbitrage price,’ which identifies a correlation between... | The scatter plots of the RSI of the Bitcoin exchange rate considering the 7, 14, and 21-day window and the daily energy consumption of crypto-mining firms, depicted in Fig. 4, show p-values of correlation coefficient of ≥\geq≥ 0.05 in all three cases. This suggests that, given the panel data concerned, cryptocurrency m... | In regards to cryptocurrency mining firms, [Rhodes et al., 2021] have already provided high-level behavioral analyses of cryptocurrency mining firms. [Menati et al., 2023] have studied the impacts of various demand response programs for cryptocurrency mining loads in Texas. [Menati et al., 2024] have also designed algo... | As shown in Fig. 1 (which is corroborated by slide 3 of [ERCOT, 2024]), the Texas electric grid is facing a rapid cryptocurrency mining data-center-driven load growth. The Electricity Reliability Commission of Texas (ERCOT)—the market operator in charge of the largest part of the Texas electricity grid—allows both gene... | A |
(Chung, Nagrani, and Zisserman 2018) is a speaker verification dataset. It contains only speaker information for over 6,000 speakers and 2,442 hours of multilingual videos. By referring to the (Shi et al. 2022), we utilize an English portion, which amounts to 1,326 hours. | Since the pseudo-labels are generated using the pre-trained ASR model, their transcription quality may be worse than human-annotated labels. To mitigate this issue, we combine VoxCeleb2 with the LRS3 dataset, which contains manually annotated transcriptions corresponding to 433 hours of videos. | (Afouras, Chung, and Zisserman 2018) is a widely used sentence-level lip reading dataset comprising 433 hours of English talking face videos and human-annotated transcriptions. These talking face videos are collected from TED and TEDx talks, thus contain diverse poses. | (Chung, Nagrani, and Zisserman 2018) is a speaker verification dataset. It contains only speaker information for over 6,000 speakers and 2,442 hours of multilingual videos. By referring to the (Shi et al. 2022), we utilize an English portion, which amounts to 1,326 hours. | To validate our proposed speaker adaptation method in real-world scenarios, we introduce a new dataset named VoxLRS-SA, derived from VoxCeleb2 (Chung, Nagrani, and Zisserman 2018) and LRS3 (Afouras, Chung, and Zisserman 2018) datasets. Together, these sources consist of about 1700 hours of YouTube videos, and as a resu... | B |
The SSR-Speech model has the same architecture as VoiceCraft, which consists of 16 Transformer layers with hidden size of 2048 and 12 attention heads. The output of the final layer is passed through four separate 2-layer MLP modules to generate prediction logits. Following VoiceCraft, we employ the ScaledAdam optimizer... | Specifically, we mask the edited segments of the original waveform with silence clips and then use a masked encoder to extract the features from this masked waveform. The masked encoder shares the same architecture as the Encodec encoder and is initialized with parameters from Encodec. Consequently, the input to the sp... | Furthermore, we observed that CFG often generates speech at an accelerated pace due to the excessive removal of silence tokens during processing. To address this issue, we propose to utilize CFG with a stride of β𝛽\betaitalic_β during inference, where β𝛽\betaitalic_β serves as a hyperparameter. | we compare the original transcript with the target transcript to identify the words that need to be masked. Using word-level forced alignment333https://github.com/m-bain/whisperX of the original transcript, we locate the corresponding masked spans of audio tokens. The phoneme tokens from the target transcript and the u... | For inference, we use nucleus sampling [29] with p=0.8𝑝0.8p=0.8italic_p = 0.8 and a temperature of 1. The extended masked span α𝛼\alphaitalic_α is set to 0.120.120.120.12 seconds. Based on initial experiments, we determined the optimal value for the hyperparameter γ𝛾\gammaitalic_γ to be 1.51.51.51.5 and β𝛽\betaital... | D |
