<p>Convolutional Neural Networks have come a long way since  AlexNet. Each year the limits of the state of the art are being pushed to new  levels. EfficientNet pushed the performance metrics to a new high and EfficientNetV2 even more so. Even so, architectures for mobile applications can benefit from improved accuracy and reduced model footprint. The classic Inverted Residual block has been the foundation upon which most mobile networks seek to improve. EfficientNet architecture is built using the same Inverted Residual block. In this paper we experiment with Harmonious Bottlenecks in  place of the Inverted Residuals to observe a reduction in the number of parameters and improvement in accuracy. The designed network is then deployed on the NXP i.MX 8M Mini board for Image classification.</p>

Vision based active safety systems have become more frequently occurring in modern vehicles to estimate depth of the objects ahead and for autonomous driving (AD) and advanced driver-assistance systems (ADAS). In this thesis a lightweight deep convolutional neural network performing real-time depth estimation on single monocular images is implemented and evaluated. Many of the vision based automatic brake systems in modern vehicles only detect pre-trained object types such as pedestrians and vehicles. These systems fail to detect general objects such as road debris and roadside obstacles. In stereo vision systems the problem is resolved by calculating a disparity image from the stereo image pair to extract depth information. The distance to an object can also be determined using radar and LiDAR systems. By using this depth information the system performs necessary actions to avoid collisions with objects that are determined to be too close. However, these systems are also more expensive than a regular mono camera system and are therefore not very common in the average consumer car. By implementing robust depth estimation in mono vision systems the benefits from active safety systems could be utilized by a larger segment of the vehicle fleet. This could drastically reduce human error related traffic accidents and possibly save many lives. The network architecture evaluated in this thesis is more lightweight than other CNN architectures previously used for monocular depth estimation. The proposed architecture is therefore preferable to use on computationally lightweight systems. The network solves a supervised regression problem during the training procedure in order to produce a pixel-wise depth estimation map. The network was trained using a sparse ground truth image with spatially incoherent and discontinuous data and output a dense spatially coherent and continuous depth map prediction. The spatially incoherent ground truth posed a problem of discontinuity that was addressed by a masked loss function with regularization. The network was able to predict a dense depth estimation on the KITTI dataset with close to state-of-the-art performance.

<div>The rise of artificial intelligence (AI) hinges on the efficient training of modern deep neural networks (DNNs) for non-convex optimization and uncertainty quantification, which boils down to a non-convex Bayesian learning problem. A standard tool to handle the problem is Langevin Monte Carlo, which proposes to approximate the posterior distribution with theoretical guarantees. However, non-convex Bayesian learning in real big data applications can be arbitrarily slow and often fails to capture the uncertainty or informative modes given a limited time. As a result, advanced techniques are still required.</div><div><br></div><div>In this thesis, we start with the replica exchange Langevin Monte Carlo (also known as parallel tempering), which is a Markov jump process that proposes appropriate swaps between exploration and exploitation to achieve accelerations. However, the na\"ive extension of swaps to big data problems leads to a large bias, and the bias-corrected swaps are required. Such a mechanism leads to few effective swaps and insignificant accelerations. To alleviate this issue, we first propose a control variates method to reduce the variance of noisy energy estimators and show a potential to accelerate the exponential convergence. We also present the population-chain replica exchange and propose a generalized deterministic even-odd scheme to track the non-reversibility and obtain an optimal round trip rate. Further approximations are conducted based on stochastic gradient descents, which yield a user-friendly nature for large-scale uncertainty approximation tasks without much tuning costs. </div><div><br></div><div>In the second part of the thesis, we study scalable dynamic importance sampling algorithms based on stochastic approximation. Traditional dynamic importance sampling algorithms have achieved successes in bioinformatics and statistical physics, however, the lack of scalability has greatly limited their extensions to big data applications. To handle this scalability issue, we resolve the vanishing gradient problem and propose two dynamic importance sampling algorithms based on stochastic gradient Langevin dynamics. Theoretically, we establish the stability condition for the underlying ordinary differential equation (ODE) system and guarantee the asymptotic convergence of the latent variable to the desired fixed point. Interestingly, such a result still holds given non-convex energy landscapes. In addition, we also propose a pleasingly parallel version of such algorithms with interacting latent variables. We show that the interacting algorithm can be theoretically more efficient than the single-chain alternative with an equivalent computational budget.</div>

