A5039317227- Neural Networks and Applications
- Advanced Memory and Neural Computing
- Adversarial Robustness in Machine Learning
- Speech Recognition and Synthesis
- Control Systems and Identification
- Network Security and Intrusion Detection
- Digital Media Forensic Detection
- Analog and Mixed-Signal Circuit Design
- Advanced Data Compression Techniques
- Music and Audio Processing
- Computer Graphics and Visualization Techniques
- Advanced Malware Detection Techniques
- Cancer Genomics and Diagnostics
- Blind Source Separation Techniques
- Oral and Maxillofacial Pathology
- Neural Networks and Reservoir Computing
- 3D Shape Modeling and Analysis
- Heme Oxygenase-1 and Carbon Monoxide
- Cancer Diagnosis and Treatment
- Algorithms and Data Compression
- Speech and Audio Processing
- Composite Material Mechanics
- Cloud Data Security Solutions
- Spam and Phishing Detection
- Anomaly Detection Techniques and Applications
Southeast University
2024
Michigan State University
1998-2024
University of Florida
2022
Zoucheng People's Hospital
2016
PLA Air Force Aviation University
2015
University of Pittsburgh
1982-2014
Johnson & Johnson (United States)
2007-2009
Tennessee Oncology
2007
Sarah Cannon
2007
University of Rochester
2007
A formulation for dynamic feedback neural networks of the Hopfield type is presented. description given general design framework used, in which a network would only have finite number memories. Some basic properties nonzero equilibria as well (unstable) equilibrium point at zero proposed are also discussed.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">></ETX>
Inaudible voice command injection is one of the most threatening attacks towards assistants. Existing aim at injecting attack signals over air, but they require access to authorized user's for activating Moreover, effectiveness can be greatly deteriorated in a noisy environment. In this paper, we explore new type channel, power line side-channel, launch inaudible injection. By audio through modified charging cable, becomes more resilient against various environmental factors and liveness...
Recent advances in natural language processing and machine learning have led to the development of chatbot models, such as ChatGPT, that can engage conversational dialogue with human users. However, understanding ability these models generate toxic or harmful responses during a non-toxic multi-turn conversation remains an open research problem. Existing focuses on single-turn sentence testing, while we find 82% individual sentences elicit behaviors are considered safe by existing tools. In...
Abstract We present a linear algorithm to reconstruct the vertex coordinates for surface mesh given its edge lengths and dihedral angles, unique up rotation translation. A local integrability condition existence of an immersion in 3D Euclidean space is provided, mirroring fundamental theorem surfaces continuous setting (i.e. Gauss's equation Mainardi–Codazzi equations) if we regard as discrete first form angles second form. The resulting sparse system solve derived from convex optimization...
Deep learning based voice synthesis technology generates artificial human-like speeches, which has been used in deepfakes or identity theft attacks. Existing defense mechanisms inject subtle adversarial perturbations into the raw speech audios to mislead models. However, optimizing perturbation not only consumes substantial computation time, but it also requires availability of entire speech. Therefore, they are suitable for protecting live streams, such as messages online meetings. In this...
Estimating the cardinality of a data stream is fundamental problem underlying numerous applications such as traffic monitoring in network or datacenter and query optimization Internet-scale P2P networks. Existing solutions suffer from high processing/query overhead memory in-efficiency, which prevents them operating online for streams with very arrival rates. This paper takes new solution path different prior art proposes self-morphing bitmap, combines operational simplicity structural...
In this paper, we propose SpecPatch, a human-in-the loop adversarial audio attack on automated speech recognition (ASR) systems. Existing attacker assumes that the users cannot notice audios, and hence allows successful delivery of crafted examples or perturbations. However, in practical scenario, intelligent voice-controlled systems (e.g., smartwatches, smart speakers, smartphones) have constant vigilance for suspicious voice, especially when they are delivering their voice commands. Once...
In this paper, we propose PhantomSound, a query-efficient black-box attack toward voice assistants. Existing adversarial attacks on assistants either apply substitution models or leverage the intermediate model output to estimate gradients for crafting audio samples. However, these approaches require significant amount of queries with lengthy training stage. PhantomSound leverages decision-based produce effective audios, and reduces number by optimizing gradient estimation. experiments,...
Abstract Background: A significant proportion of ER-positive tumors exhibit resistance to endocrine therapy. Clinical studies in LNN/ER+ patients show that the absolute distant disease-free survival benefit tamoxifen therapy is only about 9%, with a relative 30%. These benefits from treatment may, however, differ substantially according e.g. patient's age and tumor level ER it therefore importance identify biomarkers allow identification those breast cancer who may actually adjuvant...
Abstract Based on a new spectral vector field analysis triangle meshes, we construct compact representation for near conformal mesh surface correspondences. Generalizing the functional map representation, our uses between low‐frequency tangent fields induced by correspondence. While is as efficient, it also capable of handling more generic class correspondence inference. We formulate preservation constraints and regularization terms inference, with function treated special case. A number...
The authors present models of feedback neural nets which are described by nonlinear differential equations. They show that their earlier proofs for convergence to bounded regions and the existence a finite number equilibria independent symmetry interconnection matrix thus also applicable more general nongradient dynamic nets. However, when interconnect is asymmetric, network not guaranteed have only (finite) as its limit set. computer simulations three-neuron demonstrate coexistence two...
Neural circuits that have the same dynamic properties as gradient continuous-time feedback neural nets are presented. These governed by a set of equations, and in addition they implementable via all MOS elements. Consequently, these suitable for analog VLSI implementation.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">></ETX>
Custom analog CMOS VLSI components have been designed to allow the construction of neural networks with arbitrary architectures. A wide-range, all-enhancement-mode, MOS four-quadrant multiplier has employed implement scalar vector product neuron outputs and corresponding weights. nonlinear MOSFET floating element is used model conductive elements; this extends modeling synaptic weights elements. realized by a simple operational amplifier which compatible I/O multiplier. The system can be...
The authors present initial results of a pattern/character recognition and association experiment using newly fabricated 50-neuron CMOS analog silicon chip with digital on-chip learning. Attention is given to the circuit architecture, VLSI chips, interface circuitry.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">></ETX>
An all-MOS realization of the linear synaptic weight for neural nets is described. The achieved via an adaptation continuous-time analog multipliers where weights are assigned as positive or negative voltage levels. Using only a single newly designed CMOS operational amplifiers, each multiplier capable realizing scalar product Sigma W/sub ij/X/sub j/, j=1, . ., n, and i fixed, X/sub j/ external input output neuron j ij/ externally assignable weight. artificial network would then be realized...
In this paper, we propose WavePurifier, an audio purification framework to defend against adversarial attacks. Audio attacks craft examples or perturbations attack the automated speech recognition (ASR) models. Although existing defense mechanisms can detect such and raise alarms, they fail recover maintain benign commands. Consequently, leads denial of users' Different than defenses, WavePurifier aims purify examples, thereby rectifying user's We find that forward diffusion process model...
Two prototype chips of an all-MOS implementation a new architecture have been designed and then fabricated tested. Real-time experiments on the were conducted to demonstrate use neural circuit as pattern/character recognizer. Both software dynamic learning schemes on-chip digital hardware circuitry successfully stored desired patterns in network. The execution time real-time was microseconds range with low power dissipation 1-mW range. Dedicated interface circuitries environments built chip...
21109 Background: Standard treatment for most patients with CUP involves empiric chemotherapy. Since specific now exists types of advanced carcinoma, precise identification the primary site could lead to improved therapy. Veridex developed an optimized set 10 gene markers, a qRTPCR assay identify tissue origin metastatic carcinoma in formalin-fixed, paraffin-embedded (FFPE) samples (J Mol Diagn 8:320, 2006). The includes markers 6 sites: lung, pancreas, colon, breast, ovary, and prostate. In...
This paper presents speech and a noise separation methods for achieving high quality of coding. In separation, pitch harmonics restoration method is proposed. can effectively suppress the so-called musical reduce distortion. noise-base estimated in process used as separated background encoding with low bit rates The proposed preprocessor adaptive multi-rate wideband (AMR-WB) coder are evaluated by degradation category rating (DCR) test. An average 0.3-point improvement performance under...
Neural circuits that have qualitatively the same dynamic properties as continuous-time feedforward/feedback artificial neural nets are introduced. These realizable by all-MOS elements, and thus they implementable analog MOS VLSI technology. The provide an efficient inexpensive electronics for various tasks in identification control.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">></ETX>