In recent years, convolutional neural networks (CNNs) have been widely used for OCTA image segmentation [5, 6, 7]. For example, UNet’s symmetric encoder-decoder structure and skip connections extract features at different levels, enabling efficient feature transformation and laying the foundation for medical segmentati... | To address these issues, we meticulously designed a U-shaped network for OCTA vasculature segmentation based on the Mamba architecture: OCTAMamba. This network effectively captures local information and models long-range dependencies while extracting multi-scale information from OCTA images to enhance feature represent... | In recent years, convolutional neural networks (CNNs) have been widely used for OCTA image segmentation [5, 6, 7]. For example, UNet’s symmetric encoder-decoder structure and skip connections extract features at different levels, enabling efficient feature transformation and laying the foundation for medical segmentati... | In this paper, we introduced OCTAMamba, an advanced network architecture designed for the efficient and precise segmentation of OCTA vasculature. By leveraging the strengths of Mamba architecture, we developed innovative modules such as the Quad Stream Efficient Mining Embedding, Multi-Scale Dilated Asymmetric Convolut... | Recently, structured state-space sequence models (SSMs) [13, 14, 15, 16, 17], such as Mamba [18], have emerged as powerful methods for long-sequence modeling, achieving effective global modeling with linear complexity. For example, VM-Unet [19] introduces the Visual State Space (VSS) module as a foundational component ... | D |
We compare SoloAudio with DPM-TSE using in-domain data, training and testing both models on the FSD-Mix dataset under identical conditions. As shown in Table I, SoloAudio significantly outperforms DPM-TSE across all metrics. Both audio-oriented and language-oriented TSE highlight the effectiveness of SoloAudio. | Table II highlights the impact of synthetic data, showing that using TangoSyn clearly improves TSE performance on both seen and unseen data. | Table I shows the impact of adding skip connections to the DiT model, resulting in a clear performance improvement. | The primary distinction between our network architecture and DiT lies in the use of long skip connections in SoloAudio, bridging shallow and deep DiT blocks as in [26]. These skip connections create shortcuts for low-level features, streamlining the training of the entire v𝑣vitalic_v-prediction network. Furthermore, w... | To further evaluate the few-shot and zero-shot capabilities of the models, we utilized the SoloAudio model trained exclusively on TangoSyn data. For the zero-shot setting, we directly tested the model on the out-of-domain FSD-Mix test set, which contains unseen labels. In the few-shot setting, we fine-tuned the model u... | B |
Figure 4 visualizes the results of motion tracking for two representative cases for all methods. Our model outperforms other baselines by producing the best alignment results with negligible errors. This demonstrates the effectiveness of our model in correcting rigid motions and geometric distortions. | The right side of Fig. 7 illustrates the optimal epoch of model training and Dice accuracy for various training dataset sizes. Our model demonstrates rapid convergence and maintains consistently high accuracy even with datasets smaller than 50505050 samples. In contrast, baseline models such as KeyMorph and DeepPose sh... | To test the sub-module of our spatial-temporal approach for 4D EPI motion tracking, we treated all baselines as static models to predict motion parameters between subsequent images of the time series. For both motion correction and tracking, we tested all models on real fMRI scans with unknown motions, reporting the Di... | The right side of Fig. 5 quantitatively shows the accuracy of motion tracking comparison over varying degrees of motions and different lengths of data sequences. Our model exhibits superiority in handling real motions ranging from small to large, and it maintains comparable motion tracking accuracy when dealing with ex... | Motivated by the need for high-accuracy motion correction for multiple modalities, we present a novel motion tracking approach that corrects motion artifacts without necessitating network retraining. Our method leverages the strengths of deep learning while circumventing the need for extensive retraining, thereby offer... | C |
The evaluation was conducted in both speaker verification and speaker diarization to show that the proposed method is suitable for general-purpose use. | For the one-vs-one protocol, VoxCeleb1-O [13] was used for evaluation, which consists of 37 611 trials. | The verification performance was degraded when attention masking was not used, especially in the one-vs-one scenario (P4). | Using overlaps in addition showed the same good performance in the one-vs-one scenario because there were no overlaps, and the performance of one-vs-many was more degraded (B2). | Speaker verification performance was evaluated under the standard one-vs-one protocol and novel one-vs-many protocol, respectively. | D |
