Real-time semantic scene understanding is a challenging computer vision task for autonomous vehicles. A limited amount of research has been done regarding forestry and off-road scene understanding, as the industry focuses on urban and on-road applications. Studies have shown that Deep Convolutional Neural Network architectures, using parameters trained on large datasets, can be re-trained and customized with smaller off-road datasets, using a method called transfer learning and yield state-of-the-art classification performance. This master’s thesis served as an extension of such existing off-road semantic segmentation studies. The thesis focused on detecting and visualizing the general trade-offs between classification performance, classification time, and the network’s number of available classes. The results showed that the classification performance declined for every class that got added to the network. Misclassification mainly occurred in the class boundary areas, which increased when more classes got added to the network. However, the number of classes did not affect the network’s classification time. Further, there was a nonlinear trade-off between classification time and classification performance. The classification performance improved with an increased number of network layers and a larger data type resolution. However, the layer depth increased the number of calculations and the larger data type resolution required a longer calculation time. The network’s classification performance increased by 0.5% when using a 16-bit data type resolution instead of an 8-bit resolution. But, its classification time considerably worsened as it segmented about 20 camera frames less per second with the larger data type. Also, tests showed that a 101-layered network slightly degraded in classification performance compared to a 50-layered network, which indicated the nonlinearity to the trade-off regarding classification time and classification performance. Moreover, the class constellations considerably impacted the network’s classification performance and continuity. It was essential that the class’s content and objects were visually similar and shared the same features. Mixing visually ambiguous objects into the same class could drop the inference performance by almost 30%. There are several directions for future work, including writing a new and customized source code for the ResNet50 network. A customized and pruned network could enhance both the application’s classification performance and classification speed. Further, procuring a task-specific forestry dataset and transferring weights pre-trained for autonomous navigation instead of generic object segmentation could lead to even better classification performance. / Se filen

The neural net computer simulations which will be presented here are based on the acceptance of a set of assumptions that for the last twenty years have been expressed in the fields of information processing, neurophysiology and cognitive sciences. First of all, neural networks and their dynamical behaviors in terms of attractors is the natural way adopted by the brain to encode information. Any information item to be stored in the neural net should be coded in some way or another in one of the dynamical attractors of the brain and retrieved by stimulating the net so as to trap its dynamics in the desired item's basin of attraction. The second view shared by neural net researchers is to base the learning of the synaptic matrix on a local Hebbian mechanism. The last assumption is the presence of chaos and the benefit gained by its presence. Chaos, although very simply produced, inherently possesses an infinite amount of cyclic regimes that can be exploited for coding information. Moreover, the network randomly wanders around these unstable regimes in a spontaneous way, thus rapidly proposing alternative responses to external stimuli and being able to easily switch from one of these potential attractors to another in response to any coming stimulus. In this thesis, it is shown experimentally that the more information is to be stored in robust cyclic attractors, the more chaos appears as a regime in the back, erratically itinerating among brief appearances of these attractors. Chaos does not appear to be the cause but the consequence of the learning. However, it appears as an helpful consequence that widens the net's encoding capacity. To learn the information to be stored, an unsupervised Hebbian learning algorithm is introduced. By leaving the semantics of the attractors to be associated with the feeding data unprescribed, promising results have been obtained in term of storing capacity.

