Дисертації з теми "FEATURE ENCODING"

Our sensory experiences are encoded in the patterns of activity of the neurons in our brain. While we know we are capable of sensing and responding to a constantly changing sensory environment, we often study neural activity by repeatedly presenting the same stimulus and analyzing the average neural response. It is not understood how the average neural response represents the dynamic neural activity that produces our perceptions. In this work, we use functional imaging of the rodent primary somatosensory cortex, specifically the whisker representations, and apply classic signal-detection methods to test the predictive power of the average neural response. Stimulus features such as intensity are thought to be perceptually separable from the average representation; however, we show that stimulus intensity cannot be reliably decoded from neural activity from only a single experience. Instead, stimulus intensity was encoded only across many experiences. We observed this probabilistic neural code in multiple classic sensory paradigms including complex temporal stimuli (pairs of whisker deflections) and multi-whisker stimuli. These data suggest a novel framework for the encoding of stimulus features in the presence of high-neural variability. Specifically we suggest that our brains can compensate for unreliability by encoding information redundantly across cortical space. This thesis predicts that a somatosensory stimulus is not encoded identically each time it is experienced; instead, our brains use multiple redundant pathways to create a reliable sensory percept.

Machine learning methods can be used for solving important binary classification tasks in domains such as display advertising and recommender systems. In many of these domains categorical features are common and often of high cardinality. Using one-hot encoding in such circumstances lead to very high dimensional vector representations, causing memory and computability concerns for machine learning models. This thesis investigated the viability of a binary encoding scheme in which categorical values were mapped to integers that were then encoded in a binary format. This binary scheme allowed for representing categorical features using log2(d)-dimensional vectors, where d is the dimension associated with a one-hot encoding. To evaluate the performance of the binary encoding, it was compared against one-hot and feature hashed representations with the use of linear logistic regression and neural networks based models. These models were trained and evaluated using data from two publicly available datasets: Criteo and Census. The results showed that a one-hot encoding with a linear logistic regression model gave the best performance according to the PR-AUC metric. This was, however, at the expense of using 118 and 65,953 dimensional vector representations for Census and Criteo respectively. A binary encoding led to a lower performance but used only 35 and 316 dimensions respectively. For Criteo, binary encoding suffered significantly in performance and feature hashing was perceived as a more viable alternative. It was also found that employing a neural network helped mitigate any loss in performance associated with using binary and feature hashed representations.
Maskininlärningsmetoder kan användas för att lösa viktiga binära klassificeringsuppgifter i domäner som displayannonsering och rekommendationssystem. I många av dessa domäner är kategoriska variabler vanliga och ofta av hög kardinalitet. Användning av one-hot-kodning under sådana omständigheter leder till väldigt högdimensionella vektorrepresentationer. Detta orsakar minnesoch beräkningsproblem för maskininlärningsmodeller. Denna uppsats undersökte användbarheten för ett binärt kodningsschema där kategoriska värden var avbildade på heltalvärden som sedan kodades i ett binärt format. Detta binära system tillät att representera kategoriska värden med hjälp av log2(d) -dimensionella vektorer, där d är dimensionen förknippad med en one-hot kodning. För att utvärdera prestandan för den binära kodningen jämfördes den mot one-hot och en hashbaserad kodning. En linjär logistikregression och ett neuralt nätverk tränades med hjälp av data från två offentligt tillgängliga dataset: Criteo och Census, och den slutliga prestandan jämfördes. Resultaten visade att en one-hot kodning med en linjär logistisk regressionsmodell gav den bästa prestandan enligt PR-AUC måttet. Denna metod använde dock 118 och 65,953 dimensionella vektorrepresentationer för Census respektive Criteo. En binär kodning ledde till en lägre prestanda generellt, men använde endast 35 respektive 316 dimensioner. Den binära kodningen presterade väsentligt sämre specifikt för Criteo datan, istället var hashbaserade kodningen en mer attraktiv lösning. Försämringen i prestationen associerad med binär och hashbaserad kodning kunde mildras av att använda ett neuralt nätverk.

To enable the identification and localization of sounds in our environment, auditory brain centers must form representations that accurately encode distinct acoustic properties, but also integrate those properties to support a unified percept of an auditory object. These parallel operations of decomposition and integration are carried out by hierarchically organized processing regions which progressively reformat peripheral electrical impulses into signals that may be integrated into higher order brain circuits. To investigate the nature of these transformations and their vulnerability to hearing loss, I recorded extracellular responses in the auditory midbrain and cortex of awake mice. The first aim of this project was to study the multiparametric tuning characteristics of single neurons using an online stimulus optimization algorithm. Closed-loop stimulus tailoring rapidly revealed diverse multiparametric tuning, and further revealed the conservation of response sparseness between the two areas. I then tracked the recovery of central feature encoding in mice with profound cochlear neuropathy. I recorded from midbrain and cortex at two timepoints after nerve degeneration, observing a progressive recovery of responsiveness in both areas, which occurred earlier and was more robust in the cortex. Concurrently, several aspects of the once-precise temporal response properties in midbrain were persistently degraded, and classification of speech tokens in the cortex did not recover to control levels of accuracy. I hypothesize that compensatory central plasticity may support the recovery of feature encoding in the auditory pathway to a large extent, although various aspects of temporal encoding remain impaired. This may underlie the observation that some human patients with auditory neuropathy have profound deficits in speech comprehension despite having normal hearing thresholds. Finally, I tested the effect of AUT3, a novel positive modulator of the Kv3.1 potassium channel, on the encoding and classification of pulse trains and speech tokens in the midbrain. I observed that adjusting the excitability of central auditory neurons with this compound can partially restore the precision and reliability of spiking responses after hearing loss.