})-f(x_{\tau,i})]italic_R start_POSTSUBSCRIPT italic_S italic_B end_POSTSUBSCRIPT ( italic_t ) = roman_min start_POSTSUBSCRIPT italic_i ∈ { 1 , 2 , … , italic_M } , italic_τ ∈ { 1 , 2 , … , italic_t } end_POSTSUBSCRIPT blackboard_E [ italic_f ( italic_x start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT ) - italic_f ( italic_... | In our theoretical analysis, we provide bounds on Bayesian average regret and Bayesian simple regret. | Our theoretical result bounds Bayesian average regret, RAB(t),subscript𝑅𝐴𝐵𝑡R_{AB}(t),italic_R start_POSTSUBSCRIPT italic_A italic_B end_POSTSUBSCRIPT ( italic_t ) , and Bayesian simple regret, RSB(t),subscript𝑅𝑆𝐵𝑡R_{SB}(t),italic_R start_POSTSUBSCRIPT italic_S italic_B end_POSTSUBSCRIPT ( italic_t ) , with ... | We analyze the performance of the distributed Thompson sampling algorithm on the Bayesian average regret and Bayesian simple regret metrics. Our regret bound depends on the number of timesteps T𝑇Titalic_T and the structure of the agent communication graph G. As in prior work, we utilize notions from information theory... | In this paper, we proposed a distributed Thompson sampling algorithm to address the multi-agent Bayesian optimization problem under constrained communication. We develop bounds on Bayesian average regret and Bayesian simple regret for this approach, where the bound is dependent on properties of the largest complete sub... | A |
Previous work [36] highlights that large audio-language models are proficient at audio captioning but not as effective in answering discriminative questions. | The cascade pipeline involves first using a specialized audio captioning model to generate captions for the audio. | Thus, our approach first involves having the model describe the audio information before prompting it to answer the question. | In the first round, the model might be prompted with “Describe the audio”, to which it responds with a caption: “The audio contains the sound of cars honking, people talking, and traffic signals beeping.” | This is also evident when models answer the “before” question correctly but fail on the “after” question (Correct-Incorrect, C-I). | B |
More specifically, we propose to utilize the CTC compressor to match speech and text representations from both directions. | This aligning can be further enforced from a representation level by sharing the text embeddings with the CTC class embeddings, i.e., the weights of the linear layer in the compressor. | This yields the proposed CJST for joint speech and text training, which further explores two unused features of the CTC compressor, namely, on-the-fly forced peaky alignment and CTC class embeddings. | This is done by combining a simple modality adaptor with several features of the CTC compressor, including sequence compression, on-the-fly forced peaky alignment and CTC class embeddings. | In all previous work on the CTC compressor, it is mainly used to bring the speech modality closer to the text modality, especially on the sequence length perspective. | C |
The V2I communication employs 𝒱𝒱\mathcal{V}caligraphic_V as the set of CAVs and ℰℰ\mathcal{E}caligraphic_E as the set of MEC servers. Each CAV v∈𝒱𝑣𝒱v\in\mathcal{V}italic_v ∈ caligraphic_V transmits batch of video frames (i.e., CAV-frames 𝐅cavsubscript𝐅cav\mathbf{F}_{\text{cav}}bold_F start_POSTSUBSCRIPT cav end_... | We validate the data compression achieved by the RFDVC framework compared to H.264 and H.265 codecs under different conditions as lighting changes, and rain. We also evaluate its performance under simulated packet loss to assess quality under challenging network conditions. | This theoretical framework establishes a benchmark for evaluating the practical effectiveness of the RFDVC framework and highlights the potential of leveraging RFs and Delta-frame compression under ideal conditions. | To systematically address the challenges of efficient data compression and transmission in V2I communication, we formalize the problem as an optimization task. The objective is to minimize the total amount of data transmitted while maintaining high-quality reconstruction of video frames and adhering to network constrai... | Traditional video compression depends on frequent I-frame transmission to maintain video quality, especially in scenes with high motion or complexity, as P-frames and B-frames rely on I-frames for reconstruction. Extending the GOP length can reduce the overall data size by decreasing the number of I-frames, which are t... | C |
