R-3 Repository :: Browsing by Author "Baraniuk, Richard G"

Browsing by Author "Baraniuk, Richard G"

Now showing 1 - 20 of 22

A New Framework for Rapid, Scalable Bacterial Diagnostics with Microfluidics and Compressed Sensing
(2022-08-02) Kota, Pavan Kumar; Drezek, Rebekah A; Baraniuk, Richard G
Critically ill patients with suspected invasive bacterial infections can be treated effectively once the responsible pathogens are characterized. Rapid and sensitive diagnostic tests rely on a specific sensor for each target pathogen, a paradigm that cannot practically cover the dozens to hundreds of plausible pathogens. As a result, sample culture is a prerequisite for downstream testing but can take days. Clinicians are forced to use broad-spectrum antibiotics in the interim which are ineffective for some patients and contribute to rising drug resistance. To address these challenges, I present a new approach with compressed sensing (CS) and microfluidics. CS focuses on the efficient recovery of sparse signals. Patient samples from normally sterile sites are sparse because among many pathogens to consider, at most a few are causing any given infection. Microfluidic partitioning technologies split a sample into thousands of compartments, capturing bacterial biomarkers across the partitions according to a Poisson distribution. Leveraging this statistical prior can yield scalable systems that break previous theoretical barriers in CS, enabling the detection of more analytes with arbitrarily few sensors. First, I cover the theory and our newly developed Sparse Poisson Recovery (SPoRe) algorithm. I present SPoRe’s superior performance over baseline CS algorithms, including its high sensitivity and tolerance of measurement noise. Second, I generalize the theoretical results to apply to many common types of biosensors and present the first in vitro realization of SPoRe with droplet digital PCR. We use five DNA probes to barcode the 16S rRNA gene to quantify nine pathogen genera within hours. Given two fluorescent channels, we measure portions of the barcodes in four groups of droplets, pool the data, and infer bacterial quantities with SPoRe. I highlight how the underlying principles of this demonstration enable sensor-constrained systems to scalably cover large panels of analytes, an advance that could unlock new biosensing solutions for multiple industries.
Algorithms to Find the Girth and Cogirth of a Linear Matroid
(2014-09-18) Arellano, John David; Hicks, Illya V; Tapia, Richard A; Yin, Wotao; Baraniuk, Richard G
In this thesis, I present algorithms to find the cogirth and girth, the cardinality of the smallest cocircuit and circuit respectively, of a linear matroid. A set covering problem (SCP) formulation of the problems is presented. The solution to the linear matroid cogirth problem provides the degree of redundancy of the corresponding sensor network, and allows for the evaluation of the quality of the network. Hence, addressing the linear matroid cogirth problem can lead to significantly enhancing the design process of sensor networks. The linear matroid girth problem is related to reconstructing a signal in compressive sensing. I provide an introduction to matroids and their relation to the degree of redundancy problem as well as compressive sensing. I also provide an overview of the methods used to address linear matroid cogirth/girth problems, the SCP, and reconstructing a signal in compressive sensing. Computational results are provided to validate a branch-and-cut algorithm that addresses the SCP formulation as well as an algorithm which uses branch decompositions and dynamic programming to find the girth of a linear matroid.
Embargo
Covariate Balancing Methods for Randomized Controlled Trials Are Not Adversarially Robust
(2023-06-20) Babaei, Hossein; Baraniuk, Richard G; Segarra, Santiago; Sabharwal, Ashutosh
The first step towards investigating the effectiveness of a treatment via a randomized trial is to split the population into control and treatment groups then compare the average response of the treatment group receiving the treatment to the control group receiving the placebo. In order to ensure that the difference between the two groups is caused only by the treatment, it is crucial that the control and the treatment groups have similar statistics. Indeed, the validity and reliability of a trial are determined by the similarity of two groups’ statistics. Covariate balancing methods increase the similarity between the distributions of the two groups’ covariates. However, often in practice, there are not enough samples to accurately estimate the groups’ covariate distributions. In this thesis, we empirically show that covariate balancing with the Standardized Means Difference (SMD) covariate balancing measure, as well as Pocock’s sequential treatment assignment method, are susceptible to worst-case treatment assignments. Worst-case treatment assignments are those admitted by the covariate balance measure, but result in highest possible ATE estimation errors. We developed an adversarial attack to find adversarial treatment assignment for any given trial. Then, we provide an index to measure how close the given trial is to the worst-case. To this end, iii we provide an optimization-based algorithm, namely Adversarial Treatment ASsignment in TREatment Effect Trials (ATASTREET), to find the adversarial treatment assignments.
