Note: This is a virtual presentation. Here is the link for where the presentation will be taking place.
Title: Circuits and Architecture for Bio-Inspired AI Accelerators
Abstract: Technological advances in microelectronics envisioned through Moore’s law have led to more powerful processors that can handle complex and computationally intensive tasks. Nonetheless, these advancements through technology scaling have come at an unfavorable cost of significantly larger power consumption, which has posed challenges for data processing centers and computers at the scale. Moreover, with the emergence of mobile computing platforms constrained by power and bandwidth for distributed computing, the necessity for more energy-efficient scalable local processing has become more significant.
Unconventional Compute-in-Memory (CiM) architectures such as the analog winner-takes-all associative-memory, the Charge-Injection Device (CID) processor, and analog-array processing have been proposed as alternatives. Unconventional charge-based computation has been employed for neural network accelerators in the past, where impressive energy efficiency per operation has been attained in 1-bit vector-vector multiplications (VMMs), and in recent work, multi-bit vector-vector multiplications. A similar approach was used in earlier work, where a charge-injection device array was utilized to store binary coded vectors, and computations were done using binary or multi-bit inputs in the charge domain; computation is carried out by counting quanta of charge at the thermal noise limit, using packets of about 1000 electrons. These systems are neither analog nor digital in the traditional sense but employ mixed-signal circuits to count the packets of charge and hence we call them Quasi-Digital. By amortizing the energy costs of the mixed-signal encoding/decoding over compute-vectors with a large number of elements, high energy efficiencies can be achieved.
In this dissertation, I present a design framework for AI accelerators using scalable compute-in-memory architectures. On the device level, two primitive elements are designed and characterized as target storage technologies: (i) a multilevel non-volatile computational cell and (ii) a pseudo Dynamic Random-Access Memory (pseudo-DRAM) computational bit-cell. Experimental results in deep-submicron CMOS processes demonstrate successful operation; subsequently, behavioral models were developed and employed in large-scale system simulations and emulations. Thereafter, at the level of circuit description, compute-in-memory crossbars and mixed-signal circuits were designed, allowing seamless connectivity to digital controllers. At the level of data representation, both binary and stochastic-unary coding are used to compute Vector-Vector Multiplications (VMMs) at the array level, demonstrating successful experimental results and providing insight into the integration requirements that larger systems may demand. Finally, on the architectural level, two AI accelerator architectures for data center processing and edge computing are discussed. Both designs are scalable multi-core Systems-on-Chip (SoCs), where vector-processor arrays are tiled on a 2-layer Network-on-Chip (NoC), enabling neighbor communication and flexible compute vs. memory trade-off. General purpose Arm/RISCV co-processors provide adequate bootstrapping and system-housekeeping and a high-speed interface fabric facilitates Input/Output to main memory.
Andreas Andreou, Department of Electrical and Computer Engineering
Ralph Etienne-Cummings, Department of Electrical and Computer Engineering
Philippe Pouliquen, Department of Electrical and Computer Engineering
Note: This is a virtual presentation. Here is the link for where the presentation will be taking place.
Title: An Efficient and Robust Multi-Stream Framework for End-to-End Speech Recognition
Abstract: In the voice-enabled domestic or meeting environments, distributed microphone arrays aim to process distant-speech interaction into text with high accuracy. However, with dynamic corruption of noises and reverberations or human movement present, there is no guarantee that any microphone array (stream) is constantly informative. In these cases, an appropriate strategy to dynamically fuse streams or select the most informative array is necessary.
The multi-stream paradigm in Automatic Speech Recognition (ASR) considers scenarios where parallel streams carry diverse or complementary task-related knowledge. Such streams could be defined as microphone arrays, frequency bands, various modalities or etc. Hence, a robust stream fusion is crucial to emphasize on more informative streams than corrupted ones, specially under unseen conditions. This thesis focuses on improving the performance and robustness of speech recognition in multi-stream scenarios.
In recent years, with the increasing use of Deep Neural Networks (DNNs) in ASR, End-to-End (E2E) approaches, which directly transcribe human speech into text, have received greater attention. In this thesis, a multi-stream framework is presented based on joint Connectionist Temporal Classification/Attention (CTC/ATT) E2E model, where parallel streams are represented by separate encoders. On top of the regular attention networks, a secondary stream-fusion network is to steer the decoder toward the most informative streams. Two representative frameworks are proposed, which are Multi-Encoder Multi-Array (MEM-Array) and Multi-Encoder Multi-Resolution (MEM-Res), respectively.