With feature selection and model design in place, they need to be deployed on the low-power microcontroller for efficient operation. Custom implementations for the extraction of each selected feature was implemented in optimized C-code. Each function was meticulously evaluated against the initial Python code to verify ... | Afterwards, the developed ANN models were optimized for microcontroller deployment. The models were converted using the TensorFlow Lite for microcontroller framework (TFLM), applying int8 quantization to model parameters. TensorFlow Lite converts models into a format suitable for resource-constrained devices while main... | The comparison of the three approaches shows a clear trade-off between model complexity, processing time, and energy consumption. Despite its larger size and inference time, the raw-input model outperforms the other models in total prediction time and with minimal energy per prediction, making it the most suitable mode... | Conversely, while being more compact, the models operating on extracted features incurred significant computational overhead during the feature extraction process. This computational overhead resulted in longer total processing times and thus higher energy consumption of 71x and 52x of the raw-signal model, respectivel... | All three ANN models, as well as the required feature extraction algorithms have been deployed on an Arduino Nano 33 BLE Sense, encompassing the nRF52840 microcontroller. The deployed models are compared based on the inference time, feature extraction time, total processing time, model size, and energy consumption, all... | A |
Continuity and robustness hold under continuous total variation convergence of the kernels (i.e. if 𝒯n(⋅|x,un)→𝒯(⋅|x,u)\mathcal{T}_{n}(\cdot|x,u_{n})\to\mathcal{T}(\cdot|x,u)caligraphic_T start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT ( ⋅ | italic_x , italic_u start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT ) → caligr... | The above has direct implications on data-driven learning and empirical consistency, where empirical models are constructed via data, and empirical models converge weakly (and under the W1subscript𝑊1W_{1}italic_W start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT distance) almost surely ([20], Theorem 11.4.1), but do not so unde... | We refer the reader to [38, Theorem 3.2] and with further refinements under filter stability [48, Theorem 3.8] for discounted cost and [48], Theorem 3.9] for average cost. These show that the problem is robust to uncertainty in priors under total variation, and for robustness under weak convergence, total variation con... | Problem P3: Empirical Consistency of Learned Probabilistic Models and Data-Driven Stochastic Control. | If the agent does not know the underlying dynamics of the observations (transitions and/or channel), then the learning of the solutions to the optimal control problem from observed data is necessary. However, learning for POMDPs has been a challenging problem. Various approaches and studies are available in the literat... | A |
The overall structure of the proposed Mamba-SEUNet is illustrated in Fig. 1(a). Given the noisy speech input y𝑦yitalic_y, we first apply the Short-Time Fourier Transform (STFT) to obtain its magnitude spectrum Ym∈ℝT×Fsubscript𝑌𝑚superscriptℝ𝑇𝐹Y_{m}\in\mathbb{R}^{T\times F}italic_Y start_POSTSUBSCRIPT italic_m end_P... | In this study, we introduce Mamba-SEUNet, a U-Net style SE network based on TS-Mamba blocks. This architecture leverages bidirectional Mamba blocks to effectively capture both past and future information, addressing the limitations of existing transformer-based methods. By integrating TS-Mamba blocks into the U-Net fra... | We employ time and frequency Mamba blocks sequentially to learn the comprehensive feature representations, as illustrated in Fig. 1(b). To effectively capture both global and local information, each Mamba block incorporates the bidirectional SSM formulation proposed in [27], allowing the model to integrate both past an... | To ensure a fair comparison of Mamba with conformer and transformer, we employ the proposed U-Net architecture, replacing TS-Mamba with TS-Conformer as introduced in [16] and TS-Transformer as described in [32], as shown in Table V. The results indicate that Mamba outperforms conformer, achieving better performance wit... | The forward and backward Mamba blocks share an identical structure. Specifically, for an input sequence xin∈ℝL×Csubscript𝑥𝑖𝑛superscriptℝ𝐿𝐶x_{in}\in\mathbb{R}^{L\times C}italic_x start_POSTSUBSCRIPT italic_i italic_n end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_L × italic_C end_POSTSUPERSCRIPT, we... | B |
∙∙\bullet∙ Numerical simulations show that FOGNA outperforms other existing FODCAs in terms of resolution for DOA estimation. | Specifically, when the difference of incidence angles for two sources is reduced to 1.6∘superscript1.61.6^{\circ}1.6 start_POSTSUPERSCRIPT ∘ end_POSTSUPERSCRIPT, | The angles of 3 uncorrelated sources are −0.8∘superscript0.8-0.8^{\circ}- 0.8 start_POSTSUPERSCRIPT ∘ end_POSTSUPERSCRIPT, 0.8∘superscript0.80.8^{\circ}0.8 start_POSTSUPERSCRIPT ∘ end_POSTSUPERSCRIPT and 6∘superscript66^{\circ}6 start_POSTSUPERSCRIPT ∘ end_POSTSUPERSCRIPT respectively in Fig 5. | It can be seen that FOGNA can distinguish two sources with the difference of incidence angles reduced to 1.6∘superscript1.61.6^{\circ}1.6 start_POSTSUPERSCRIPT ∘ end_POSTSUPERSCRIPT, while others can not. | while the angle of one source fixed at −0.8∘superscript0.8-0.8^{\circ}- 0.8 start_POSTSUPERSCRIPT ∘ end_POSTSUPERSCRIPT, and the angle of other sources vary from 0 to 10. | A |