<p dir="ltr">The genetic perturbations caused by spaceflight on biological systems tend to have a system-wide effect which is often difficult to deconvolute into individual signals with specific points of origin. Single cell multi-omic data can provide a profile of the perturbational effects but does not necessarily indicate the initial point of interference within a network. The objective of this project is to take advantage of large scale and genome-wide perturbational or Perturb-Seq datasets by using them to pre-train a generalist machine learning model that is capable of predicting the effects of unseen perturbations in new data. Perturb-Seq datasets are large libraries of single cell RNA sequencing data collected from CRISPR knock out screens in cell culture. The advent of generative machine learning algorithms, particularly transformers, make it an ideal time to re-assess large scale data libraries in order to grasp cell and even organism-wide genomic expression motifs. By tailoring an algorithm to learn the downstream effects of the genetic perturbations, we present a pre-trained generalist model capable of predicting the effects of multiple perturbations in combination, locating points of origin for perturbation in new datasets, predicting the effects of known perturbations in new datasets, and annotation of large-scale network motifs. We demonstrate the utility of this model by identifying key perturbational signatures in RNA sequencing data from spaceflown biological samples from the NASA Open Science Data Repository.</p>

<p>Synthesized speech may be used for nefarious purposes, such as fraud, spoofing, and misinformation campaigns. We present several speech forensics methods based on deep learning to protect against such attacks. First, we use a convolutional neural network (CNN) and transformers to detect synthesized speech. Then, we investigate closed set and open set speech synthesizer attribution. We use a transformer to attribute a speech signal to its source (i.e., to identify the speech synthesizer that created it). Additionally, we show that our approach separates different known and unknown speech synthesizers in its latent space, even though it has not seen any of the unknown speech synthesizers during training. Next, we explore machine learning for an objective in the aerospace domain.</p> <p><br></p> <p>Compared to conventional ballistic vehicles and cruise vehicles, hypersonic glide vehicles (HGVs) exhibit unprecedented abilities. They travel faster than Mach 5 and maneuver to evade defense systems and hinder prediction of their final destinations. We investigate machine learning for identifying different HGVs and a conic reentry vehicle (CRV) based on their aerodynamic state estimates. We also propose a HGV flight phase prediction method. Inspired by natural language processing (NLP), we model flight phases as “words” and HGV trajectories as “sentences.” Next, we learn a “grammar” from the HGV trajectories that describes their flight phase transition patterns. Given “words” from the initial part of a HGV trajectory and the “grammar”, we predict future “words” in the “sentence” (i.e., future HGV flight phases in the trajectory). We demonstrate that this approach successfully predicts future flight phases for HGV trajectories, especially in scenarios with limited training data. We also show that it can be used in a transfer learning scenario to predict flight phases of HGV trajectories that exhibit new maneuvers and behaviors never seen before during training.</p>