Denoising-based Approximate Message Passing
(2014-12-05) Metzler, Chris A; Baraniuk, Richard G; Veeraraghavan, Ashok; Zhang, Yin
A denoising algorithm seeks to remove perturbations or errors from a signal. The last three decades have seen extensive research devoted to this arena, and as a result, today's denoisers are highly optimized algorithms that effectively remove large amounts of additive white Gaussian noise. A compressive sensing (CS) reconstruction algorithm seeks to recover a structured signal acquired using a small number of randomized measurements. Typical CS reconstruction algorithms can be cast as iteratively estimating a signal from a perturbed observation. This thesis answers a natural question: How can one effectively employ a generic denoiser in a CS reconstruction algorithm? In response, in this thesis, I propose a denoising-based approximate message passing (D-AMP) algorithm that is capable of high-performance reconstruction. I demonstrate that, when used with a high performance denoiser, D-AMP offers state-of-the-art CS recovery performance for natural images while operating tens of times faster than the only competitive method. In addition, I explain the exceptional performance of D-AMP by analyzing some of its theoretical features. A critical insight into this approach is the use of an appropriate Onsager correction term in the D-AMP iterations, which coerces the signal perturbation at each iteration to be distributed approximately like the white Gaussian noise that denoisers are typically designed to remove. In doing so, this feature enables the algorithm to effectively use nearly any denoiser.
Face detection and verification with FlatCam lensless imaging system
(2018-08-09) Tan, Jasper; Baraniuk, Richard G
Progress in any technological area requires distinct breakthrough ideas. In the field of imaging, lensless imaging technology is a disruptive concept that allows cameras to continue getting thinner and cheaper. The FlatCam lensless imaging system demonstrates this by replacing the thick and expensive lens of a conventional camera with a thin and cheap aperture mask and a reconstruction algorithm. Indeed, such a design allows recognizable image capture, albeit with much lower resolution and greater noise than conventional lens-based cameras. The true disruptive ability of FlatCam in society is its potential to fuel a machine's capability of obtaining a wealth of information from the world via images, a common step in the pipeline of machine intelligence. In this work, I rigorously demonstrate and evaluate performing face detection and verification, two such intelligence tasks, with FlatCam images. To perform face detection and verification, I propose and adapt basic deep learning techniques to handle the resolution, noise, and artifacts inherent with the FlatCam. I show with common evaluation protocols that there is only a small decrease in accuracy when using FlatCam images compared to the original lens-based images. Furthermore, I describe the construction of a face dataset captured with a FlatCam prototype containing 23,368 lensless camera images of 92 subjects in a range of different operating conditions. Further evaluating face verification on this dataset verifies the FlatCam's potential for performing inference tasks in real-world deployment.
FlatCam: Lensless Imaging, Principles, Applications and Fabrication
(2015-12-03) Ayremlou, Ali; Baraniuk, Richard G
FlatCam is a thin form-factor camera that consists of a coded mask placed on top of a bare, conventional sensor array. The design is inspired by coded aperture imaging principles; each sensor pixel records a linear combination of the scene in front of the camera, and a computational algorithm reconstructs the image. A key design feature of the FlatCam is its slim form-factor, which enables imaging using extremely thin, even flexible surfaces that operate over a wide spectral range and are amenable to monolithic fabrication. We demonstrate the potential of the FlatCam design using two prototypes: one at visible wavelengths and one at infrared wavelengths.