For Speech Translation (ST), we used both ASR and ST data from CoVoST2 [23] and Europarl-ST [24]. The training data was limited to the following languages: Portuguese (pt), French (fr), Italian (it), German (de), Spanish (es), and English (en). The X→→\rightarrow→en (X ∈\in∈ PFIGS) ST training data comprised approximat... | Despite the successes of SLMs, most of the work has focused on understanding the generalization capabilities of instruction-tuned LLMs in supervised speech processing tasks, where the spoken languages used at test time are seen during the training of the SLM. Therefore, little to no effort has been put into understandi... | Table I presents the speech translation results from both text-only LLMs and our spoken language models (SLMs). For the text-only LLMs, we utilized the original ground-truth transcriptions as input. Firstly, comparing rows 1 and 2, the performance of mT0-XXL significantly outperforms mT0-XL across all language pairs, l... | Prior to assessing the performance of SLMs for speech translation on zero-resource language pairs, we establish an upper bound (i.e., text-only translation) for this work by utilizing the same LLMs employed for training the SLMs. Two different models are evaluated: mT0-XL and mT0-XXL. Furthermore, we include a fine-tun... | To align speech and text representations, we employ a CNN adapter for speech and incorporate LoRA [21] for text LLMs, as outlined in [22]. These components are jointly trained on speech translation and ASR data. To further improve the alignment, prior to the joint training of adapters, we pre-train the CNN adapter on A... | C |
The global and local features of each modality and current text feature are set to 256 dimensions. And, we use two attention heads for GAE, and the fused features are also set to 256 dimensions. Each video clip is processed at 25 FPS with a sampling rate of 16000 Hz. Mel-spectrogram features are extracted with a 40ms w... | (1) Gross Pitch Error (GPE) [27]: percentage of frames where pitch error exceeds 20 % with voicing present in both synthesized and ground-truth speech. | (2) F0 Frame Error (FFE) [28]: percentage of frames with either a voicing decision error or a pitch error exceeding 20%. | (4) Lip Sync Error-Distance (LSE-D) [29, 30]: average error by measuring the distance between lip and audio representations. | (2) MOS-Similarity (MOS-S): Evaluates the similarity of the synthesized speech’s prosody to the ground-truth. | A |
We introduce a novel cross-modality constraint, PABA, which explores fine-grained coherence representations while retaining intrinsic discriminative information. | To validate the effectiveness of SPEG-Net, several state-of-the-art methods are introduced for comparison [5, 6, 23, 24, 12, 25, 26, 27, 28, 13]. Table I reports the optimal performance achieved by the proposed SPEG-Net, which surpasses the second-best method on the RegDB dataset by 3.5% and 4.1% in Rank-1 and by 2.3% ... | TABLE I: The Comparison Results of SEPG-Net and State-of-The-Art methods on RegDB and SYSU-MM01 Datasets. | SEPG-Net, a simple yet effective architecture, is developed for VI-ReID task. First, we observe and analyse the spectral discrepancies between infrared and visible images and propose an end-to-end spectral enhancement strategy which uses frequency domain information and greyscale space to bridge the gap from RGB images... | Experimental results on the SYSU-MM01 [21] and RegDB [20] datasets demonstrate that our method outperforms other state-of-the-art methods. | D |
\cdot a_{ik}+\sin(k\mathbf{x}_{i})\cdot b_{ik}\right)italic_ϕ start_POSTSUBSCRIPT italic_F end_POSTSUBSCRIPT ( bold_x ) = ∑ start_POSTSUBSCRIPT italic_i = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_d end_POSTSUPERSCRIPT ∑ start_POSTSUBSCRIPT italic_k = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_g end_POSTSU... | The Hepatic Vessel dataset is sourced from the Medical Segmentation Decathlon (MSD) challenge, a competition focused on medical image segmentation. Its primary goal is to segment hepatic vessels and tumors from Hepatic CT images. The dataset contains 443 cases of 3D CT data, with each slice image having a resolution of... | Medical image segmentation [1] plays a crucial role in extracting important information from computed tomography (CT) images, providing significant support for disease diagnosis, radiation therapy localization, and surgical planning[2]. Therefore, in recent years, various computational models have been proposed to tack... | Figure 2: The core algorithm framework of SegKAN, illustrates the 3D image slices and encoding of 3D patches, as well as the learning of spatial relationships between 3D patches. The unique elongated structure of Hepatic blood vessels is highlighted in the top-left corner. The bottom-right image shows the impact of dif... | The experimental results on the Hepatic Vessel dataset indicate that, due to the large volume of the dataset, models based on self-attention mechanisms significantly outperform pure convolutional models, such as nnUNet, especially in the recent nnFormer and TransUNet models. Thanks to the introduction of the self-atten... | A |