Thesis (Ph. D.)--Harvard-MIT Division of Health Sciences and Technology, 2011.
Cataloged from PDF version of thesis.
Includes bibliographical references (p. 150-159).
This thesis builds on previous efforts to develop tactile speech-reception aids for the hearing-impaired. Whereas conventional hearing aids mainly amplify acoustic signals, tactile speech aids convert acoustic information into a form perceptible via the sense of touch. By facilitating visual speechreading and providing sensory feedback for vocal control, tactile speech aids may substantially enhance speech communication abilities in the absence of useful hearing. Research for this thesis consisted of several lines of work. First, tactual detection and temporal order discrimination by congenitally deaf adults were examined, in order to assess the practicability of encoding acoustic speech information as temporal relationships among tactual stimuli. Temporal resolution among most congenitally deaf subjects was deemed adequate for reception of tactually-encoded speech cues. Tactual offset-order discrimination thresholds substantially exceeded those measured for onset-order, underscoring fundamental differences between stimulus masking dynamics in the somatosensory and auditory systems. Next, a tactual speech transduction scheme was designed with the aim of extending the amount of articulatory information conveyed by an earlier vocoder-type tactile speech display strategy. The novel transduction scheme derives relative amplitude cues from three frequency-filtered speech bands, preserving the cross-channel timing information required for consonant voicing discriminations, while retaining low-frequency modulations that distinguish voiced and aperiodic signal components. Additionally, a sensorimotor training approach ("directed babbling") was developed with the goal of facilitating tactile speech acquisition through frequent vocal imitation of visuo-tactile speech stimuli and attention to tactual feedback from one's own vocalizations. A final study evaluated the utility of the tactile speech display in resolving ambiguities among visually presented consonants, following either standard or enhanced sensorimotor training. Profoundly deaf and normal-hearing participants trained to exploit tactually-presented acoustic information in conjunction with visual speechreading to facilitate consonant identification in the absence of semantic context. Results indicate that the present transduction scheme can enhance reception of consonant manner and voicing information and facilitate identification of syllableinitial and syllable-final consonants. The sensorimotor training strategy proved selectively advantageous for subjects demonstrating more gradual tactual speech acquisition. Simple, low-cost tactile devices may prove suitable for widespread distribution in developing countries, where hearing aids and cochlear implants remain unaffordable for most severely and profoundly deaf individuals. They have the potential to enhance verbal communication with minimal need for clinical intervention.
by Theodore M. Moallem.
Ph.D.

This thesis contains a review of previous work in the fields of encoding sequential healthcare data and predicting graft- versus- host disease, a medical condition, based on patient history using machine learning. A new encoding of such data is proposed for machine learning purposes. The proposed encoding, called bag of binned weighted events, is a combination of two strategies proposed in previous work, called bag of binned events and bag of weighted events. An empirical experiment is designed to evaluate the predictive performance of the proposed encoding over various binning windows to that of the previous encodings, based on the area under the receiver operating characteristic curve (AUC) metric. The experiment is carried out on real- world healthcare data obtained from Swedish registries, using the random forest and the logistic regression algorithms. After filtering the data, solving quality issues and tuning hyperparameters of the models, final results are obtained. These results indicate that the proposed encoding strategy performs on par, or slightly better than the bag of weighted events, and outperforms the bag of binned events in most cases. However, differences in metrics show small differences. It is also observed that the proposed encoding usually performs better with longer binning windows which may be attributed to data noise. Future work is proposed in the form of repeating the experiment with different datasets and models, as well as changing the binning window length of the baseline algorithms.
Denna avhandling innehåller en recension av tidigare arbete inom områden av kodning av sekventiell sjukvårdsdata och förutsägelse av transplantat- mot- värdsjukdom, ett medicinskt tillstånd, baserat på patienthistoria med maskininlärning. En ny kodning av sådan data föreslås i maskininlärningssyfte. Den föreslagna kodningen, kallad bag of binned weighted events, är en kombination av två strategier som föreslagits i tidigare arbete, kallad bag of binned events och bag of weighted events. Ett empiriskt experiment är utformat för att utvärdera den föreslagna prestandan för den föreslagna kodningen över olika binningfönster jämfört med tidigare kodningar, baserat på AUC- måttet. Experimentet utförs på verkliga sjukvårdsdata som erhållits från svenska register, med random forest och logistic regression. Efter filtrering av data, lösning av kvalitetsproblem och justering av hyperparametrar för modellerna, erhålls slutliga resultat. Dessa resultat indikerar att den föreslagna kodningsstrategin presterar i nivå med, eller något bättre än bag of weighted events, och överträffar i de flesta fall bag of binned events. Skillnader i mått är dock små. Det observeras också att den föreslagna kodningen vanligtvis fungerar bättre med längre binningfönster som kan tillskrivas dataljud. Framtida arbete föreslås i form av att upprepa experimentet med olika datamängder och modeller, samt att ändra binningfönstrets längd för basalgoritmerna.

