Görsel algıya ilişkin bir korteks modeli
A cortex model of visual cognition
- Tez No: 540243
- Danışmanlar: PROF. DR. NESLİHAN SERAP ŞENGÖR
- Tez Türü: Yüksek Lisans
- Konular: Elektrik ve Elektronik Mühendisliği, Nöroloji, Electrical and Electronics Engineering, Neurology
- Anahtar Kelimeler: Belirtilmemiş.
- Yıl: 2018
- Dil: Türkçe
- Üniversite: İstanbul Teknik Üniversitesi
- Enstitü: Fen Bilimleri Enstitüsü
- Ana Bilim Dalı: Elektronik ve Haberleşme Mühendisliği Ana Bilim Dalı
- Bilim Dalı: Elektronik Mühendisliği Bilim Dalı
- Sayfa Sayısı: 87
Özet
Görsel algı, beyinde incelenen en etraflıca çalışılmış ve ulaşılabilir algısal konulardan biridir. Görsel içerik oldukça kolay manipüle edilebildiğinden, yanlış algılaması ilüzyon adı altında kültürel içerik bile olabilmiştir. Son zamanlarda görsel sanal gerçeklik insanı tamamen sarabilmeye başladığından, görsel algının diğer bilişsel konular ile ilintisini araştırabilmek için farklı imkanlar doğmaktadır. Beyinin içindeki görsel devre, kısa dönem belleği, dikkat, alt katman görsel işleme gibi vasıflar arasında, obje farkındalığı gibi yüksek katman fonksiyonları oluşturacak şekilde bir köprü vasfı görür. Beyindeki belli yerlerin gerek deney ile gerek bir hastalık ile fiziksel olarak zarar görmesinin sonucunda elde edilen bilgiler ışığında fizyoloji, hangi bölgelerin ve sinir demetlerinin hangi işleve katkıda bulunduğu hakkında açıklamalara sahiptir. Bu bağlantıların görsel algının ve dikkatin temel yapı taşlarını sergileyebileceğinin alçak katmandaki açıklaması halen bir araştırma konusudur. İhtiyacımız olan, süregelen etkinlik, çeldirici uyartılara bağışıklık gibi temel özelliklerin sağlanabilmesi için biyolojik akla yatkınlığa sahip bir modelle oluşturulmuş bir benzetim kullanılabilir. Doğrusal olmayan topla ve ateşle nöron modelleri ile bu fenomenlerin bir çoğu basite indirgenmiş halleri ile yakalanabilmektedir. Aynı nöron modellerinin görsel algı ve dikkat için de açıklayıcı güce sahip olma ihtimali bulunmaktadır. Bu çalışmada, bu ihtimal üzerinde durulacak ve uygunluğu teste tabii tutulacaktır.
Özet (Çeviri)
Visual perception is one of the most widely studied and accessible cognitive subjects in the brain. Since visual sensory stimuli are easily craftable, the misperception of them have been subject to human content such as illusions. It also has been lately possible to encapsulate a person within a visual virtual reality, to dive deeper into the endeavor of understanding visual cognition, and how it is connected to other subjects of scientific interest. The visual circuitry within brain exhibits a tight gameplay between short term memory, attention, and low-level visual processing, facilitating higher level functions such as object recognition. Physiology has explanations on how this circuitry exhibits these higher-level functions by singling out zones within the brain and bundles of neurons through losses of functionality by either experimental means or by an intrusive malady. The low-level explanation of how an interconnect of neurons can exhibit the primitives of visual cognition and attention is still under research. Simulations with marginal biological plausibility can be used in order to implement core features that might be required, such as persistence of activity, distractor immunity for cortical networks. Many of these phenomena are exhibited in a simplistic manner with integrate and fire neurons with nonlinearity. There is a possibility that same neuron models might be explanatory for the case of visual attention and perception. In this work, this possibility will be pursued and put to test. In this work, we construct a simulation of the visual cortex that relies on the Brunel and Wang integrate and fire neurons and put them into a trial to whether they can exhibit some higher neuroscientific phenomena, such as backward masking, or biased competition. The neuron is modeled with an electrical model, in which every neuron has a leaking membrane capacitance, individual voltage-dependent non-linear channel conductivities, and uses integrated spike trains in as excitation for channel gating variables, a figure that is based on the ratio of open channels. Several receptors play role on the dynamics, where particularly NMDA, an excitatory receptor that is activated nonlinearly with respect to the membrane potential, which the consensus is that it plays key role in mechanisms of attention within the brain. The experiment uses a group of 1000 neurons that are all modeled as mentioned, and all connected to each other (except self-connections), making up close to 1000000 connections, thus synapses, to compute. 200 of these neurons are interneurons that suppress excitation, which excites only GABA receptors that depolarize, and move the postsynaptic neuron away from excitation. The rest are excitatory neurons, which use AMPA and NMDA receptors that hyperpolarize, thus initiate excitation in the postsynaptic neuron. This interaction is modeled with delta functions which signify the events of spiking in the presynaptic neuron. These delta function are integrated, thus instantaneously increment postsynaptic channel activation variables, and this activity is damped down with a measured time constant of that particular channel. Each activated postsynaptic channel lets through a current that depolarizes the membrane potential. When these excitations accumulate in the postsynaptic neuron such that it results in a membrane potential higher than a threshold value of -50mV, the postsynaptic neuron“fires”and instantaneously polarizes back its membrane to a steady state of -70mV. This is the singular phenomenon of how a spike or an excitation is generated. The experiments include several neuron groups that have been assumed to have undergone a process of Hebbian learning, which has