Submitted by PPG Ci?ncia da Computa??o (ppgcc@pucrs.br) on 2018-06-15T16:40:47Z No. of bitstreams: 1 LEANDRO PEREIRA DA SILVA_DIS.pdf: 16008947 bytes, checksum: 327a925ea56fcca0a86530a0eb3b1637 (MD5) / Approved for entry into archive by Sheila Dias (sheila.dias@pucrs.br) on 2018-06-26T13:25:28Z (GMT) No. of bitstreams: 1 LEANDRO PEREIRA DA SILVA_DIS.pdf: 16008947 bytes, checksum: 327a925ea56fcca0a86530a0eb3b1637 (MD5) / Made available in DSpace on 2018-06-26T13:34:22Z (GMT). No. of bitstreams: 1 LEANDRO PEREIRA DA SILVA_DIS.pdf: 16008947 bytes, checksum: 327a925ea56fcca0a86530a0eb3b1637 (MD5) Previous issue date: 2018-03-27 / Computer vision is the science that aims to give computers the capability of see- ing the world around them. Among its tasks, object recognition intends to classify objects and to identify where each object is in a given image. As objects tend to occur in particular environments, their contextual association can be useful to improve the object recognition task. To address the contextual awareness on object recognition task, the proposed ap- proach performs the identification of the scene context separately from the identification of the object, fusing both information in order to improve the object detection. In order to do so, we propose a novel architecture composed of two convolutional neural networks running in parallel: one for object identification and the other to the identification of the context where the object is located. Finally, the information of the two-streams architecture is concatenated to perform the object classification. The evaluation is performed using PASCAL VOC 2007 and MS COCO public datasets, by comparing the performance of our proposed approach with architectures that do not use the scene context to perform the classification of the ob- jects. Results show that our approach is able to raise in-context object scores, and reduces out-of-context objects scores. / A vis?o computacional ? a ci?ncia que permite fornecer aos computadores a ca- pacidade de verem o mundo em sua volta. Entre as tarefas, o reconhecimento de objetos pretende classificar objetos e identificar a posi??o onde cada objeto est? em uma imagem. Como objetos costumam ocorrer em ambientes particulares, a utiliza??o de seus contex- tos pode ser vantajosa para melhorar a tarefa de reconhecimento de objetos. Para utilizar o contexto na tarefa de reconhecimento de objetos, a abordagem proposta realiza a iden- tifica??o do contexto da cena separadamente da identifica??o do objeto, fundindo ambas informa??es para a melhora da detec??o do objeto. Para tanto, propomos uma nova arquite- tura composta de duas redes neurais convolucionais em paralelo: uma para a identifica??o do objeto e outra para a identifica??o do contexto no qual o objeto est? inserido. Por fim, a informa??o de ambas as redes ? concatenada para realizar a classifica??o do objeto. Ava- liamos a arquitetura proposta com os datasets p?blicos PASCAL VOC 2007 e o MS COCO, comparando o desempenho da abordagem proposta com abordagens que n?o utilizam o contexto. Os resultados mostram que nossa abordagem ? capaz de aumentar a probabili- dade de classifica??o para objetos que est?o em contexto e reduzir para objetos que est?o fora de contexto.

Current approaches to generating rhyming English poetry with a neural network involve constraining output to enforce the condition of rhyme. We investigate whether this approach is necessary, or if recurrent neural networks can learn rhyme patterns on their own. We compile a new dataset of amateur poetry which allows rhyme to be learned without external constraints because of the dataset’s size and high frequency of rhymes. We then evaluate models trained on the new dataset using a novel framework that automatically measures the system’s knowledge of poetic form and generalizability. We find that our trained model is able to generalize the pattern of rhyme, generate rhymes unseen in the training data, and also that the learned word embeddings for rhyming sets of words are linearly separable. Our model generates a couplet which rhymes 68.15% of the time; this is the first time that a recurrent neural network has been shown to generate rhyming poetry a high percentage of the time. Additionally, we show that crowd-source workers can only distinguish between our generated couplets and couplets from our dataset 63.3% of the time, indicating that our model generates poetry with coherency, semantic meaning, and fluency comparable to couplets written by humans. / Graduate