GEM Incorporating Context into Genomic Distance Estimation
(2019-06-04) Barberan, CJ; Baraniuk, Richard G
A pivotal question in computational biology is how similar two organisms are based on their genomic sequences. Unfortunately, classical sequence alignment-based methods for estimating genomic distances do not scale well to the massive number of organisms that have been sequenced to date. Recently, composition-based methods have gained interest due to their computational efficiencies for massive distance estimation problems. However, these methods reduce the computation time at the cost of distorting the genomic distances. The main problem with composition-based methods is their reliance on the occurrence of length-k subsequences of the genome, known as k-mers, which ignores their ordering, i.e., their context in the genome. In this thesis, we take inspiration from computational linguistics to develop a new genomic distance estimation approach that exploits not only the frequency of the k-mers but also their context. In our Genomic distance EstiMation (GEM) algorithm, we first learn a context-aware, low-dimensional embedding for k-mers by training on a large corpus of FASTA files comprising 159 million bases of whole genome sequence data from microbial organisms in the National Center of Biotechnology Information (NCBI) repository. We then define the distance between two organisms using a generalization of the Jaccard similarity that incorporates the context-aware embedding of the constituent k-mers. A range of experiments demonstrate that GEM estimates the distance between unseen organisms with up to 2 times less error compared to state-of-art algorithms while incurring a similar running time. As a bonus, the GEM context reveals a distinct structure in the ordering of k-mers in bacteria, viruses, and fungi, a finding that motivates follow-up evolutionary studies.
Imaging Plasmons with Compressive Hyperspectral Microscopy
(2015-04-23) Lu, Liyang; Kelly, Kevin F; Baraniuk, Richard G; Landes, Christy
With the ability of revealing the interactions between objects and electromagnetic waves, hyperspectral imaging in optical microscopy is of great importance in the study of various micro/nano-scale physical and chemical phenomena. The conventional methods, however, require various scanning processes to acquire a complete set of hyperspectral data because of its 3-dimensional structure. As such the quality and efficiency of the data acquisition using these conventional scanning techniques is greatly limited by the detector sensitivity and low signal light intensity from the sample. To overcome such limitations, we applied compressive sensing theory to the hyperspectral imaging. The compressive imaging enhances the measurement signal-to-noise ratio by encoding and combining the spatial information of the sample to the detector, and a recovery algorithm is used to decode the detector outputs and reconstruct the image. A microscopy system based on this compressive hyperspectral imaging scheme was designed and implemented. Further analysis and discussion on the diffraction and interference phenomenon and a solution to the spectral distortion in this compressive sensing microscopy system are also presented. Experimental results of compressive dark-field scattering from gold nanobelts are presented, followed with an analysis on signal-to-noise ratio and a comparison with conventional scanning methods in measuring the plasmon resonances.
Machine Learning in Large-scale Genomics: Sensing, Processing, and Analysis
(2017-04-19) Aghazadeh Mohandesi, Amir Ali; Baraniuk, Richard G
Advances in the field of genomics, a branch of biology concerning with the structure, function, and evolution of genomes, has led to dramatic reductions in the price of sequencing machines. As a result, torrents of genomic data is being produced every day which pose huge challenges and opportunities for engineers, scientists, and researchers in various fields. Here, we propose novel machine learning tools and algorithms to more efficiently sense, process, and analyze large-scale genomic data. To begin with, we develop a novel universal microbial diagnostics (UMD) platform to sense microbial organisms in an infectious sample, using a small number of random DNA probes that are agnostic to the target genomic DNA sequences. Our platform leverages the theory of sparse signal recovery (compressive sensing) to identify the composition of a microbial sample that potentially contains thousands of novel or mutant species. We next develop a new sensor selection algorithm that finds the subset of sensors that best recovers a sparse vector in sparse recovery problems. Our proposed algorithm, Insense, minimizes a coherence-based cost function that is adapted from classical results in sparse recovery theory and outperforms traditional selection algorithms in finding optimal DNA probes for microbial diagnostics problem. Inspired by recent progress in robust optimization, we then develop a novel hashing algorithm, dubbed RHash, that minimizes the worst-case distortion among pairs of points in a dataset using an \ell_infinity-norm minimization technique. We develop practical and efficient implementations of RHash based on the alternating direction method of multipliers (ADMM) framework and column generation that scale well to large datasets. Finally, we develop a novel machine learning algorithm using techniques in deep learning and natural language processing literature to embed DNA sequences of arbitrary length into a single low-dimensional space. Our so-called Kmer2Vec platform learns biological concepts such as drug-resistance by parsing raw DNA sequences of microbial organisms with no prior biology knowledge.