Previous evidence has shown that older adults may have specific declines in prefrontal cortex (PFC)-mediated processes supported source memory retrieval, such as strategic retrieval and post-retrieval monitoring. This decline may manifest in the form of attenuated late-frontal ERP effects. Behavioral research suggests that explicitly integrating a target context, or source, with a stimulus during encoding will improve subsequent source memory performance for both younger and older adults. Explicit item-feature binding instructions during encoding may alleviate source memory impairments, in part, by reducing the need for strategic processing during episodic retrieval. The present ERP study investigated whether explicit direction of attention toward item-feature integration may reduce age-related deficits in source memory by alleviating the necessity of frontally-mediated strategic processing at retrieval. Results demonstrated that explicit direction of attention improved source memory accuracy for both young and older adults, but older adults benefited less than the young, indicating additional age-related deficits. ERPs revealed that explicit encoding support attenuated post-retrieval monitoring effects in the young. In the old, explicit encoding instruction resulted in earlier onset of early frontal effects, possibly related to familiarity. Results suggest explicit direction of attention toward item-source integration at encoding may improve source memory by alleviating the need for strategic retrieval, but age-related deficits persist.

Thesis: Ph. D., Massachusetts Institute of Technology, Department of Chemical Engineering, February 2018.
Cataloged from PDF version of thesis. "February 2017." Handwritten on title page "February 2018."
Includes bibliographical references.
It is both uncontroversial and unassailable to assert that small things can often go where big things cannot. It is similarly prosaic to note that at smaller length scales, matter behaves differently than at larger length scales. This thesis exploits these intuitive and simple axioms to yield advances in three independent lines of enquiry: (i) robust and practically accessible encoding of information within microparticles, (ii) rapid, quantitative sensing of hydrophobic colloids and (iii) immunologically-focused drug delivery. Specifically, upconversion nanocrystals are used as the foundation of a novel spatial-spectral patterning motif to produce polymer microparticles with unique, decodable identities. With large single-particle encoding capacities (>10-⁶), an ultralow decoding false alarm rate (<10-⁹), and pronounced insensitivity to both particle chemistry and harsh processing conditions, this architecture enables practical deployment of encoded microparticles in applications with orthogonal requirements, including multiplexed bioassays and covert labeling of objects and packaging for anti-counterfeiting. Next, the large specific surface area of nanoscale objects is exploited by a family of zwitterionic, surfactant-like molecular rotors to develop a broadly applicable tool for sensitive, quantitative, and accessible nanoscale metrology. This tool is shown to address multiple challenges in nanometrology of self-assembled structures, including (i) quantification of surfactant adsorption isotherms on metal oxide surfaces, (ii) determination of self-assembly onset concentration, and (iii) high-throughput readout of drug delivery nanoparticle mass concentration. Finally, the combination of small size and large interfacial area was exploited to design nanoscale formulations for (i) ex vivo delivery to human neutrophils, a significant element of the innate immune system and (ii) targeted delivery of therapeutics to the asthmatic lung.
by Paul W. Bisso.
Ph. D.

Thesis: S.M., Massachusetts Institute of Technology, Department of Mechanical Engineering, 2015.
Cataloged from PDF version of thesis.
Includes bibliographical references (pages 121-124).
Skin imaging is a powerful, noninvasive method used with potential to aid in the diagnosis of various dermatological diseases and assess overall skin health. This thesis discusses imaging platforms that were developed to aid in studying skin features and characteristics at different time and length scales to characterize and monitor skin. Two applications are considered: (1) using natural skin features as a position encoding system and an aid for volume reconstruction of ultrasound imaging and (2) studying natural skin feature evolution or stability over time to aid in assessing skin health. A 5-axis, rigid translational scanning system was developed to capture images at specific locations and to validate skin based body registration algorithms. We show that natural skin features could be used to perform ultrasound based reconstruction accurate to 0.06 mm. A portable, handheld scanning device was designed to study skin characteristics at different time and length scales. With this imaging platform, we analyze skin features at different length scales: [mu]m (for microreliefs), mm (for moles and pores), and cm (for distances between microreliefs and other features). Preliminary algorithms are used to automatically identify microreliefs. Further work in image processing is required to assess skin variation using these images.
by Ina Annesha Kundu.
S.M.