This research employs five predicting variables to implementing the market timing strategy. These five variables are E/P1, E/P2, B/M, CP and GM. The investment performances of market timing under a variety of investment horizons are examined. There are four different forecasting horizons, which are one-month, three-month, six-month, and twelve-month investment horizons. Both the linear approach and artificial neural networks are employed to forecasting the market. The artificial neural network is employed with a view to capture the non-linearity property embedded in the market. The results are summarized as follows. (1) Both the linearity and nonlinear approaches are able to outperform the market. According to the results of Cumby-Modest test, they do have the market timing ability. (2) In the simple regression models, the performance of CP is relatively well compared to those of other variables. (3) The correct prediction rate increases as the investment horizon increases. (4) The performance of the expanding window approach is on average inferior to that of the moving window approach. (5) In the simulations of timing abilities over the period of May, 1991 to December, 1997. The multiple regression models has the best performance for the cases of one-month, three-month, and six-month investment horizons. On the other hand, BP(1) has the best performance for the case of one-year investment horizon. Contents Chapter 1 Introduction ……………………………………… 1 1.1 Background……………………………………………………………. 1 1.2 Motivations and objectives…………………………………………….3 1.3 Thesis organization ………………………………………………….. 4 Chapter 2 Literature Review…………………………………6 2.1 Previous studies on market timing……………………………………. 6 2.2 Predicting variables…………………………………………………… 8 2.3 Artificial Neural Networks……………………………………………10 2.4 Back Propagation Neural Networks…………………………………..11 2.5 Applications of ANNs to financial fields………………….………….12 Chapter 3 Data and Methodology……………………….….15 3.1 Data………………………………………………………………..….15 3.2 Linear approaches to implementing market timing strategy……….…18 3.3 ANNs to implementing market timing strategy…………..…………..23 Chapter 4 Results on Timing Performance……………..…26 4.1 Performance of linear approach………………………………………26 4.2 Performance of ANNs………………………………………………...38 4.3 Performance evaluation……………………………………………….39 Chapter 5 Summary…………………………………………54 5.1 Conclusions……………………………………………………….….54 5.2 Future works…………………………………………………………55 Appendix……………………………………………………..56 References……………………………………………………57 / This research employs five predicting variables to implementing the market timing strategy. These five variables are E/P1, E/P2, B/M, CP and GM. The investment performances of market timing under a variety of investment horizons are examined. There are four different forecasting horizons, which are one-month, three-month, six-month, and twelve-month investment horizons. Both the linear approach and artificial neural networks are employed to forecasting the market. The artificial neural network is employed with a view to capture the non-linearity property embedded in the market. The results are summarized as follows. (1) Both the linearity and nonlinear approaches are able to outperform the market. According to the results of Cumby-Modest test, they do have the market timing ability. (2) In the simple regression models, the performance of CP is relatively well compared to those of other variables. (3) The correct prediction rate increases as the investment horizon increases. (4) The performance of the expanding window approach is on average inferior to that of the moving window approach. (5) In the simulations of timing abilities over the period of May, 1991 to December, 1997. The multiple regression models has the best performance for the cases of one-month, three-month, and six-month investment horizons. On the other hand, BP(1) has the best performance for the case of one-year investment horizon.

Previous works on classification committees have shown that an efficient committee should consist of networks that are not only very accurate, but also diverse. In this work, aiming to explore trade-off between the diversity and accuracy of committee networks, the steps of neural network training, aggregation of the networks into a committee, and elimination of irrelevant input variables are integrated. To accomplish the elimination, an additional term to the Negative correlation learning error function, which forces input weights connected to the irrelevant input variables to decay, is added.