The MEM-Array model aims at improving the far-field ASR robustness using multiple microphone arrays which are activated by separate encoders. With an increasing number of streams (encoders) requiring substantial memory and massive amounts of parallel data, a practical two-stage training strategy is desgnated to address these issues. Furthermore, a two-stage augmentation scheme is present to improve the robustness of the multi-stream model, where small amount of parallel data is sufficient to achieve competitive results. In MEM-Res, two heterogeneous encoders with different architectures, temporal resolutions and separate CTC networks work in parallel to extract complementary information from same acoustics. Compared with the best single-stream performance, both models have achieved substantial improvement, which also outperform various conventional fusion strategies.
While proposed framework optimizes information in multi-stream scenarios, this thesis also studies the Performance Monitoring (PM) measures to predict if recognition result of an end-to-end model is reliable, without growth-truth knowledge. Four different PM techniques are investigated, suggesting that PM measures on attention distributions and decoder posteriors are well-correlated with true performances.
Hynek Hermansky, Department of Electrical and Computer Engineering
Shinji Watanabe, Department of Electrical and Computer Engineering
Najim Dehak, Department of Electrical and Computer Engineering
Gregory Sell, JHU Human Language Technology Center of Excellence
Note: This is a virtual presentation. Here is the link for where the presentation will be taking place.
Title: Localizing Seizure Foci with Deep Neural Networks and Graphical Models
Abstract: Worldwide estimates of the prevalence of epilepsy range from 1-3% of the total population, making it one of the most common neurological disorders. With its wide prevalence and dramatic effects on quality of life, epilepsy represents a large and ongoing public health challenge. Critical to the treatment of focal epilepsy is the localization of the seizure onset zone. The seizure onset zone is defined as the region of the cortex responsible for the generation of seizures. In the clinic, scalp electroencephalography (EEG) recording is the first modality used to localize the seizure onset zone.
My work focuses on developing machine learning techniques to localize this zone from these recordings. Using Bayesian techniques, I will present graphical models designed to captures the observed spreading of seizures in clinical EEG recordings. These models directly encode clinically observed seizure spreading phenomena to capture seizure onset and evolution. Using neural networks, the raw EEG signal is evaluated is evaluated for seizure activity. In this talk I will propose extensions to these techniques employing semi-supervised learning and architectural improvements for training sophisticated neural networks designed to analyze scalp EEG signals. In addition, I will propose modeling improvements to current graphical models for evaluating the confidence of localization results.
Archana Venkataraman (Department of Electrical and Computer Engineering)
Sri Sarma (Department of Biomedical Engineering)
Rene Vidal (Department of Biomedical Engineering)
Richard Leahy (Department of Electrical Engineering Systems – University of Southern California)
Title: Leveraging Inverter-Based Frequency Control in Low-Inertia Power Systems
Abstract: The shift from conventional synchronous generation to renewable converter-interfaced sources has led to a noticeable degradation of power system frequency dynamics. Fortunately, recent technology advancements in power electronics and electric storage facilitate the potential to enable higher renewable energy penetration by means of inverter-interfaced storage units. With proper control approaches, fast inverter dynamics can ensure the rapid response of storage units to mitigate degradation. A straightforward choice is to emulate the damping effect and/or inertial response of synchronous generators through droop control or virtual inertia, yet they do not necessarily fully exploit the benefits of inverter-interfaced storage units. For instance, droop control sacrifices steady-state effort share to improve dynamic performance, while virtual inertia amplifies frequency measurement noise. This work thus seeks to challenge this naive choice of mimicking synchronous generator characteristics and instead advocate for a principled control design perspective. To achieve this goal, we build our analysis upon quantifying power network dynamic performance using $\mathcal L_2$ and $\mathcal L_\infty$ norms so as to perform a systematic study evaluating the effect of different control approaches on both frequency response metrics and storage economic metrics. The main contributions of this project will be as follows: (i) We will propose a novel dynamic droop control approach, for grid following inverters, that can be tuned to achieve low noise sensitivity, fast synchronization, and Nadir elimination, without affecting the steady-state performance; (ii) We will propose a new frequency shaping control approach that allows to trade-off between the rate of change of frequency (RoCoF) and storage conrol effort; (iii) We will further extend the proposed solutions to operate in a grid-forming setting that is suitable for a non-stiff power grid where the amplitude and frequency of grid voltage is not well-regulated.