The QAE achieves a lower steady-state BLER after 20 epochs, and this trend remains consistent throughout the remaining range, demonstrating superior BLER convergence. The result further reveals that the parallel quantum circuits at the transmitter can learn a more effective mapping for encoding the one-hot vector to th... | The QAE achieves a lower steady-state BLER after 20 epochs, and this trend remains consistent throughout the remaining range, demonstrating superior BLER convergence. The result further reveals that the parallel quantum circuits at the transmitter can learn a more effective mapping for encoding the one-hot vector to th... | The primary advantage of QML over the classical DL technique is that it requires significantly fewer trainable parameters, resulting in memory efficiency in training and the deployment on quantum computers. The parameter analysis for AE is presented in Table I. At the transmitter, the classical AE consists of two FC la... | Recently, quantum machine learning (QML) has gained significant attention in the field of wireless communications. The principles of quantum superposition and quantum entanglement offer a fundamental rethinking of traditional approaches, providing new insights to enhance the mainstream DNN-based architecture. Due to th... | In this paper, we introduce a hybrid quantum-classical autoencoder framework that exhibits advantages in both BLER performance and parameter savings compared to state-of-the-art solutions. Leveraging the quantum superposition principle, the quantum circuit encodes the large dimensional one-hot representation to a small... | D |
We present measurement setup and details of the measurement campaign, which analyzes over 25,0002500025,00025 , 000 directional power delay profile (PDP) to reveal key channel characteristics, where distances between transmitter (Tx) and receiver (Rx) range between 60606060 and 400 mtimes400meter400\text{\,}\mathrm{... | In this section, we discuss the results of our measurement campaign, focusing on key observations and analyses derived from the data. First, the PDPs for selected LoS and OLoS cases are examined to highlight the delay dispersion characteristics of the measured channels. Next, the angular power spectra are analyzed to u... | AS provides insights into the spatial characteristics of the channel at both the Tx and Rx. The CDFs of AS are presented in Fig. 11(a) for the Tx and Fig. 12(a) for the Rx, while the corresponding modeling as a function of log10(d)subscript10𝑑\log_{10}(d)roman_log start_POSTSUBSCRIPT 10 end_POSTSUBSCRIPT ( italic_d )... | We analyze sample results both to verify the correctness of the measurements and point out key propagation effects; we particularly note the importance of vegetation attenuation on the link characteristics, and argue that obstructed line-of-sight (OLoS) must be treated as separate category from LoS and non line-of-si... | The PL results reveal several key trends. First, path loss increases with frequency, as expected, due to the frequency-dependent nature of different types of attenuation. Similarly, path loss increases with distance, reflecting the fundamental behavior of signal attenuation over larger propagation ranges. However, the ... | C |
In the final simulation, we observe the standard deviation (SD) of loss versus epoch to evaluate the convergence speed of different algorithms. It is worth mentioning that we take the convergence value as the expectation to calculate the standard deviation. As shown in Figure 5(c), the SD of MoD-DNN converges rapidly c... | In order to further compare the estimation accuracy, we observe the RMSE of different algorithms versus SNR. As shown in Figure 5(a), the performance of MUSIC and DeepMUSIC almost remain unchanged as SNR varies. This phenomenon indicates that these algorithm are statistically invalid with the existence of hardware impa... | In this section, we demonstrate the effectiveness of AoA estimation with existence of hardware impairments using the proposed MoD-DNN, especially in comparison with | In this subsection, we compare the AoA estimation performance of different methods in an anechoic chamber. Experiments conducted in an anechoic chamber provide valuable insights as NLoS and multipath propagation are absent. In this controlled environment, hardware impairments are the primary challenges to the accuracy ... | As shown in the overall iteration illustrated in Figure 2(a), the AoA resolution of the CSS is relatively low, and the position of the peak is inaccurate. To enhance the AoA estimation performance, the reconstruction of the spatial spectrum is expressed as the following optimization problem | C |
