Gamze Erten

Real time separation of audio signals using digital signal processors

This work presents a practical real time execution of a family of dynamic blind signal separation algorithms using commercial digital signal processors (DSP). The implementation adopts discrete-time formulations with a view towards on line commercial use. Many applications exist in voice based computer interfaces and telephony.

Exact entropy series representation for blind source separation

An explicit infinite series for the marginal entropy of a probability density function is developed. The series includes all orders of statistics and employs both the Gram-Charlier and the Edgeworth expansions for its derivation. The derivation exploits the fact that the two expansions are equivalent for the same probability density. The developed entropy series expression can be used to express the averaged mutual information to any degree of accuracy. This measure is then used in the derivation of the update laws of the blind separation of sources.

Using neural networks to control the process of plasma etching and deposition

Neural architectures are proposed to model and control plasma etching and deposition processes in semiconductor wafer manufacturing. Static and dynamic neural networks are used to develop plant models and inverse models. A single-hidden layer feedforward neural network model learns to identify the systems input-output relationship. Another single-hidden layer feedforward neural controller learns to model the inverse relationship of the plant. The trained controller, in series with appropriate filters, is then used to control the plasma machine in etching and deposition processes. The paper demonstrates how neural networks can learn both the modeling and control tasks in this nonlinear and complex process.

Modeling of a plasma processing machine for semiconductor wafer etching using energy-functions-based neural networks

The complex processing of plasma etching and deposition is highly nonlinear and its modeling is intractable by analytical basic-principles techniques. Neural network approaches have shown initial success for specific plasma processes in extracting implicit relations/models based on input-output measurements. The resulting modeling techniques naturally depend on the neural structure, the adopted learning algorithms, and the specific plasma process and machine. We describe a plasma processing machine designed and in operation at Michigan State University, East Lansing, which has been equipped with select sensing devices. The machine exhibits a hysteresic nonlinearity in the desirable processing modes of operation. The experimental data characterize a testbed plasma etching process using Argon gas with control inputs including incident microwave power, pressure, and cavity size. The internal states and the outputs include reflected power, electric field, and ion density. We employ several tailored networks with novel learning algorithms derived from functions that include the polynomial and the exponential energy functions. It is shown that the learning algorithms enable fast and satisfactory convergence of parameters (weights and biases) in several scenarios of modeling and generalizing the input-state-output relations of the plasma process.

Formulation and algorithms for blind signal recovery

We review some recent approximations of the averaged mutual information criterion and its use as a measure of signal independence. We describe an update law and its comparison with previous work in the literature. We also identify the link between the minimization of the mutual information and the information-maximization of the output entropy function of a (nonlinear) neural network. Example simulations demonstrate the performance of our recently developed algorithm in static and dynamic environments.

Sensor fusion by principal and independent component decomposition using neural networks

The paper describes a view to use both principal component analysis (PCA) and independent component analysis (ICA) within the context of sensor fusion. A nonlinear version of PCA would be appropriate for representing signals/data which span a submanifold structure in its coordinate space. The nonlinear PCA is a candidate for data reduction/compression where multisensors are measurement the same type of signal, e.g., image or sound. In contrast the ICA is a candidate for fusing different types of signals, e.g., image, sound, acceleration, etc., to generate independent components. The PCA approach can be used to transfer compressed data then reconstruct the information bearing signal for use. While the ICA may be used to infer the condition/state of the environment, e.g., office building, airport, etc. Thus the two approaches can be integrated to form a complementary sensory fusion system.

Voice signal extraction for enhanced speech quality in noisy vehicle environments

The voice extraction (VE) system described in this paper is an enabling technology that provides pure voice signals to speech processing systems in noisy vehicle environments. The technology is used to extract a single voice signal of interest from a mixture of sounds, including background noise, music, and multiple speakers. The technology embodies both hardware and software elements. Tests conducted inside a noisy vehicle are presented.

Low power analog chips for the computation of the maximal principal component

Test results of two prototype circuit implementations that compute the maximal principal component are described. The implementations are designed to be compact and operate in the subthreshold regime for low power consumption. The prototypes use direct realization of a nonlinear self-learning circuit models which we have developed.<<ETX>>

Nonlinear projection to submanifolds using neural networks with circuit realization and its application to data reduction

The Principal Component Analysis (PCA) approach and its variations compute eigenvalues and eigenvectors and hence planar surfaces. An extension of the PCA approach, which computes nonplanar (folded) surfaces, is a computational tool that computes submanifolds in higher dimensional spaces. The focus of this work is an application of this nonlinear projection approach to nonlinear data reduction/compression of color images and 3-band radar signals. Results of various image processing from both computer simulations of the circuit model and chip experiments are reported. Furthermore, we describe how certain neural networks can be constructed as computational tools for general submanifold charts and parameterization maps of nonlinear dynamical systems. Moreover, the architectural framework is compatible with basic circuit realization using micropower microelectronic components.

Hybrid analog/digital VLSI chip for template matching

A template matching algorithm, its hybrid analog/digital VLSI implementation, and test results from prototype chips are presented. The inputs to the system are analog values of the pixels of both the template and the region to be matched. The output is a digital code identifying the best match. This hybrid implementation offers two advantages: First, the low-level computationally-intensive template matching task is carried out in analog VLSI with circuits operating in the subthreshold region of transistor operation. The chip consequently consumes very little power. Second, the coded output is digital and hence provides direct interface capability to conventional digital computers. A complete system based on this chip would find numerous applications in edge detection, stereo and motion correspondence, as well as in pattern recognition.