Enrique Mallada (Department of Electrical & Computer Engineering)
Pablo A. Iglesias (Department of Electrical & Computer Engineering)
Dennice F. Gayme (Department of Mechanical Engineering)
Title: Accelerating Magnetic Resonance Imaging using Convolutional Recurrent Neural Networks
Abstract: Fast and accurate MRI image reconstruction from undersampled data is critically important in clinical practice. Compressed sensing based methods are widely used in image reconstruction but the speed is slow due to the iterative algorithms. Deep learning based methods have shown promising advances in recent years. However, recovering the fine details from highly undersampled data is still challenging. Moreover, Current protocol of Amide Proton Transfer-weighted (APTw) imaging commonly starts with the acquisition of high-resolution T2-weighted (T2w) images followed by APTw imaging at particular geometry and locations (i.e. slice) determined by the acquired T2w images. Although many advanced MRI reconstruction methods have been proposed to accelerate MRI, existing methods for APTw MRI lack the capability of taking advantage of structural information in the acquired T2w images for reconstruction. In this work, we introduce a novel deep learning-based method with Convolutional Recurrent Neural Networks (CRNN) to reconstruct the image from multiple scales. Finally, we explore the use of the proposed Recurrent Feature Sharing (RFS) reconstruction module to utilize intermediate features extracted from the matched T2w image by CRNN so that the missing structural information can be incorporated into the undersampled APT raw image thus effectively improving the image quality of the reconstructed APTw image.
Vishal M. Patel, Department of Electrical and Computer Engineering
Rama Chellappa, Department of Electrical and Computer Engineering
Shanshan Jiang, Department of Radiology and Radiological Science
Title: Deep Learning-based Heterogeneous Face Recognition
Abstract: Face Recognition (FR) is one of the most widely studied problems in computer vision and biometrics research communities due to its applications in authentication, surveillance, and security. Various methods have been developed over the last two decades that specifically attempt to address the challenges such as aging, occlusion, disguise, variations in pose, expression, and illumination. In particular, convolutional neural network (CNN) based FR methods have gained significant traction in recent years. Deep CNN-based methods have achieved impressive performances on the current FR benchmarks. Despite the success of CNN-based methods in addressing various challenges in FR, they are fundamentally limited to recognizing face images that are collected near-infrared spectrum. In many practical scenarios such as surveillance in low-light conditions, one has to detect and recognize faces that are captured using thermal cameras. However, the performance of many deep learning-based methods degrades significantly when they are presented with thermal face images.
Thermal-to-visible face verification is a challenging problem due to the large domain discrepancy between the modalities. Existing approaches either attempt to synthesize visible faces from thermal faces or extract robust features from these modalities for cross-modal matching. We present a work in which we use attributes extracted from visible images to synthesize the attribute-preserved visible images from thermal imagery for cross-modal matching. A pre-trained VGG-Face network is used to extract the attributes from the visible image. Then, a novel multi-scale generator is proposed to synthesize the visible image from the thermal image guided by the extracted attributes. Finally, a pre-trained VGG-Face network is leveraged to extract features from the synthesized image and the input visible image for verification.
Carlos Castillo, Department of Electrical and Computer Engineering
Vishal Patel, Department of Electrical and Computer Engineering
Title: Retina OCT image analysis using deep learning methods
Abstract: Optical coherence topography (OCT) is a non-invasive imaging modality which uses low-coherence light waves to take cross-sectional images of optical scattering media (e.g., the human retina). OCT has been widely used in diagnosing retinal and neural diseases by imaging the human retina. The thickness of retina layers are important biomarkers for neurological diseases like multiple sclerosis (MS). The peripapillary retinal nerve fiber layer (pRNFL) and ganglion cell plus inner plexiform layer (GCIP) thickness can be used to assess global disease progression of MS patient. Automated OCT image analysis tools are critical for quantitatively monitoring disease progression and explore biomarkers. With the development of more powerful computational resources, deep learning based methods have achieved much better performance in accuracy, speed, and algorithm flexibility for many image analysis tasks. However, these emerging deep learning methods are not satisfactory when directly applied to OCT image analysis tasks like retinal layer segmentation if not using task specific knowledge.
This thesis aims to develop a set of novel deep learning based methods for retinal OCT image analysis. Specifically, we are focusing on retinal layer segmentation from macular OCT images. Image segmentation is the process of classifying each pixel in a digital image into different classes. Deep learning methods are powerful classifiers in pixel classification, but it is hard to incorporate explicit rules. For retinal layer OCT images, pixels belonging to different layer classes must satisfy the anatomical hierarchy (topology): pixels of the upper layers should have no overlap or gap with pixels of layers beneath it. This topological criterion is usually achieved by sophisticated post-processing methods, which current deep learning method cannot guarantee. To solve this problem, we aim to:
The deep learning model’s performance will degrade badly when test data is generated differently from the training data; thus, we aim to
The deep learning pipeline will be used to analyze longitudinal OCT images for MS patients, where the subtle changes due to the MS should be captured; thus, we aim to:
Title: Medical Image Modality Synthesis and Resolution Enhancement Based on Machine Learning Techniques
Abstract: To achieve satisfactory performance from automatic medical image analysis algorithms such as registration or segmentation, medical imaging data with the desired modality/contrast and high isotropic resolution are preferred, yet they are not always available. We addressed this problem in this thesis using 1) image modality synthesis and 2) resolution enhancement.