Les objets perceptuels sont les unités élémentaires utilisées par le cerveau pour construire une représentation interne du monde a partir de signaux physiques, comme la lumière ou les ondes sonores. Alors que ces signaux sont d'abord traduit, par les récepteurs dans les organes périphériques, en signaux neuronaux, l'émergence d'objets perceptuels nécessite un traitement intensif dans le système nerveux central qui n'est pas encore entièrement connu. Il est intéressant de noter que les progrès récents de deep learning montrent qu'une séries d'opérations non linéaires et linéaires est très efficace pour catégoriser les objets perceptuels visuels et auditifs de la même manière que les humains. En revanche, la plupart des connaissances actuelles sur le système auditif se concentrent sur les transformations linéaires. Afin de comprendre la contribution des non-linéarités du système auditif à la perception, nous avons étudié l'encodage des sons avec une intensité croissante et une intensité décroissante dans le cortex auditif de la souris. Ces deux sons sont perçus avec une importance inégale malgré le fait qu'ils ont la même énergie physique et le même contenu spectral, un phénomène incompatible avec le traitement linéaire. En enregistrant l'activité de grandes populations corticales pour les sons montants et descendants, nous avons constaté que le cortex les encode avec des populations distinctes qui détectent des caractéristiques non linéaires, ce qui explique l'asymétrie perceptuelle. Nous avons également montré que, dans les modèles de reinforcement learning, la quantité d'activité neuronale déclenchée par un son impacte la vitesse et la stratégie d'apprentissage. Des effets très similaires ont été observés dans plusieurs taches de discrimination ou les sons provoquaient des réponses neuronales de différentes intensités. Ceci établit que les non-linéarités du système auditif ont un impact sur la perception et le comportement. Pour mieux identifier les non-linéarités qui influencent le codage des sons, nous avons ensuite enregistré l'activité d'environ 60 000 neurones échantillonnant toute la superficie du cortex auditif. Au-delà de l'organisation tonotopique à fine échelle découverte avec cet ensemble de données, nous avons identifié et quantifié 7 non-linéarités. Il est aussi intéressant de constater que différentes non-linéarités peuvent interagir entre elles d'une manière non triviale. La connaissance de ces interactions est importante pour affiner le modèle de traitement auditif. Enfin, nous nous sommes demandé si les processus non linéaires sont également importants pour l'intégration multisensorielle. Nous avons mesuré, par imagerie calcique, comment les images et les sons se combinent dans le cortex visuel et auditif. Nous n'avons trouvé aucune modulation du cortex auditif (L2/3) en réponse à des stimuli visuels. Nous avons observé que les entrées du cortex auditif dans le cortex visuel affectent les réponses visuelles concomitantes à un son. Nous avons constaté que les projections du cortex auditif au cortex visuel encode de préférence une caractéristique non linéaire particulière : l'apparition soudaine de sons fort. Par conséquent, l'activité du cortex visuel pour une image et un son fort est plus élevée que pour l'image seule ou combinée à un son faible. Ce résultat suggère que les sons forts sont pertinents du point de vue de comportement multisensoriel, peut-être pour indiquer la présence de nouveaux objets dans le champ visuel, ce qui pourrait représenter des menaces potentielles. En conclusion, nos résultats montrent que les non-linéarités sont omniprésentes dans le traitement du son par le cerveau et jouent également un rôle dans l'intégration de l'information auditive avec l'information visuelle. Il est non seulement crucial de tenir compte de ces non-linéarités pour comprendre comment se forment les représentations perceptuelles, mais aussi pour prédire l'impact de ces représentations sur le comportement
Perceptual objects are the elementary units used by the brain to construct an inner world representation of the environment from multiple physical sources, like light or sound waves. While the physical signals are first encoded by receptors in peripheral organs into neuroelectric signals, the emergence of perceptual object require extensive processing in the central nervous system which is not yet fully characterized. Interestingly, recent advances in deep learning shows that implementing series of nonlinear and linear operations is a very efficient way to create models that categorize visual and auditory perceptual objects similarly to humans. In contrast, most of the current knowledge about the auditory system concentrates on linear transformations. In order to establish a clear example of the contribution of auditory system nonlinearities to perception, we studied the encoding of sounds with an increasing intensity (up ramps) and a decreasing intensity (down ramps) in the mouse auditory cortex. Two behavioral tasks showed evidence that these two sounds are perceived with unequal salience despite carrying the same physical energy and spectral content, a phenomenon incompatible with linear processing. Recording the activity of large cortical populations for up- and down-ramping sounds, we found that cortex encodes them into distinct sets of non-linear features, and that asymmetric feature selection explained the perceptual asymmetry. To complement these results, we also showed that, in reinforcement learning models, the amount of neural activity triggered by a stimulus (e.g. a sound) impacts learning speed and strategy. Interestingly very similar effects were observed in sound discrimination behavior and could be explain by the amount of cortical activity triggered by the discriminated sounds. This altogether establishes that auditory system nonlinearities have an impact on perception and behavior. To more extensively identify the nonlinearities that influence sounds encoding, we then recorded the activity of around 60,000 neurons sampling the entire horizontal extent of auditory cortex. Beyond the fine scale tonotopic organization uncovered with this dataset, we identified and quantified 7 nonlinearities. We found interestingly that different nonlinearities can interact with each other in a non-trivial manner. The knowledge of these interactions carry good promises to refine auditory processing model. Finally, we wondered if the nonlinear processes are also important for multisensory integration. We measured how visual inputs and sounds combine in the visual and auditory cortex using calcium imaging in mice. We found no modulation of supragranular auditory cortex in response to visual stimuli, as observed in previous others studies. We observed that auditory cortex inputs to visual cortex affect visual responses concomitant to a sound. Interestingly, we found that auditory cortex projections to visual cortex preferentially channel activity from neurons encoding a particular non-linear feature: the loud onset of sudden sounds. As a result, visual cortex activity for an image combined with a loud sound is higher than for the image alone or combine with a quiet sound. Moreover, this boosting effect is highly nonlinear. This result suggests that loud sound onsets are behaviorally relevant in the visual system, possibly to indicate the presence of a new perceptual objects in the visual field, which could represent potential threats. As a conclusion, our results show that nonlinearities are ubiquitous in sound processing by the brain and also play a role in the integration of auditory information with visual information. In addition, it is not only crucial to account for these nonlinearities to understand how perceptual representations are formed but also to predict how these representations impact behavior