resulted in a static nonhomogenous connectivity between these groups. It is important to emphasize that the learned information, or association is a given trait in this experiment. Execution of a learning process might be a way to extend the scope of this experiment. Inhibitory neurons are connected to every excitatory neuron without differentiating which group it belongs to, because it is necessary for reasons of general stability that every excitatory neuron be repressed when excited. On the other hand, excitatory neurons are separated into 5 groups of 80 (S1, S2, S3, S4 and S5) and a group of 400 (Sn). The groups of 80 each represent a trait, which might be a shape or position, whereas the group of 400 is in place just to keep the excitation present in the population. All weights (that multiply the gating variables thus affect the induced current) of these 1000000 synapses are predetermined representative of this phase of a Hebbian training. The neurons in the same groups (e.g. S1 to S1) that have object representation are connected to each other in an elevated weight. The groups that represent different representations have a diminished weight (e.g. S1 to S2). This way, it becomes advantageous for groups to exhibit recurrent spiking, and when one group is excited the others are effectively inhibited, therefore the overall collective dynamic forms attractors (steady states) in which one of the representative groups are in recurrent spiking. In order to keep the population in the verge of excitation, a noise with a Poisson distribution of 2.4 kHz is supplied to every neuron. This elevates the rate of excitation to a level that only a few incoming spikes is enough to initiate a post-synaptic spike. This frequency of cumulative spontaneous excitation is found to be biologically plausible, shown to be consistent with in-vivo measurements. After the model is constructed, several computational complications are eradicated, and seemingly, a dynamical plausibility is achieved, a test experiment is executed. In this test experiment, a stripped down version of binary behavior is exhibited. When the aforementioned setup is built, when S1 is excited by an elevated noise source, the excitation rate of S1 elevates, and remains elevated after the excitation is lifted. When a distracting excitation floods every single neuron within the setup, this persistent behavior is eliminated, and the system goes back to its initial state. Therefore, the population can exhibit the function of a memory, and the behavior is in line with the literature that focuses on the dynamical aspect of these neurons. To exhibit biased competition, two groups of 1000 neurons with the same arrangement, but with 2 groups of 80 (S1 and S2) and a group of 640 (Sn). One group represents V2, and the other V4 in the visual neural pathway. The V4 group of 1000 has the same arrangement, but the groups are denoted with S1', S2' and Sn'. In this arrangement, individual groups are connected in a similar manner with the previous test experiment, and the groups are connected that favor associated groups over cross-associated groups (e.g. $S1 \rightarrow S1'$ to $S1 \rightarrow S2'$) and favors forward connections to backward connections. (e.g. $S1 \rightarrow S1'$ to $S1' \rightarrow S1$) In this experiment, it is investigated whether a bias in one of the representations (e.g. S1) creates easier excitability in the associated V4 group, S1'. We compared this case with the case where a bias in one of the representations creates easier excitability in the cross-associated V4 group. Although there were single occurrences of induced biases, we failed to create a statistically relevant case for biased competition in this arrangement. When an average is taken over multiple instances of this experiment, these two cases fail to exhibit a difference. In this work, the simulations are executed with Brian2 libraries, which opens up possibilities for further work to be done with similiar experimental arrangements, with similiar underlying biological principals. The work can be downloaded and reproduced by downloading the digital notebook, and the work can be seamlessly improved. Possible improvements might be inclusion of a Hebbian training phase, use of real images with their Gabor representations, or further work in order to reveal top-down biasing with the models that have been introduced and coded in this work.
Benzer Tezler
- Zaman algısı ve işleyen bellek ilişkisinin işlevsel manyetik rezonans görüntüleme ile incelenmesi
Evaluation of the relationship between time perception and working memory with functional magnetic resonance imaging
SERTAÇ ÜSTÜN
- Grafik tasarımda yapısal öğelerin algısı ile sosyo demografik değişkenler arasındaki ilişki Selçuk Üniversitesi örneği
The relation between sociodemographic variables and perception of structural elements in graphic design: The case of Selçuk University
ABDULGANİ ARIKAN
Doktora
Türkçe
2008
Güzel SanatlarSelçuk ÜniversitesiGüzel Sanatlar Eğitimi Ana Bilim Dalı
PROF. DR. ADNAN TEPECİK
- Keşfe yönelen davranışlar: Olanaklılık teorisine bir katkı
Exploration driven behaviors: A contrubiton to the theory of affordance
ÖZGÜN YÜCETÜRK
Yüksek Lisans
Türkçe
2016
Mimarlıkİstanbul Teknik ÜniversitesiMimarlık Ana Bilim Dalı
YRD. DOÇ. SEMA ALAÇAM
- Kentsel tasarım rehberlerinin görsel algı bağlamında araştırılması
Investigation of urban design guidelines in the context of visual perception
İBRAHİM EREN
Yüksek Lisans
Türkçe
2019
Şehircilik ve Bölge PlanlamaYıldız Teknik ÜniversitesiŞehir ve Bölge Planlama Ana Bilim Dalı
DOÇ. DR. NİLGÜN ÇOLPAN ERKAN
- Fuarlarda görsel algılama açısından grafik tasarım sorunları ve bir stand uygulaması
Problems in graphical design in terms of visual perception in fairs and an experiment wtih a stand exhibitions
EROL TURGUT