Extensive experiments demonstrate that SECodec outperforms EnCodec in speech reconstruction. Furthermore, we developed a Structural Entropy-based Speech Language Model (SESLM) that leverages SECodec, yielding superior results in terms of generated speech content accuracy and quality. Additionally, the experiments show ... | This work was supported in part by the National Natural Science Foundation of China (Nos. U24A20334, 62376111, U23A2038 and 61972186), Yunnan provincial major scienceand technology special plan projects (Nos. 202302AD080003, 202402AG050007), Yunnan Provincial Key Researchand provincial major scienceand technology speci... | We validate the effectiveness of the structural entropy-based speech language model on the zero-shot TTS task. During inference, text input is converted to a phoneme sequence and the speech prompt to speech tokens. These are concatenated to form the prompts for both the AR and NAR models. The tokens generated by the AR... | The ideal speech representation for speech language models should meet two key characteristics: i) Effective preservation of speech information; ii) Sufficient compressiveness for efficient training of speech language models. | To demonstrate the compressive and informative nature of the codebook learned by SECodec, we first visualized each column vector in the codebooks initialized by different methods using t-SNE. As illustrated in Figure 2(a) and Figure 2(b), the codebook initialized with k-means exhibits an uneven distribution in space, w... | A |
The sounds produced by these touch movements are recorded by a microphone (the black device in Fig. 2) to form the sound dataset in this paper. | Table I shows the results of MTRCNN for classifying 6 gestures, as well as arousal and valence of emotions. Arousal is usually classified as low, neutral, and high. Valence is classified as negative, neutral, and positive. The AVDM of arousal-valence joint classification has four quadrants and an origin, so the arousal... | In Fig. 5 (a), MTRCNN performs better in identifying high arousal than low and neutral states. This implies that high arousal associated with touch sounds is easier to distinguish. Fig. 5 (c) implies that touch-based emotions in positive valence are more distinguishable than those in negative valence. In the Aro-Val sp... | According to the Circumplex Model [23], the distribution of the 10 emotions involved in arousal-valence dimensions is shown in Fig. 3. | Emotions are typically analysed along two independent dimensions (arousal and valence), like in Russell’s model [22][23], and neuroimaging studies [24][25] support these two-dimensional representations. Hence, the arousal-valence dimensional model (AVDM) is more commonly used for sound-related affective computing, e.g.... | C |
Re-mapping text tokens as audio tokens. Next, we replace some of the text tokens in the LLM vocabulary with audio tokens. We do that by mapping the 1024 audio cluster ids to the last 1024 vocabulary indices corresponding to the least frequent text-tokens. This design choice is motivated from the fact that the last few ... | Since our goal is to not only train the model to recognize speech but also adapt it to more diverse disordered speech, we compare fine-tuning of the LLM to recognize speech on different mixtures of standard speech datasets and disordered speech data. With the tokenized audio data we use standard fine tuning recipes to ... | vocabulary of the LLM prior to tuning it, we simply repurpose low frequency text tokens in the vocabulary for audio and tune the LLM to enable the model to recognize speech. | Prior to training, we tokenize the audio datasets. We learn the clusters for the audio tokens based only on the LibriSpeech dataset. With the tokenized audio, we perform supervised finetuning of the Gemma model with a learning rate of 5×10−65superscript1065\times 10^{-6}5 × 10 start_POSTSUPERSCRIPT - 6 end_POSTSUPERSCR... | Our eventual goal is to adapt the LLM based ASR model to recognize disordered speech. This can be done by tuning the model further on the disordered speech data. In their seminal work [27] demonstarted that RL can be used to finetune LLM outputs on a relatively small number of examples to maximize metrics learnt from h... | A |