Дисертаційна робота присвячена вирішенню актуальної науково-практичної проблеми в телекомунікаційних системах, а саме підвищення пропускної здатності каналу передачі семантичних мовних даних за рахунок ефективного їх кодування, тобто формулюється питання підвищення ефективності семантичного кодування, а саме – з якою мінімальною швидкістю можливо кодувати семантичні ознаки мовних сигналів із заданою ймовірністю безпомилкового їх розпізнавання? Саме на це питання буде дана відповідь у даному науковому дослідженні, що є актуальною науково-технічною задачею враховуючи зростаючу тенденцію дистанційної взаємодії людей і роботизованої техніки за допомогою мови, де безпомилковість функціонування даного типу систем безпосередньо залежить від ефективності семантичного кодування мовних сигналів. У роботі досліджено відомий метод підвищення ефективності семантичного кодування мовних сигналів на основі мел-частотних кепстральних коефіцієнтів, який полягає в знаходженні середніх значень коефіцієнтів дискретного косинусного перетворення прологарифмованої енергії спектра дискретного перетворення Фур'є обробленого трикутним фільтром в мел-шкалі. Проблема полягає в тому, що представлений метод семантичного кодування мовних сигналів на основі мел-частотних кепстральних коефіцієнтів не дотримується умови адаптивності, тому було сформульовано основну наукову гіпотезу дослідження, яка полягає в тому що підвищити ефективність семантичного кодування мовних сигналів можливо за рахунок використання адаптивного емпіричного вейвлет-перетворення з подальшим застосуванням спектрального аналізу Гільберта. Під ефективністю кодування розуміється зниження швидкості передачі інформації із заданою ймовірністю безпомилкового розпізнавання семантичних ознак мовних сигналів, що дозволить значно знизити необхідну смугу пропускання, тим самим підвищуючи пропускну здатність каналу зв'язку. У процесі доведення сформульованої наукової гіпотези дослідження були отримані наступні результати: 1) вперше розроблено метод семантичного кодування мовних сигналів на основі емпіричного вейвлетперетворення, який відрізняється від існуючих методів побудовою множини адаптивних смугових вейвлет-фільтрів Мейера з подальшим застосуванням спектрального аналізу Гільберта для знаходження миттєвих амплітуд і частот функцій внутрішніх емпіричних мод, що дозволить визначити семантичні ознаки мовних сигналів та підвищити ефективність їх кодування; 2) вперше запропоновано використовувати метод адаптивного емпіричного вейвлет-перетворення в задачах кратномасштабного аналізу та семантичного кодування мовних сигналів, що дозволить підвищити ефективність спектрального аналізу за рахунок розкладання високочастотного мовного коливання на його низькочастотні складові, а саме внутрішні емпіричні моди; 3) отримав подальший розвиток метод семантичного кодування мовних сигналів на основі мел-частотних кепстральних коефіцієнтів, але з використанням базових принципів адаптивного спектрального аналізу за допомогою емпіричного вейвлет-перетворення, що підвищує ефективність даного методу.
The thesis is devoted to the solution of the actual scientific and practical problem in telecommunication systems, namely increasing the bandwidth of the semantic speech data transmission channel due to their efficient coding, that is the question of increasing the efficiency of semantic coding is formulated, namely – at what minimum speed it is possible to encode semantic features of speech signals with the set probability of their error-free recognition? It is on this question will be answered in this research, which is an urgent scientific and technical task given the growing trend of remote human interaction and robotic technology through speech, where the accurateness of this type of system directly depends on the effectiveness of semantic coding of speech signals. In the thesis the well-known method of increasing the efficiency of semantic coding of speech signals based on mel-frequency cepstral coefficients is investigated, which consists in finding the average values of the coefficients of the discrete cosine transformation of the prologarithmic energy of the spectrum of the discrete Fourier transform treated by a triangular filter in the mel-scale. The problem is that the presented method of semantic coding of speech signals based on mel-frequency cepstral coefficients does not meet the condition of adaptability, therefore the main scientific hypothesis of the study was formulated, which is that to increase the efficiency of semantic coding of speech signals is possible through the use of adaptive empirical wavelet transform followed by the use of Hilbert spectral analysis. Coding efficiency means a decrease in the rate of information transmission with a given probability of error-free recognition of semantic features of speech signals, which will significantly reduce the required passband, thereby increasing the bandwidth of the communication channel. In the process of proving the formulated scientific hypothesis of the study, the following results were obtained: 1) the first time the method of semantic coding of speech signals based on empirical wavelet transform is developed, which differs from existing methods by constructing a sets of adaptive bandpass wavelet-filters Meyer followed by the use of Hilbert spectral analysis for finding instantaneous amplitudes and frequencies of the functions of internal empirical modes, which will determine the semantic features of speech signals and increase the efficiency of their coding; 2) the first time it is proposed to use the method of adaptive empirical wavelet transform in problems of multiscale analysis and semantic coding of speech signals, which will increase the efficiency of spectral analysis due to the decomposition of high-frequency speech oscillations into its low-frequency components, namely internal empirical modes; 3) received further development the method of semantic coding of speech signals based on mel-frequency cepstral coefficients, but using the basic principles of adaptive spectral analysis with the application empirical wavelet transform, which increases the efficiency of this method. Conducted experimental research in the software environment MATLAB R2020b showed, that the developed method of semantic coding of speech signals based on empirical wavelet transform allows you to reduce the encoding speed from 320 to 192 bit/s and the required passband from 40 to 24 Hz with a probability of error-free recognition of about 0.96 (96%) and a signal-to-noise ratio of 48 dB, according to which its efficiency increases 1.6 times in contrast to the existing method. The results obtained in the thesis can be used to build systems for remote interaction of people and robotic equipment using speech technologies, such as speech recognition and synthesis, voice control of technical objects, low-speed encoding of speech information, voice translation from foreign languages, etc.