Mapping Questions to Textbook Content
(2018-04-18) Burmeister, Ryan J; Baraniuk, Richard G
Learning is an iterative process which consists of knowledge acquisition, assessment of supposed knowledge, identification of misconceptions, and refinement of understanding. Within courses, teachers often employ textbook review questions as a method of assessing student knowledge retention. However, student attempts to resolve misconceptions or reinforce concepts following these questions may leave them searching large expanses of textbook content. This project aims to alleviate this problem by providing formative feedback in the form of textbook passages. By utilizing question answering and reading comprehension methodologies, a search algorithm for selecting relevant material within textbooks is developed. The algorithm capitalizes on the inherent structure of textbooks which, intuitively, reduces the search space. Evaluation of question answering and information retrieval practices in the education domain is performed beyond simple factoid questions. Results are compared to data collected via subject matter experts, and limitations of these models, as they pertain to textbooks, are explored.
NeuroView: Explainable Deep Network Decision Making
(2022-07-06) Barberan, CJ; Baraniuk, Richard G; Balakrishnan, Guha
Deep neural networks (DNs) provide superhuman performance in numerous computer vision tasks, yet it remains unclear exactly which of a DN's units contribute to a particular decision. A deep network’s prediction cannot be explained in a formal mathematical manner such that you know how all the parameters contribute to the decision. NeuroView is a new family of DN architectures that are explainable by design. Each member of the family is derived from a standard DN architecture by concatenating all of the activations and feeding them into a global linear classifier. The resulting architecture establishes a direct, causal link between the state of each unit and the classification decision. We validate NeuroView on multiple datasets and classification tasks to show that it has on par performance to a typical DN. Also, we inspect how it’s unit/class mapping aids in understanding the decision-making process. In this thesis, we propose using NeuroView in other architectures such as convolutional and recurrent neural networks to show how it can aid in providing additional understanding in applications that need more explanation.
oASIS: Adaptive Column Sampling for Kernel Matrix Approximation
(2015-04-21) Patel, Raajen; Baraniuk, Richard G; Veeraraghavan, Ashok; Sorensen, Danny C
Kernel or similarity matrices are essential for many state-of-the-art approaches to classification, clustering, and dimensionality reduction. For large datasets, the cost of forming and factoring such kernel matrices becomes intractable. To address this challenge, we introduce a new adaptive sampling algorithm called Accelerated Sequential Incoherence Selection (oASIS) that samples columns without explicitly computing the entire kernel matrix. We provide conditions under which oASIS is guaranteed to exactly recover the kernel matrix with an optimal number of columns selected. Numerical experiments on both synthetic and real-world datasets demonstrate that oASIS achieves performance comparable to state-of-the-art adaptive sampling methods at a fraction of the computational cost. The low runtime complexity of oASIS and its low memory footprint enable the solution of large problems that are simply intractable using other adaptive methods.