The high costs and political tensions associated with Ballistic Missile Defense Systems (BMDS) has driven much of the testing and evaluation of BMDS to be performed through high fidelity Modeling and Simulation (M&S). In response, the M&S environments have become highly complex, extremely computationally intensive, and far too slow to be of use to systems engineers and high level decision makers. Regression models can be used to map the system characteristics to the metrics of interest, bringing about large quantities of data and allowing for real-time interaction with high-fidelity M&S environments, however the abundance of discontinuities and non-unique solutions makes the application of regression techniques hazardous. Due to these ambiguities, the transfer function from the characteristics to the metrics appears to have multiple solutions for a given set of inputs, which combined with the multiple inputs yielding the same set of outputs, causes troubles in creating a mapping. Due to the abundance of discontinuities, the existence of a neural network mapping from the system attributes to the performance metrics is not guaranteed, and if the mapping does exist, it requires a large amount of data to be for creating a regression model, making regression techniques less suitable to BMDS analysis. By employing Nested Neural Networks (NNNs), intermediate data can be associated with an ambiguous output which can allow for a regression model to be made. The addition of intermediate data incorporates more knowledge of the design space into the analysis. Nested neural networks divide the design space to form a piece-wise continuous function, which allows for the user to incorporate system knowledge into the surrogate modeling process while reducing the size of a data set required to form the regression model. This thesis defines nested neural networks along with methods and techniques for using NNNs to relieve the effects of discontinuities and non-unique solutions. To show the benefit of the approach, these techniques are applies them to a BMDS simulation. Case studies are performed to optimize the system configurations and assess robustness which could not be done without the regression models.

A new method is developed to determine a set of informative and refined interface assertions satisfied by functions that are represented by feed-forward neural networks. Neural networks have often been criticized for their low degree of comprehensibility.It is difficult to have confidence in software components if they have no clear and valid interface description. Precise and understandable interface assertions for a neural network based software component are required for safety critical applications and for theintegration into larger software systems. The interface assertions we are considering are of the form &quote if the input x of the neural network is in a region (alpha symbol) of the input space then the output f(x) of the neural network will be in the region (beta symbol) of the output space &quote and vice versa. We are interested in computing refined interface assertions, which can be viewed as the computation of the strongest pre- and postconditions a feed-forward neural network fulfills. Unions ofpolyhedra (polyhedra are the generalization of convex polygons in higher dimensional spaces) are well suited for describing arbitrary regions of higher dimensional vector spaces. Additionally, polyhedra are closed under affine transformations. Given a feed-forward neural network, our method produces an annotated neural network, where each layer is annotated with a set of valid linear inequality predicates. The main challenges for the computation of these assertions is to compute the solution of a non-linear optimization problem and the projection of a polyhedron onto a lower-dimensional subspace.

With self-driving cars on the horizon, vehicle autonomy and its problems is a hot topic. In this study we are using convolutional neural networks to make a robot car avoid obstacles. The robot car has a monocular camera, and our approach is to use the images taken by the camera as input, and then output a steering command. Using this method the car is to avoid any object in front of it. In order to lower the amount of training data we use models that are pretrained on ImageNet, a large image database containing millions of images. The model are then trained on our own dataset, which contains of images taken directly by the robot car while driving around. The images are then labeled with the steering command used while taking the image. While training we experiment with using different amounts of frozen layers. A frozen layer is a layer that has been pretrained on ImageNet, but are not trained on our dataset. The Xception, MobileNet and VGG16 architectures are tested and compared to each other. We find that a lower amount of frozen layer produces better results, and our best model, which used the Xception architecture, achieved 81.19% accuracy on our test set. During a qualitative test the car avoid collisions 78.57% of the time.

In this thesis the verification of a neural network’s proficiency at labeling ionizing radiation particles from the unsorted output of a timepix3 camera is attempted. Focus is put on labeling single particles in separate data sequences with slightly preprocessed input data. Preprocessing of input data is done to simplify the patterns that should be recognized. Two major choices were available for this project, Elman-network and Jordan-network. A more complicated type was not an option because of the longer time needed to implement them. The network type chosen was Elman because of freedom in context size. The neural network is created and trained with the TensorFlow API in python with labeled data that was not created by hand. The network recognized the length difference between gamma particles and alpha particles. Beta particles were not considered by the network. It is concluded that the Elman-style network is not proficient in labeling the sequences, which were considered short enough and to have simple enough input data. A more modern network type is therefore likely required to solve this problem.