碩士
國立臺灣科技大學
資訊工程系
103
With the development of technology, Security of information becomes an important issue. Biometrics recognition is popular in recent years, Biometrics is of considerable research interest in recent years, it use humans face、iris、voice、fingerprint recognition, and the biometrics recognition mainstream become vein recognition. This paper is set a recognition system by using finger vein. The characteristics of finger vein are small and portability. Nowadays, the issue we have to discussion is recognize immediately with big database. Although many recognition algorithms have the high recognition rate, but they always cost too much time. Our proposed encode the new feature by using SURF information. The method can reduce the execution time, decrease the dimensions of vector and keep the feature of the vein. We calculate two distances to matching two images. First, computing the feature of second and third level by Hamming distance, it can filter the false candidates quickly. Second, SURF feature using Euclidean distance to determine each pattern vector matching or not. Using this method can retain the original characteristics, and have efficient recognition in the big databases.

Yes
Human memory is content addressable—i.e., contents of the memory can be accessed using partial information about the bound features of a stored item. In this study, we used a cross-feature cuing technique to examine how the human visual system encodes, binds, and retains information about multiple stimulus features within a set of moving objects. We sought to characterize the roles of three different features (position, color, and direction of motion, the latter two of which are processed preferentially within the ventral and dorsal visual streams, respectively) in the construction and maintenance of object representations. We investigated the extent to which these features are bound together across the following processing stages: during stimulus encoding, sensory (iconic) memory, and visual shortterm memory. Whereas all features examined here can serve as cues for addressing content, their effectiveness shows asymmetries and varies according to cue–report pairings and the stage of information processing and storage. Position-based indexing theories predict that position should be more effective as a cue compared to other features. While we found a privileged role for position as a cue at the stimulus-encoding stage, position was not the privileged cue at the sensory and visual short-term memory stages. Instead, the pattern that emerged from our findings is one that mirrors the parallel processing streams in the visual system. This stream-specific binding and cuing effectiveness manifests itself in all three stages of information processing examined here. Finally, we find that the Leaky Flask model proposed in our previous study is applicable to all three features.

碩士
國立中興大學
生物科技學研究所
106
Predicting their attributes is an essential task in computational biology for the advancement of the proteomics. However, the existing methods did not consider the integration of heterogeneous coding and the accuracy of subunit categories with low data number. To end this, we proposed a predictive tool which can predicting more than 12 subunit protein oligomers, QUATgo. At the same time, three kinds of sequence coding were used, including dipeptide composition which was first time using to predict protein quaternary structural attributes, protein half-life characteristics and we modified the coding method of the Functional Domain Composition which proposed by the predecessors to solve the problem of large feature vectors. QUATgo solves the problem of insufficient data in a single subunit using a two-stage architecture and uses 10 times cross-validation to test the predictive accuracy of the classifier, the first-stage prediction model uses a random forest algorithm to generate sixteen homologous, heterologous oligomers and monomer respectively. The accuracy of the first-stage classifier is 63.4%. However, the number of training data of the hetero-10mer is insufficient so the training data of the hetero-10mer and the hetero-more than 12mer is regarded as the same category X. If the result of the first stage classifier is class X the sequence will sent to second stage classifier which was constructed with support vector machines, and can the prediction result of the hetero-10mer and hetero-more than 12mer with an accuracy of 97.5%, QUATgo will eventually have 61.4% cross-validation accuracy and 63.4% independent test accuracy. In case study, QUATgo can accurately predicts the variable complex structure of the MERS-CoV ectodomains.

Iris Recognition is one of the most important biometric recognition systems to authenticate people. Iris recognition works by matching the iris patterns of two individuals. In this paper an efficient iris recognition system is proposed. The process used herein concentrates on image segmentation, normalization, feature extraction and encoding, and verification. Segmentation is done with the help of a great mix of Integro-Differential operator and Hough transformation. This approach provides you with better noise reduction and better iris deduction. Normalization of iris region is performed using the rubber sheet model of Daugman. Next, the properties of Gabor Filters are used to extract the features from the iris. Finally, matching of two templates or two iriscodes is performed by using the hamming distance formula to decide among imposters and genuine ones. The principle of operation used behind iris recognition system is the failure of a test of statistical independence and distinctiveness of the iris texture. This system gives better results than Libor Masek’s system, in terms of efficiency and in terms of time complexity. The proposed system gives better results than Libor Masek’s system, in terms of efficiency and time complexity. The proposed system performed more than 52900 comparisons for authentication on a set of 230 eye images. These images were taken from CASIA-IrisV4-Sync database.