On the Momentum-based Methods for Training and Designing Deep Neural Networks
(2020-09-14) Nguyen, Minh Tan; Baraniuk, Richard G
Training and designing deep neural networks (DNNs) are an art that often involves expensive search over candidate architectures and optimization algorithms. In my thesis, we develop novel momentum-based methods to speed up deep neural networks training and facilitate the process of designing them. For training DNNs, stochastic gradient descent (SGD) algorithms with constant momentum and its variants such as Adam are the optimization methods of choice for training DNNs. There is great interest in speeding up the convergence of these methods due to their high computational expense. Nesterov accelerated gradient (NAG) improves the convergence rate of gradient descent (GD) for convex optimization using a specially designed momentum; however, it accumulates error when an inexact gradient is used (such as in SGD), slowing convergence at best and diverging at worst. We propose scheduled restart SGD (SRSGD), a new NAG-style scheme for training DNNs. SRSGD replaces the constant momentum in SGD by the increasing momentum in NAG but stabilizes the iterations by resetting the momentum to zero according to a schedule. Using a variety of models and benchmarks for image classification, we demonstrate that, in training DNNs, SRSGD significantly improves convergence and generalization; for instance, in training ResNet-200 for ImageNet classification, SRSGD achieves an error rate of 20.93% vs. the benchmark of 22.13%. These improvements become more significant as the network grows deeper. Furthermore, on both CIFAR and ImageNet, SRSGD reaches similar or even better error rates with significantly fewer training epochs compared to the SGD baseline. For designing DNNs, we focus on the recurrent neural networks (RNNs) and establish a connection between the hidden state dynamics in an RNN and gradient descent (GD). We then integrate momentum into this framework and propose a new family of RNNs, called MomentumRNNs. We theoretically prove and numerically demonstrate that MomentumRNNs alleviate the vanishing gradient issue in training RNNs. We also demonstrate that MomentumRNN is applicable to many types of recurrent cells, including those in the state-of-the-art orthogonal RNNs. Finally, we show that other advanced momentum-based optimization methods, such as Adam and NAG with a restart, can be easily incorporated into the MomentumRNN framework for designing new recurrent cells with even better performance.
PIE: Perceptual Image Enhancement via Local Manifold Sampling on Pretrained Diffusion Models
(2023-12-01) Mayer, Paul Michael; Baraniuk, Richard G
PIE: Perceptual Image Enhancement via Local Manifold Sampling on Pretrained Diffusion Models
Random Acquisition Sensors and Receivers: Design, Architectures, and Applications
(2017-08-07) El Smaili, Sami; Baraniuk, Richard G
The emergence of concepts such as the internet of things (IoT) is but a manifestation of the increasing integration of communication and data acquisition systems in a wide range of devices and in a plethora of applications. Pushing the limits of traditional Nyquist-based systems, such widespread integration of acquisition systems is coupled with the increasing demand for lower power, higher bandwidth, multi-standard and reconfigurable systems. Compressive sensing promises to alleviate much of the constraints facing Nyquist-based systems and provide efficient solutions in all these areas. However, the current approaches for implementing compressive sensing suffer from several limitations: - A gap between theory and application: reconstruction algorithms use an ideal model of the system that the hardware is designed to mimic. However, in most cases, the ideal model might not be the only one that guarantees reconstruction accuracy, which results in an unduly constrained design. In fact, the theory frames the condition for successful reconstruction in terms of the very general concept of restricted isometry property (RIP) that can be applied to any system model. The ideality of the ideal model does not necessarily has its roots originating in the theory; a more practical model that is easier to mimic can still satisfy the RIP and be considered ideal. - General and widely used architectures have as parameters the number of acquired measurements and signal sparsity. Because these parameters are defined by the application rather than designed, there is little room to tweak the design, at the system level, to manage various system trade-offs and to overcome practical challenges such as the resetting frequency of an integrator or the window length in each projection channel. - Applicability to realms beyond that of sparse signals: In many communication applications, signals are not sparse (and should not be for bandwidth utilization), but random acquisition architecture can still provide tremendous benefits and provide novel solutions. Reaping the benefits of random acquisition in such domains requires new architectures that go beyond the traditional application scope of compressive sensing. The work we propose as part of this thesis aims at overcoming these limitations and expanding the realm of compressive sensing into new areas and applications. In this work we - Provide a practical approach to compressive sensing that starts from a practical system model to derive system requirements for successful reconstruction. We consider the basic building block of compressive sensing, the projection channel, and assume a general filter is used rather than an integrator (the ideal model). We derive the conditions that such a model should have