The first contribution of this thesis is computed tomography (CT)-to-magnetic resonance imaging (MRI) image synthesis method, which was developed to provide MRI when CT is the only modality that is acquired. The main challenges are that CT has poor contrast as well as high noise in soft tissues and that the CT-to-MR mapping is highly nonlinear. To overcome these challenges, we developed a convolutional neural network (CNN) which is a modified U-net. With this deep network for synthesis, we developed the first segmentation method that provides detailed grey matter anatomical labels on CT neuroimages using synthetic MRI.
The second contribution is a method for resolution enhancement for a common type of acquisition in clinical and research practice, one in which there is high resolution (HR) in the in-plane directions and low resolution (LR) in the through-plane direction. The challenge of improving the through-plane resolution for such acquisitions is that the state-of-art convolutional neural network (CNN)-based super-resolution methods are sometimes not applicable due to lack of external LR/HR paired training data. To address this challenge, we developed a self super-resolution algorithm called SMORE and its iterative version called iSMORE, which are CNN-based yet do not require LR/HRpaired training data other than the subject image itself. SMORE/iSMOREcreate training data from the HR in-plane slices of the subject image itself, then train and apply CNNs to through-plane slices to improve spatial resolution and remove aliasing. In this thesis, we perform SMORE/iSMORE on multiple simulated and real data sets to demonstrate their accuracy and generalizability. Also, SMORE as a preprocessing step is shown to improve segmentation accuracy.
In summary, CT-to-MR synthesis, SMORE, and iSMORE were demonstrated in this thesis to be effective preprocessing algorithms for visual quality and other automatic medical image analysis such as registration or segmentation.
Jerry Prince, Department of Electrical and Computer Engineering
John Goutsias, Department of Electrical and Computer Engineering
Trac Tran, Department of Electrical and Computer Engineering
Title: Detecting Unknown Instances Using CNNs
Abstract: Deep convolutional neural networks (DCNNs) have shown impressive performance improvements for object detection and recognition problems. However, a vast majority of DCNN-based recognition methods are designed for a closed world, where the primary assumption is that all categories are known a priori. In many real-world applications, this assumption does not necessarily hold. Generally, incomplete knowledge of the world is present at training time, and unknown classes can be submitted to an algorithm during testing. The goal of a visual recognition system is then to reject samples from unknown classes and classify samples from known classes.
In the first part of my talk, I will present new DCNNs for anomaly detection based on one-class classification. The main idea is to use a zero centered Gaussian noise in the feature space as the pseudo-negative class and train the network using the cross-entropy loss. Also, a method in which both classifier and feature representations are learned together in an end-to-end fashion will be presented. In the second part of the talk, I will present a multi-class category detection using a network which utilizes both global and local information to predict whether the test image belongs to one of the known classes or an unknown category. Specifically, the models is trained using a network to perform image-level category prediction and another network to perform patch-level category prediction. We evaluate the effectiveness all these methods on multiple publicly available datasets and show that these approaches achieve better performance compared to previous state-of-the-art methods.
Title: Student-Teacher Learning Techniques for Bilingual and Low Resource OCR
Abstract: Optical Character Recognition (OCR) is the automatic generation of a transcription given a line image of text. Current methods have been very successful on printed English text, with Character Error Rates of less than 1¥%. However, clean datasets are not commonly seen in real life applications. There is a move in OCR towards `text in the wild’, conditions where there are lower resolution images like store fronts, street sign, and billboards. Oftentimes these texts contain multiple scripts, especially in countries where multiple languges are spoken. In addition, Latin characters are wildly seen no matter what language. The presence of multilingual text poses a unique challenge.
Traditional OCR methods involve text localization, script identification, and then text recognition. A separate system is used in each task and the results from one system are passed to the next. However, the downside of this pipeline approach is that errors propagate downstream and there is no way of providing feedback upstream. These downsides can be mitigated with fully integrated approaches, where one large system does text localization, script identification, and text recognition jointly. These approaches are also sometimes known as end-to-end approaches in literature.
With larger and larger networks, there is also a need for a greater amount of training data. However, this data may be difficult to obtain if the target language is low resource. There are also problems if the data that is obtained is in a slightly different domain, for example, printed versus handwritten text. This is where synthetic data generation techniques and domain adaptation techniques can be helpful.
Given these current challenges in OCR, this thesis proposal is focused on training an integrated (ie: end-to-end) bilingual systems and domain adaptation techniques. Both these objectives can be achieved using student-teacher learning methods. The basics of this approach is to have a trained teacher model add an additional loss function while training a student model. The outputs of the teacher will be used as soft targets for the student to learn. The following experiments will be performed: