Motion in the outside world forms one of the primary uses of visual information for many animals. The ability to interpret motion quickly and accurately permits interaction with and response to events in the outside world. While much is known about some aspects of motion perception, there is less agreement about how feature selectivity leading to motion perception is actually formed in the convergent and divergent pathways of the visual system. It is even less clear how these classical understandings of motion processing, often driven by artificial stimuli with little resemblance to the outside world, correspond to responses of neurons when using more natural stimuli. In this thesis, we probe these gaps, first by demonstrating that synchronization within the visual thalamus leads to efficient representations of motion (through tuning properties) in primary visual cortex, exploiting precise timing across populations in a unique manner compared to traditional models. We then create a novel “minimally-natural” stimulus with the appearance of an infinite hallway wallpapered with sinusoidal gratings, to probe how such minimally natural features modulate our predictions of neural responses based upon feature tuning properties. Through encoding and decoding models we find that measuring a restricted tuning parameter space limits our ability to capture all response properties but preserves relevant information for decoding. We finish with an exploration of ethologically relevant natural features, perspective and complex motion, and show that even moderate amounts of each feature within or near the classical V1 receptive field changes the neural response from what classical feature tuning would predict and improves stimulus classification tremendously. Together all of these results indicate that capturing information about motion in the outside world through visual stimuli requires a more advanced model of feature selectivity that incorporates parameters based on more complex spatial relationships.

<p>One of the most important aspects of visual perception is inference of 3D shape from a 2D retinal image of the real world. The existence of several valid mapping functions from object to data makes this inverse problem ill-posed and therefore computationally difficult. In human vision, the retinal image is a 2D projection of the 3D real world. The visual system imposes certain constraints on the family of solutions in order to efficiently solve this inverse problem. This project specifically focuses on the aspect of minimization of standard deviation of all 3D angles (MSDA) for 3D perception. Our goal is to use a Deep Convolutional Neural Network based on biological principles derived from visual area V4 to solve 3D reconstruction using constrained minimization of MSDA. We conduct an experiment with novel shapes with human participants to collect data to test our model.</p>

Negation is a common property of languages, in that there are few languages, if any, that lack means to revert the truth-value of a statement. A challenge to cross-lingual studies of negation lies in the fact that languages encode and use it in different ways. Although this variation has been extensively researched in linguistics, little has been done in automated language processing. In particular, we lack computational models of processing negation that can be generalized across language. We even lack knowledge of what the development of such models would require. These models however exist and can be built by means of existing cross-lingual resources, even when annotated data for a language other than English is not available. This thesis shows this in the context of detecting string-level negation scope, i.e. the set of tokens in a sentence whose meaning is affected by a negation marker (e.g. 'not'). Our contribution has two parts. First, we investigate the scenario where annotated training data is available. We show that Bi-directional Long Short Term Memory (BiLSTM) networks are state-of-the-art models whose features can be generalized across language. We also show that these models suffer from genre effects and that for most of the corpora we have experimented with, high performance is simply an artifact of the annotation styles, where negation scope is often a span of text delimited by punctuation. Second, we investigate the scenario where annotated data is available in only one language, experimenting with model transfer. To test our approach, we first build NEGPAR, a parallel corpus annotated for negation, where pre-existing annotations on English sentences have been edited and extended to Chinese translations. We then show that transferring a model for negation scope detection across languages is possible by means of structured neural models where negation scope is detected on top of a cross-linguistically consistent representation, Universal Dependencies. On the other hand, we found cross-lingual lexical information only to help very little with performance. Finally, error analysis shows that performance is better when a negation marker is in the same dependency substructure as its scope and that some of the phenomena related to negation scope requiring lexical knowledge are still not captured correctly. In the conclusions, we tie up the contributions of this thesis and we point future work towards representing negation scope across languages at the level of logical form as well.

O presente trabalho descreve o projeto de modelos neurais pulsados em tecnologia CMOS. Foram projetados dois modelos: um neurônio baseado em condutâncias e um neurônio do tipo integra e dispara. O primeiro gera impulsos elétricos similares aos potenciais de ação gerados pelo neurônio biológico. Mediante simulação, foram observadas as seguintes características: disparo do impulso quando se atinge a tensão de limiar, hiperpolarização após o potencial de ação, retorno passivo à tensão de repouso, presença de período refratário e relação sigmoide entre a frequência de disparo e a intensidade do estímulo. Da mesma maneira, foi reproduzida a curva mínima duração x amplitude de estímulo típico dos neurônios biológicos. O segundo realiza a codificação de uma grandeza analógica na fase relativa dos impulsos elétricos gerados. Os impulsos gerados pelo circuito estão afastados em relação a um sinal periódico, em um intervalo que apresenta uma dependência logarítmica de uma corrente de entrada. John Hopfield propus esse tipo de codificação para explicar o reconhecimento de padrões com independência de escala, realizado pelo cérebro humano. No decorrer da pesquisa, foi necessário desenvolver algumas expressões analíticas para o projeto de circuitos de baixa frequência em CMOS, não encontradas na literatura estudada. As expressões estão baseadas na equação da corrente do transistor MOS proposta no modelo conhecido como Advanced Compact Mosfet (ACM). O projeto, implementação e testes de um transcondutor linearizado, e os resultados das simulações dos modelos neurais projetados, demonstram a validade das expressões desenvolvidas. / This work describes the design of pulsed neural models in CMOS technology. Two models were designed: a conductance based neuron and an integrate and fire neuron. The first generates electrical impulses similar to action potentials generated by the biological neuron. Through simulation, the following characteristics were observed: pulse trigger after reaching threshold voltage, hyperpolarization after the action potential, passive return to resting potential, presence of refractory period and sigmoid relationship between the firing rate and the stimulus intensity. Likewise, the curve minimal duration vs stimulus amplitude typical of biological neurons was reproduced. The second one performs the encoding of an analog input in the relative phase of electrical impulses. The impulses generated by the circuit are delayed with respect to a reference periodic signal, in a range that has a logarithmic dependence on an input current. John Hopfield proposed this type of encoding to explain the scale independent pattern recognition performed by the human brain. During the research, it was necessary to develop some analytical expressions for the design of low-frequency circuits in CMOS, not found in the literature studied. The expressions are based on the Advanced Compact MOSFET (ACM) model. The design, implementations and testing of a linearized transconductor, and the simulations results of the neural models designed, demonstrate the validity of the expressions developed.

O presente trabalho descreve o projeto de modelos neurais pulsados em tecnologia CMOS. Foram projetados dois modelos: um neurônio baseado em condutâncias e um neurônio do tipo integra e dispara. O primeiro gera impulsos elétricos similares aos potenciais de ação gerados pelo neurônio biológico. Mediante simulação, foram observadas as seguintes características: disparo do impulso quando se atinge a tensão de limiar, hiperpolarização após o potencial de ação, retorno passivo à tensão de repouso, presença de período refratário e relação sigmoide entre a frequência de disparo e a intensidade do estímulo. Da mesma maneira, foi reproduzida a curva mínima duração x amplitude de estímulo típico dos neurônios biológicos. O segundo realiza a codificação de uma grandeza analógica na fase relativa dos impulsos elétricos gerados. Os impulsos gerados pelo circuito estão afastados em relação a um sinal periódico, em um intervalo que apresenta uma dependência logarítmica de uma corrente de entrada. John Hopfield propus esse tipo de codificação para explicar o reconhecimento de padrões com independência de escala, realizado pelo cérebro humano. No decorrer da pesquisa, foi necessário desenvolver algumas expressões analíticas para o projeto de circuitos de baixa frequência em CMOS, não encontradas na literatura estudada. As expressões estão baseadas na equação da corrente do transistor MOS proposta no modelo conhecido como Advanced Compact Mosfet (ACM). O projeto, implementação e testes de um transcondutor linearizado, e os resultados das simulações dos modelos neurais projetados, demonstram a validade das expressões desenvolvidas. / This work describes the design of pulsed neural models in CMOS technology. Two models were designed: a conductance based neuron and an integrate and fire neuron. The first generates electrical impulses similar to action potentials generated by the biological neuron. Through simulation, the following characteristics were observed: pulse trigger after reaching threshold voltage, hyperpolarization after the action potential, passive return to resting potential, presence of refractory period and sigmoid relationship between the firing rate and the stimulus intensity. Likewise, the curve minimal duration vs stimulus amplitude typical of biological neurons was reproduced. The second one performs the encoding of an analog input in the relative phase of electrical impulses. The impulses generated by the circuit are delayed with respect to a reference periodic signal, in a range that has a logarithmic dependence on an input current. John Hopfield proposed this type of encoding to explain the scale independent pattern recognition performed by the human brain. During the research, it was necessary to develop some analytical expressions for the design of low-frequency circuits in CMOS, not found in the literature studied. The expressions are based on the Advanced Compact MOSFET (ACM) model. The design, implementations and testing of a linearized transconductor, and the simulations results of the neural models designed, demonstrate the validity of the expressions developed.

No description available.