to satisfy the restricted isometry property and show that the new requirements are far less restrictive and more practical than traditional requirements stemming from an integrator-based model. - Develop an approach for quantifying hardware variability and model uncertainty and its effect on reconstruction accuracy. Particularly, we study the effect of filter pole variability on reconstruction, which might be due to model approximations or hardware variability. - Propose a multi-channel random demodulator that bridges the gap between the two main architectures, the random demodulator consisting of one projection channel, and the random modulator pre integrator, which uses M channels for M measurements. The multi-channel random demodulator has the number of channels as a design parameter, which can be used to manage practical trade-offs such as the integrator reset frequency and the window length in each channel. We utilize this architecture in the reconfigurable receiver architecture that we also present in this work. - Develop a framework for random acquisition reconfigurable receivers, which expands the realm of compressive sensing to communication systems where signals are dense but their supports are known. We develop a design methodology for such systems, linking the various system parameters to system metrics. - Propose and analyze an ADC-less architecture for sensors that breaks from the conventional compressive sensing approach of digitizing measurements at the acquisition site. We study when this approach is more beneficial than the traditional approach of digitizing on the acquisition site.
Ridge Regularization by Randomization in Linear Ensembles
(2022-11-21) LeJeune, Daniel; Baraniuk, Richard G
Ensemble methods that average over a collection of independent predictors that are each limited to random sampling of both the examples and features of the training data command a significant presence in machine learning, such as the ever-popular random forest. Combining many such randomized predictors into an ensemble produces a highly robust predictor with excellent generalization properties; however, understanding the specific nature of the effect of randomization on ensemble method behavior has received little theoretical attention. We study the case of an ensembles of linear predictors, where each individual predictor is a linear predictor fit on a randomized sample of the data matrix. We first show a straightforward argument that an ensemble of ordinary least squares predictors fit on a simple subsampling can achieve the optimal ridge regression risk in a standard Gaussian data setting. We then significantly generalize this result to eliminate essentially all assumptions on the data by considering ensembles of linear random projections or sketches of the data, and in doing so reveal an asymptotic first-order equivalence between linear regression on sketched data and ridge regression. By extending this analysis to a second-order characterization, we show how large ensembles converge to ridge regression under quadratic metrics.
Self-Consuming Generative Models Go MAD
(2024-07-30) Casco-Rodriguez, Josue; Baraniuk, Richard G
Seismic advances in generative AI algorithms for imagery, text, and other data types has led to the temptation to use synthetic data to train next-generation models. Repeating this process creates an autophagous (self-consuming) loop whose properties are poorly understood. We conduct a thorough analytical and empirical analysis using state-of-the-art generative image models of three families of autophagous loops that differ in how fixed or fresh real training data is available through the generations of training and in whether the samples from previous generation models have been biased to trade off data quality versus diversity. Our primary conclusion across all scenarios is that without enough fresh real data in each generation of an autophagous loop, future generative models are doomed to have their quality (precision) or diversity (recall) progressively decrease. We term this condition Model Autophagy Disorder (MAD), making analogy to mad cow disease.
Single-Frame 3D Fluorescence Microscopy with Ultra-Miniature Lensless FlatScope
(2017-06-01) Adams, Jesse K; Robinson, Jacob T; Baraniuk, Richard G; Landes, Christy
Modern biology increasingly relies on fluorescence microscopy, which is driving demand for smaller, lighter, and cheaper microscopes. However, traditional microscope architectures suffer from a fundamental tradeoff: as lenses become smaller, they must either collect less light or image a smaller field of view. To break this fundamental tradeoff between device size and performance, we present a new concept for 3D fluorescence imaging that replaces lenses with an optimized amplitude mask placed a few hundred microns above the sensor and an efficient algorithm that can convert a single frame of captured sensor data into high-resolution 3D images. The result is FlatScope: a lensless microscope that is scarcely larger than an image sensor (roughly 0.2 grams in weight and less than 1 mm thick) and yet able to produce micron-resolution, high-frame-rate, 3D fluorescence movies covering a total volume of several cubic millimeters. The ability of FlatScope to reconstruct full 3D images from a single frame of captured sensor data allows us to image 3D volumes roughly 40,000 times faster than a laser scanning confocal microscope. We envision that this new flat fluorescence microscopy paradigm will lead to implantable endoscopes that minimize tissue damage, arrays of imagers that cover large areas, and bendable, flexible microscopes that conform to complex topographies.