碩士
國立臺北科技大學
電機工程系
106
In this thesis, we investigate principal components analysis (PCA) on both well-known speech features simultaneously at the front-end of the distributed speech recognition (DSR) systems for discarding insignificant feature dimensions. One speech feature is the Mel-frequency cepstral coefficients (MFCC) after the cepstrum mean and variance normalization (CMVN) preprocessing, the other feature is the line spectral frequencies (LSF). After removal of less important features, we apply the singular value decomposition (SVD) to encode the significant speech features for further reducing the transmission bandwidth. At the back-end we employ the histogram equalization (HEQ) method on the decoded speech features and their first and second delta counterparts. In the experiments we use the Aurora-2 database and a basic front-end of the European Telecommunication Standards Institute (ETSI) distributed speech recognition system for evaluation. The experimental results show that the proposed SVD method (9 frames/group) can promote the word accuracy by 14.8 % in the clean condition and 2.65 % in the multi condition, respectively, as compared to the ETSI baseline system. In terms of transmission bitrate, the proposed SVD method (9 frames/group) can perform an average 12.4 bits/frame reduction (approximately 31.79%) than that of the full frame rate (FFR) ETSI baseline system.

碩士
國立臺北科技大學
電子工程系
106
High Efficiency Video Coding (HEVC) is a novel video coding standard. It inherits and improves the previous standard H.264/AVC video coding architecture. With the same video quality, it has better compression efficiency. HEVC applies many architectures and technologies, such as recursive architecture constituted by Coding Unit (CU), Prediction Unit (PU), Transformation Unit (TU), Block Merging. It adapts technologies such as Sample Adaptive Offset (SAO) and supports higher resolution video. Although the performance is increased, each unit from both intra-frame and inter-frame predictions must undergo a large number of computations of Rate-Distortion Cost (RDC) to find the best mode of encoding. Therefore, the complexity in the prediction is rather high, and the time cost is also increased when compared with that in H.264/AVC. In order to reduce the amount of computational complexity required for inter-frame prediction, we propose a fast mode decision algorithm for inter-picture prediction. By analyzing the prediction mode of neighboring blocks, the complexity of the current block is predicted, and those less likely prediction modes will be pruned. Experimental results show that a very good trade-off on the image quality and the coding time can be obtained with the proposed algorithm.

With the volume of data generated in today's internet-of-things, learning algorithms to extract and understand the underlying relations between the various attributes of data have gained momentum. This thesis is focused on learning algorithms to extract meaningful relations from the data using both unsupervised and supervised learning algorithms. Vector quantization techniques are popularly used for applications in contextual data clustering, data visualization and high-dimensional data exploration. Existing vector quantization techniques, such as, the K-means and its variants and those derived from the self-organizing maps consider the input data vector as a whole without prioritizing over individual coordinates. Motivated by applications requiring priorities over data coordinates, we develop a theory for clustering data with different priorities over the coordinates called the data-dependent priority-based soft vector quantization. Based on the input data distribution, the priorities over the data coordinates are learnt by estimating the marginal distributions over each coordinate. The number of neurons approximating each coordinate based on the priority are determined through a reinforcement learning algorithm. Analysis on the convergence of the proposed algorithm and the probability of misclassification are presented along with simulation results on various data sets. Self-organizing maps (SOM) are popularly used for applications in learning features, vector quantization, and recalling spatial input patterns. The adaptation rule in SOMs is based on the Euclidean distance between the input vector and the neuronal weight vector along with a neighborhood function that brings in topological arrangement of the neurons in the output space. It is capable of learning the spatial correlations among the data but fails to capture temporal correlations present in a sequence of inputs. We formulate a potential function based on a spatio-temporal metric and create hierarchical vector quantization feature maps by embedding memory structures similar to long short-term memories across the feature maps to learn the spatio-temporal correlations in the data across clusters. Error correction codes such as low density parity check codes are popularly used to enhance the performance of digital communication systems. The current decoding framework relies on exchanging beliefs over a Tanner graph, which the encoder and decoder are aware of. However, this information may not be available readily, for example, in covert communication. The main idea is to build a neural network to learn the encoder mappings in the absence of knowledge of the Tanner graph. We propose a scheme to learn the mappings using the back-propagation algorithm. We investigate into the choice of different cost functions and the number of hidden neurons for learning the encoding function. The proposed scheme is capable of learning the parity check equations over a binary field towards identifying the validity of a codeword. Simulation results over synthetic data show that our algorithm is indeed capable of learning the encoder mappings. We also propose an approach to identify noisy codes using uncertainty estimation and to decode them using autoencoders. In the next work, we consider the convolutional neural networks which are widely used in natural language processing, video analysis, and image recognition. However, the popularly used max-pooling layer discards most of the data, which is a drawback in applications, such as, prediction of video frames. We propose an adaptive prediction and classification network based on a data-dependent pooling architecture. We formulate a combined cost function for minimizing the prediction and classification errors. We also detect the presence of an unseen class during testing for digit prediction in videos.

