Amirhossein Sayyafan

Bio: Amirhossein Sayyafan is an academic researcher from Washington State University. The author has contributed to research in topics: Equalization (audio) & Noise. The author has an hindex of 3, co-authored 7 publications receiving 30 citations. Previous affiliations of Amirhossein Sayyafan include Sharif University of Technology.

Deep Neural Network Based Media Noise Predictors for Use in High-Density Magnetic Recording Turbo-Detectors

Abstract: This article presents a combined Bahl–Cocke–Jelinek–Raviv (BCJR) and deep neural network (DNN) turbo-detection architecture for 1-D hard disk drive (HDD) magnetic recording. Simulated HDD readings based on a grain flipping probabilistic (GFP) model are input to a linear filter equalizer with a 1-D partial response (PR) target. The equalizer output is provided to the BCJR detector in order to minimize the intersymbol interference (ISI) due to the PR mask. The BCJR detector’s log-likelihood-ratio (LLR) outputs (along with the linear equalizer outputs) are then input to the DNN detector, which estimates the signal-dependent media noise. The media noise estimate is then fed back to the BCJR detector in an iterative manner. Several DNN media noise estimation architectures based on fully connected (FC) and convolutional neural networks (CNNs) are investigated. For GFP data at 48 nm track pitch and 11 nm bit length, the CNN-based BCJR-DNN turbo detector reduces the detector bit error rate (BER) by $0.334\times $ and the per bit computational time by $0.731\times $ compared to a BCJR detector that incorporates 1-D pattern-dependent noise prediction (PDNP). The proposed BCJR-DNN turbo detection architecture can be generalized for two-dimensional magnetic recording (TDMR).

Deep Neural Network-Based Detection and Partial Response Equalization for Multilayer Magnetic Recording

Abstract: To increase the storage capacity limit of magnetic recording channels, recent studies proposed multilayer magnetic recording (MLMR): the vertical stacking of magnetic media layers. MLMR readback waveforms consist of the superposition of signals from each layer recovered by a read head placed above the upper layer. This article considers the problem of equalization and detection for MLMR comprising two layers. To this end, we use MLMR waveforms generated using a grain switching probability (GSP) model that is trained on realistic micromagnetic simulations. We propose three systems for equalization and detection. The first is a convolutional neural network (CNN) equalizer followed by an MLMR Viterbi algorithm (VA) for detection. We show that this system outperforms the traditional 2-D linear minimum mean squared error (2-D-LMMSE) equalizer. The second system uses CNNs for equalization and separation of signals from each layer, which is followed by a regular VA. The third system contains CNNs trained to directly provide soft bit estimates. By interfacing the CNN detector with a channel decoder, we show that an areal density gain of 16.2% can be achieved by a two-layer MLMR system over a one-layer system.

Pixel-level alignment of facial images for high accuracy recognition using ensemble of patches.

Abstract: The variation of pose, illumination, and expression continues to make face recognition a challenging problem As a pre-processing step in holistic approaches, faces are usually aligned by eyes The proposed method tries to perform a pixel alignment rather than eye alignment by mapping the geometry of faces to a reference face while keeping their own textures The proposed geometry alignment not only creates a meaningful correspondence among every pixel of all faces, but also removes expression and pose variations effectively The geometry alignment is performed pixel-wise, ie, every pixel of the face is corresponded to a pixel of the reference face In the proposed method, the information of intensity and geometry of faces is separated properly, trained by separate classifiers, and finally fused together to recognize human faces Experimental results show a great improvement using the proposed method in comparison to eye-aligned recognition For instance, at the false acceptance rate (FAR) of 0001, the recognition rates are respectively improved by 24% and 33% in the Yale and AT&T datasets In the labeled faces in the wild dataset, which is a challenging, big dataset, improvement is 20% at a FAR of 01

Deep Neural Network Media Noise Predictor Turbo-Detection System for 1-D and 2-D High-Density Magnetic Recording

Abstract: This article presents a concatenated Bahl–Cocke–Jelinek–Raviv (BCJR) detector, low-density parity-check (LDPC) decoder, and deep neural network (DNN) architecture for a turbo-detection system for 1-D and 2-D magnetic recording (1DMR and TDMR). The input readings first are fed to a partial response (PR) equalizer. Two types of the equalizer are investigated: a linear filter equalizer with a 1-D/2-D PR target and a convolutional neural network (CNN) PR equalizer that is proposed in this work. The equalized inputs are passed to the BCJR to generate the log-likelihood-ratio (LLR) outputs. We input the BCJR LLRs to a CNN noise predictor to predict the signal-dependent media noise. Two different CNN interfaces with the channel decoder are evaluated for TDMR. Then, the second pass of the BCJR is provided with the estimated media noise, and it feeds its output to the LDPC decoder. The system exchanges LLRs between BCJR, LDPC, and CNN iteratively to achieve higher areal density. The simulation results are performed on a grain flipping probabilistic (GFP) model with 11.4 Teragrains per square inch (Tg/in2). For the GFP data with 18 nm track pitch (TP) and 11 nm bit length (BL), the proposed method for TDMR achieves 27.78% areal density gain over the 1-D pattern-dependent noise prediction (PDNP). The presented BCJR-LDPC-CNN turbo-detection system obtains 3.877 Terabits per square inch (T/bin2) areal density for 11.4 Tg/in2 GFP model data, which is among the highest areal densities reported to date.