Sound-to-Touch Sensory Substitution and Beyond
(2015-06-29) Novich, Scott David; Eagleman, David M; Baraniuk, Richard G; Burrus, Charles S; Cox, Steven J; O'Malley, Marcia K
Deafness affects an estimated 2 million people in the United States and 53 million worldwide. Cochlear implants are an effective therapeutic solution but suffer from a number of drawbacks: they are expensive, require an invasive surgery, have low efficacy in early-onset deaf adults, and not everyone who needs them may qualify for them. This creates a large unmet need for a solution that is affordable, non-surgical, and works in adults. "Sensory substitution"--the concept that information can be effectively mapped from one sense to another--has the potential to overcome all of these issues: it is non-invasive (and therefore inexpensive and less regulated) and leverages a developed sensory modality. This is realized by the understanding that (1) the nervous system ultimately encodes information as electrical signals and (2) the brain has the remarkable capability of cortical reorganization. Sensory substitution has previously been successfully applied as a solution for blindess via vision-to-touch substitutions. To this end, a sound-to-touch sensory substitution device, The Versatile Extra-Sensory Transducer (VEST), has been developed for this thesis-work as a means for overcoming deafness. The device consists of a smartphone that takes sound from the environment and converts this information to patterns of vibration on the torso. This occurs via an array of vibratory motors embedded on a vest that is worn under the user's clothing. It is capable of giving congenitally deaf individuals the ability to intuit speech. The development of this device serves as a motivating example for a more general guiding framework that applies to sensory substitution and augmentation devices.
The Deep Rendering Model: Bridging Theory and Practice in Deep Learning
(2018-10-10) Nguyen, Minh Tan; Baraniuk, Richard G
A grand challenge in machine learning is the development of computational algorithms that match or outperform humans in perceptual inference tasks such as visual object and speech recognition. The key factor complicating such tasks is the presence of numerous nuisance variables, for instance, the unknown object position, orientation, and scale in object recognition or the unknown voice pronunciation, pitch, and speed in speech recognition. Recently, a new breed of deep learning algorithms has emerged for high-nuisance inference tasks; they are constructed from many layers of alternating linear and nonlinear processing units and are trained using large-scale algorithms and massive amounts of training data. The recent success of deep learning systems is impressive— they now routinely yield pattern recognition systems with near or super-human capabilities — but a fundamental question remains: Why do they work? Intuitions abound, but a coherent framework for understanding, analyzing, and synthesizing deep learning architectures has remained elusive. We answer this question by developing a new probabilistic framework for deep Learning, namely the Deep Rendering Model (DRM), based on a Bayesian generative probabilistic model that explicitly captures variation due to nuisance variables. The graphical structure of the model enables it to be learned from data using classical expectation-maximization techniques. Furthermore, by relaxing the generative model to a discriminative one, we can recover deep convolutional neural networks (DCNs) as well as its variants including the deep residual networks (ResNet) and the densely connected convolutional networks (DenseNet), providing insights into their successes and shortcomings as well as a principled route to their improvement. The DRMM is also applicable to semi-supervised and unsupervised learning tasks, achieving results that are state-of-the-art in several categories on the MNIST benchmark and comparable to state of the art on the CIFAR10 benchmark.

Browsing by Author "Baraniuk, Richard G"

Results Per Page

Sort Options