The DDD-GenDT framework introduces the LLMs as a digital twin, the powerful function of LLM makes it more flexible and simple to build a digital twin system with DDDAS concept. DDD-GenDT does not require retraining the model, and only needs to adjust the LLMs input to generate digital twin prediction output. | Building upon this conceptual model, we propose a new dynamic data-driven generative digital twin framework, DDD-GenDT, to enable coupling and interaction between the dynamic digital twin and physical systems. The key idea of this framework is to take advantage of the broad problem solving capabilities of large languag... | Many researchers have recently proposed LLM as the bridge between reality and virtuality, and its impact spans from laboratories to industry. Coscientist, an artificial intelligence system for automated chemical experiments proposed by Boiko et al. [18], which can semi-automatically plan the overall experiment and oper... | In this study, we proposed the DDD-GenDT framework using LLM as a dynamic digital twin to capture and map the physical state behavior of complex systems and proposed the CNC machining power consumption curve as an application case to make LLM Dynamic DT closely related to the physical system interact and take control. ... | We used GPT-4 and GPT 3.5 Turbo as LLM-based dynamic DTs in the DDD-GenDT framework in experiments. For the GPT 3.5 Turbo and GPT-4 models, we use the LLMTime method proposed by Gruver et al. [12] to generate time series data corresponding to the physical state. Our scenario is to apply a digital twin to map the CNC cu... | B |
The integration of MC dropout with Inception-v3 follows the design adopted by Dolezal et al. [14]. At the inference stage, uncertainty is estimated by generating an ensemble of predictions, where each image tile is passed through the dropout-enabled Inception-v3 five times. | In the scenario of using foundation models for NSCLC subtyping, which involves training a separate model using latent tile representations generated by foundation models, the TCGA and CPTAC datasets were utilized independently for the entire model training and evaluation process. Each dataset was partitioned into three... | Image preprocessing. Regardless of distinct model input requirements, all WSIs were segmented and split into appropriate tiles using a standard preprocessing pipeline implemented in Slideflow (version 2.1.0) [54]. Directly feeding large gigapixel WSIs into a neural network is computationally impractical, so image tiles... | We trained models from scratch using a 24GB NVIDIA RTX 4090 GPU with a batch size of 64, where image tiles are resized to 299×299299299299\times 299299 × 299 pixels. We adopted a learning rate scheduler with an initial learning rate of 0.0003 and a decay rate of 0.98 every 512 steps. Each model is trained for four epoc... | To adapt tile-level foundation models for NSCLC subtyping, we utilized the image encoders of UNI and CONCH to extract tile embedding (i.e., latent representations), concatenate the resulting representations for each slide, and apply the widely used attention-based multiple instance learning (ABMIL) for weakly supervise... | C |
Table I summarizes the performance of the AST models trained using various Mixup methods across three source domains: ICBHI, SPR, and HF. The performance is also evaluated on the combined test data, denoted as COMB, to simulate real-world distribution. The top half of the table presents previous methods, while the bott... | The bolded results represent the best performance, and the underlined results represent the second-best performance in the target domains. | The results demonstrate that Lungmix is the most effective method, significantly enhancing performance across the different source and test domains. Even though results are slightly inverted when using SPRSound as the source domain, Lungmix still achieves the best overall performance. When Lungmix is applied without th... | Figure 2: Visualization of Lungmix. An a crackle and wheeze are mixed into both. The grey parts denote the random mask, and the white parts denote the loudness mask. The zoomed-in section highlights the short and discontinuous crackle sound. The part under the zoomed-in is randomly generated padding. | Table I summarizes the performance of the AST models trained using various Mixup methods across three source domains: ICBHI, SPR, and HF. The performance is also evaluated on the combined test data, denoted as COMB, to simulate real-world distribution. The top half of the table presents previous methods, while the bott... | B |
README.md exists but content is empty.
- Downloads last month
- 4