A perspective on deep neural network-based detection for multilayer magnetic recording

Abstract: This paper describes challenges, solutions, and prospects for data recovery in multilayer magnetic recording (MLMR)—the vertical stacking of magnetic media layers to increase information storage density. To this end, the channel model for MLMR is discussed. Data recovery is described in terms of the readback stage followed by equalization and then detection. We illustrate how deep neural networks (DNNs) can be used to design systems for equalization and detection for MLMR. We show that such DNN-based systems outperform the conventional baseline and provide a good trade-off between complexity and performance. To achieve additional density gains, several prospective methods are discussed. On a physical level, the selective reading of tracks on different layers can be achieved by resonant reading. Resonant reading promises reduced interference from different layers, enabling higher storage densities. Regarding the signal processing, DNNs can be used to estimate the media noise and iteratively exchange soft-bit information with the decoder. Also, to ameliorate partial erasures, an auto-encoder-based system is proposed as a modulation coding scheme.

Equalization For Maximum Likelihood Detectors

Abstract: In this paper, the minimum mean-square error (MMSE) technique has been used to equalize the recording channel in order to facilitate the application of the Viterbi detector. The resulting performance has been compared with that of the optimal equalization system which yields the minimum probability of error at the output of the Viterbi detector. The results indicate that depending on the constraint used in the MMSE design, the amount of noise correlation varies significantly at the equalizer output, which in turn makes a large difference in the performance of the Viterbi detector. In particular, in the jitter-dominant channel where unconditioned channel noise samples are highly correlated, the monic constraint on the equalizer target response tends to whiten the noise samples at the equalizer output. This results in a significant performance improvement of the monic constraint upon the fixed-energy constraint as well as the popular partial response targets of the form (1-D)(1+D)/sup n/. >

Feature Selection and Feature Extraction in Pattern Analysis: A Literature Review.

Abstract: Pattern analysis often requires a pre-processing stage for extracting or selecting features in order to help the classification, prediction, or clustering stage discriminate or represent the data in a better way. The reason for this requirement is that the raw data are complex and difficult to process without extracting or selecting appropriate features beforehand. This paper reviews theory and motivation of different common methods of feature selection and extraction and introduces some of their applications. Some numerical implementations are also shown for these methods. Finally, the methods in feature selection and extraction are compared.

Principles of digital communication and coding

Abstract: two listed are 1) the reference model is a nonlinear, time-varying system, and 2) the adjustable system also contains nonlinearities. These situations may lead to violations of the basic hypothesis in the parameter identification and adaptive observer problems, but not in the adaptive control problem. Another consequence is that here seems to be a serious omission of one assumption required in the control problem but not in the other two. That is that Ahe sign of the leading coefficient in the plant’s numerator polynomial b , in equation (6.4-35), must be known. Solutions which do not require this assumption have been discussed by various people, but are not very satisfactory because the algorithms invariably involve pogntial division by zero. The requirement for knowledge of the sign of bm in the control case. is obscured in the general discussion of Section 4.2 of Chapter Iv. Chapters 7 and 8 give a good development of the model reference adaptive approach to the parameter identification problem and the simultaneous adaptive state observation and parameter identification problem, respectively. Chapter 7 contains the development for both continuous time and discrete time identification algorithms, and considers the importance of variable (time decreasing) adaption g a i n s in the presence of measurement noise. The advantages and disadvantages of parallel identifiers (output error method) and series-parallel identifiers (equation error method) are discussed. Two case studies are included, one using the continuous time algorithm and one the discrete time algorithm. Chapter 8 deals only with discrete time algorithms in a deterministic context. The last statement in Chapter 7 is, “ M U S identifiers can be used for implementing adaptive control schemes.” Some discussion concerning the potential stability problems associated with this approach would have been helpful. The author should be commended for even attempting a book in this extremely active area of research where many concepts have not yet completely solidified. He has achieved to a considerable degree his stated goals of making the book interesting to practicing engineers, useful as a textbook for graduate students, and a starting point for further investigation by researchers.

Deep Neural Network a Posteriori Probability Detector for Two-Dimensional Magnetic Recording

Abstract: In two-dimensional magnetic recording (TDMR) channels, intersymbol interference (within and between tracks) and pattern-dependent media noise are impediments to reaching higher areal density. We propose a novel deep neural network (DNN)-based a posteriori probability (APP) detection system with parallel multi-track detection for TDMR channels. The proposed DNN-based APP detector replaces the trellis-based Bahl–Cocke–Jelinek–Raviv (BCJR) or Viterbi algorithm and pattern-dependent noise prediction (PDNP) in a typical TDMR scenario, in which it directly outputs log-likelihood ratios of the coded bits and iteratively exchanges them with a subsequent channel decoder to minimize bit error rate (BER). We investigate three DNN architectures—fully connected DNN, convolutional neural network (CNN), and long short-term memory (LSTM) network. The DNN’s complexity is limited by employing linear partial response (PR) equalizer pre-processing. The best performing DNN architecture, CNN, is selected for iterative decoding with a channel decoder. Simulation results on a grain-flipping-probability (GFP) media model show that all three DNN architectures yield significant BER reductions over a recently proposed 2D-PDNP system and a previously proposed local area influence probabilistic (LAIP)-BCJR system. On a GFP model with 18 nm track pitch and 11.4 Teragrains/in2, the CNN detection system achieves an information areal density of 3.08 Terabits/in2, i.e., a 21.72% density gain over a standard BCJR-based 1D-PDNP; the CNN-based system also has $3\times $ the throughput of 1D-PDNP, yet requires only 1/10th the computer run time.