碩士
國立高雄大學
電機工程學系碩士班
101
For image clustering, homogeneous and meaningful image pixels with specific features are clustered. The homogeneity of various features is usually calculated by the Euclidean distances among features. For features that have continuous variations, such as color, luminance, saturation, gradient of intensity, distance-based clustering can give effective results. When textures are used as features for clustering, an encoding scheme that describes the variations of textures in terms of distances can produce effective clustering results. This study proposes a genetic-based encoding method to deal with the abovementioned problem where the local binary pattern (LBP) is employed as the texture for clustering. The genetic algorithm (GA) is used to implement the optimal encoding scheme of LBP-based textures. In the encoding scheme, similar LBP-textures are required to have shorter distances, and vice versa. The GA process is separated into two stages. The first stage arranges the locations for all LBP patterns so that they can have continuous variations. The second stage assigns each LBP pattern a unique integer in a manner that similar (dissimilar) patterns have short (long) distances in the same Euclidean scale. A fitness function describing these requirements is defined. In this study, fuzzy c-means is used as the clustering method. Various encoding methods are compared with the proposed method. From the experimental results, the genetic-based encoding method finds a feasible set of encodes for LBP-based textures and improves the quality of image clustering. Some images are tested and the results are analyzed.

Orientation – the process by which animals determine their position in an environment – can be accomplished by using the visually distinct properties of objects or surfaces, known as features (i.e., colour or pattern) or the relationship among objects and surfaces, known as geometry (i.e., wall length or angular information). Although features have been shown to facilitate the encoding of geometry, little is known as to whether restricting one’s viewpoint to include fewer features will still facilitate the encoding of geometry. During this experiment, men and women were trained to search near either an acute or an obtuse corner of a virtual parallelogram-shaped room that contained either three or four discrete and distinctive features. Participants were subsequently tested for their encoding of wall length and angles when the cues were presented in isolation, together, or in conflict. Results showed that the number of features present during training did not influence the encoding of geometry. However, the discrete and distinctive properties of the features overshadowed the encoding of angles by women as well as by participants who were trained with the obtuse corner. Although some groups of participants did not encode angular information when this was the only available geometric cue, all groups weighed angles more heavily than wall length when the cues provided conflicting information. This result suggests that one type of geometric cue (i.e., wall length) can facilitate the encoding of another (i.e., angles).

Confrontation with signal non-stationarity is a rule rather than an exception in the analysis of natural signals, such as speech, animal vocalization, music, bio-medical, atmospheric, ans seismic signals. Interestingly, our auditory system analyzes signal non-stationarity to trigger our perception. It does this with a performance which is unparalleled when compared to any man-made sound analyzer. Non-stationary signal analysis is a fairly challenging problem in the expanse of signal processing. Conventional approaches to analyze non-stationary signals are based on short-time quasi- stationary assumptions. Typically, short-time signal segments are analyzed using one of several transforms, such as Fourier, chirplets, and wavelets, with a predefined basis. However, the quasi-stationary assumption is known to be a serious limitation in recognizing fine temporal and spectral variations in natural signals. An accurate analysis of embedded variations can provide for more insightful understanding of natural signals. Motivated from the sensory mechanisms associated with the peripheral auditory system, this thesis proposes an alternate approach to analyze non-stationary signals. The approach builds on the intuition (and findings from auditory neuroscience literature) that a sequence of zero-crossings (ZCs) of a sine -wave provides its frequency information. Building over this, we hypothesize that sampling an arbitrary signal at some signal specific time instants, instead of uniform Nyquist-rate sampling, can obtain a compact and informative dataset for representation of the signal. The information-richness of the dataset can be quantified by the accuracy to characterize the time-varying attributes of the signal using the sample dataset. We systematically analyze this hypothesis for synthetic signals modeled by time-varying sinusoids and their additive mixtures. A restricted but rich class of non-stationary signals can be modeled using time-varying sinusoids. These sinusoids are characterized by their instantaneous-amplitude (IA) and instantaneous -frequency (IF) variations. It is shown that using ZCs of the signal and its higher-order derivatives, referred to as higher-order ZCs (HoZCs), we can obtain an accurate estimate of IA and IF variations of the sinusoids contained in the signal. The estimation is verified on synthetic signals and natural signal recordings of vocals and birdsong. On comparison of the approach with empirical mode decomposition, a popular technique for non-stationary signal analysis, and we show that the proposed approach has both improved precision and resolution. Building on the above finding on information-richness in the HoZCs instant, we evaluate signal reconstruction using this dataset. The sampling density of this dataset is time-varying in a manner adapting to the temporally evolving spectral content of the signal. Reconstruction is evaluated for speech and audio signals. It is found that for the same number of captured samples, HoZCs corresponding to the first derivative of the signal (extrema samples) provide maximum information compared to other derivatives. This is found to be true even in a comparison of signals reconstructed from an equal number of randomly sampled measurements. Based on these ideas we develop an analysis-modification-synthesis technique for a purely non-stationary modeling of speech signals. This is unlike the existing quasi-stationary analysis techniques. Instead, we propose to model the time- varying quasi -harmonic nature of speech signals. The proposed technique is not constrained by signal duration which helps to avoid blocking artifacts, and at the same time also provides fine temporal resolution of the time-varying attributes. The objective and subjective evaluations show that the technique has better naturalness post modification. It allows controlled modification of speech signals, and can provide for synthesizing novel speech stimuli for probing perception.