[ES] La presente tesis doctoral propone la arquitectura de un sistema de Geo-visualización Espaciotemporal de actividad delictiva y criminal, para ser aplicada a Sistemas de Comando y Control (C2S) específicamente dentro de sus Sistemas de Información de Comando y Control (C2IS). El sistema de Geo-visualización Espaciotemporal se basa en el análisis masivo de datos reales de actividad delictiva, proporcionado por la Policía Nacional Colombiana (PONAL) y está compuesto por dos aplicaciones diferentes: la primera permite al usuario geo-visualizar espaciotemporalmente de forma dinámica, las concentraciones, tendencias y patrones de movilidad de esta actividad dentro de la extensión de área geográfica y el rango de fechas y horas que se precise, lo cual permite al usuario realizar análisis e interpretaciones y tomar decisiones estratégicas de acción más acertadas; la segunda aplicación permite al usuario geo-visualizar espaciotemporalmente las predicciones de la actividad delictiva en periodos continuos y cortos a modo de tiempo real, esto también dentro de la extensión de área geográfica y el rango de fechas y horas de elección del usuario. Para estas predicciones se usaron técnicas clásicas y técnicas de Machine Learning (incluido el Deep Learning), adecuadas para el pronóstico en multiparalelo de varios pasos de series temporales multivariantes con datos escasos. Las dos aplicaciones del sistema, cuyo desarrollo se muestra en esta tesis, están realizadas con métodos novedosos que permitieron lograr estos objetivos de efectividad a la hora de detectar el volumen y los patrones y tendencias en el desplazamiento de dicha actividad, mejorando así la conciencia situacional, la proyección futura y la agilidad y eficiencia en los procesos de toma de decisiones, particularmente en la gestión de los recursos destinados a la disuasión, prevención y control del delito, lo cual contribuye a los objetivos de ciudad segura y por consiguiente de ciudad inteligente, dentro de arquitecturas de Sistemas de Comando y Control (C2S) como en el caso de los Centros de Comando y Control de Seguridad Ciudadana de la PONAL. / [CA] Aquesta tesi doctoral proposa l'arquitectura d'un sistema de Geo-visualització Espaitemporal d'activitat delictiva i criminal, per ser aplicada a Sistemes de Comandament i Control (C2S) específicament dins dels seus Sistemes d'informació de Comandament i Control (C2IS). El sistema de Geo-visualització Espaitemporal es basa en l'anàlisi massiva de dades reals d'activitat delictiva, proporcionada per la Policia Nacional Colombiana (PONAL) i està composta per dues aplicacions diferents: la primera permet a l'usuari geo-visualitzar espaitemporalment de forma dinàmica, les concentracions, les tendències i els patrons de mobilitat d'aquesta activitat dins de l'extensió d'àrea geogràfica i el rang de dates i hores que calgui, la qual cosa permet a l'usuari fer anàlisis i interpretacions i prendre decisions estratègiques d'acció més encertades; la segona aplicació permet a l'usuari geovisualitzar espaciotemporalment les prediccions de l'activitat delictiva en períodes continus i curts a mode de temps real, això també dins l'extensió d'àrea geogràfica i el rang de dates i hores d'elecció de l'usuari. Per a aquestes prediccions es van usar tècniques clàssiques i tècniques de Machine Learning (inclòs el Deep Learning), adequades per al pronòstic en multiparal·lel de diversos passos de sèries temporals multivariants amb dades escasses. Les dues aplicacions del sistema, el desenvolupament de les quals es mostra en aquesta tesi, estan realitzades amb mètodes nous que van permetre assolir aquests objectius d'efectivitat a l'hora de detectar el volum i els patrons i les tendències en el desplaçament d'aquesta activitat, millorant així la consciència situacional , la projecció futura i l'agilitat i eficiència en els processos de presa de decisions, particularment en la gestió dels recursos destinats a la dissuasió, prevenció i control del delicte, la qual cosa contribueix als objectius de ciutat segura i per tant de ciutat intel·ligent , dins arquitectures de Sistemes de Comandament i Control (C2S) com en el cas dels Centres de Comandament i Control de Seguretat Ciutadana de la PONAL. / [EN] This doctoral thesis proposes the architecture of a Spatiotemporal Geo-visualization system of criminal activity, to be applied to Command and Control Systems (C2S) specifically within their Command and Control Information Systems (C2IS). The Spatiotemporal Geo-visualization system is based on the massive analysis of real data of criminal activity, provided by the Colombian National Police (PONAL) and is made up of two different applications: the first allows the user to dynamically geo-visualize spatiotemporally, the concentrations, trends and patterns of mobility of this activity within the extension of the geographic area and the range of dates and times that are required, which allows the user to carry out analyses and interpretations and make more accurate strategic action decisions; the second application allows the user to spatially visualize the predictions of criminal activity in continuous and short periods like in real time, this also within the extension of the geographic area and the range of dates and times of the user's choice. For these predictions, classical techniques and Machine Learning techniques (including Deep Learning) were used, suitable for multistep multiparallel forecasting of multivariate time series with sparse data. The two applications of the system, whose development is shown in this thesis, are carried out with innovative methods that allowed achieving these effectiveness objectives when detecting the volume and patterns and trends in the movement of said activity, thus improving situational awareness, the future projection and the agility and efficiency in the decision-making processes, particularly in the management of the resources destined to the dissuasion, prevention and control of crime, which contributes to the objectives of a safe city and therefore of a smart city, within architectures of Command and Control Systems (C2S) as in the case of the Citizen Security Command and Control Centers of the PONAL. / Salcedo González, ML. (2023). Arquitectura de un sistema de geo-visualización espacio-temporal de actividad delictiva, basada en el análisis masivo de datos, aplicada a sistemas de información de comando y control (C2IS) [Tesis doctoral]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/192685