Refine
Year of publication
- 2022 (5) (remove)
Document Type
- Doctoral Thesis (5)
Language
- English (5)
Has Fulltext
- yes (5)
Is part of the Bibliography
- no (5)
Keywords
Institute
- Informatik und Mathematik (3)
- Physik (2)
Machine learning (ML) techniques have evolved rapidly in recent years and have shown impressive capabilities in feature extraction, pattern recognition, and causal inference. There has been an increasing attention to applying ML to medical applications, such as medical diagnosis, drug discovery, personalized medicine, and numerous other medical problems. ML-based methods have the advantage of processing vast amounts of data.
With an ever increasing amount of medical data collection and large, inter-subject variability in the medical data, automated data processing pipelines are very much desirable since it is laborious, expensive, and error-prone to rely solely on human processing. ML methods have the potential to uncover interesting patterns, unravel correlations between complex features, learn patient-specific representations, and make accurate predictions. Motivated by these promising aspects, in this thesis, I present studies where I have implemented deep neural networks for the early diagnosis of epilepsy based on electroencephalography (EEG) data and brain tumor detection based on magnetic resonance spectroscopy (MRS) data.
In the project for early diagnosis of epilepsy, we are dealing with one of the most common neurological disorders, epilepsy, which is characterized by recurrent unprovoked seizures. It can be triggered by a variety of initial brain injuries and manifests itself after a time window which is called the latent period. During this period, a cascade of structural and functional brain alterations takes place leading to an increased seizure susceptibility.
The development and extension of brain tissue capable of generating spontaneous seizures is defined as epileptogenesis (EPG).
Detecting the presence of EPG provides a precious opportunity for targeted early medical interventions and, thus, can slow down or even halt the disease progression. In order to study brain signals in this latent window, animal epilepsy models are used to provide valuable data as it is extremely difficult to obtain this data from human patients. The aim of this study is to discover biomarkers of EPG using animal models and then to find the equivalent and counterparts in human patients' data. However, the EEG features for EPG are not well-understood and there is not a sufficiently large amount of annotated data for ML-based algorithms. To approach this problem, firstly, I utilized the timestamp information of the recorded EEG from an animal epilepsy model where epilepsy is induced by an electrical stimulation. The timestamp serves as a form of weak supervision, i.e., before and after the stimulation. Secondly, I implemented a deep residual neural network and trained it with a binary classification task to distinguish the EEG signals from these two phases. After obtaining a high discriminative ability on the binary classification task, I proposed to divide further the time span after the stimulation for a three-class classification, aiming to detect possible stages of the progression of the latent EPG phase. I have shown that the model can distinguish EEG signals at different stages of EPG with high accuracy and generalization ability. I have also demonstrated that some of the learned features from the network are clinically relevant.
In the task of detecting brain tumors based on MRS data, I first proposed to apply a deep neural network on the MRS data collected from over 400 patients for a binary classification task. To combat the challenge of noisy labeling, I developed a distillation step to filter out relatively ``cleanly'' labeled samples. A mixing-based data augmentation method was also implemented to expand the size of the training set. All the experiments were designed to be conducted with a leave-patient-out scheme to ensure the generalization ability of the model. Averaged across all leave-patient-out cross-validation sets, the proposed method performed on par with human neuroradiologists, while outperforming other baseline methods. I have demonstrated the distillation effect on the MNIST data set with manually-introduced label noise as well as providing visualization of the input influences on the final classification through a class activation map method.
Moreover, I have proposed to aggregate information at the subject level, which could provide more information and insights. This is inspired by the concept of multiple instance learning, where instance-level labels are not required and which is more tolerant to noisy labeling. I have proposed to generate data bags consisting of instances from each patient and also proposed two modules to ensure permutation invariance, i.e., an attention module and a pooling module. I have compared the performance of the network in different cases, i.e., with and without permutation-invariant modules, with and without data augmentation, single-instance-based and multiple-instance-based learning and have shown that neural networks equipped with the proposed attention or pooling modules can outperform human experts.
In the human brain, the incoming light to the retina is transformed into meaningful representations that allow us to interact with the world. In a similar vein, the RGB pixel values are transformed by a deep neural network (DNN) into meaningful representations relevant to solving a computer vision task it was trained for. Therefore, in my research, I aim to reveal insights into the visual representations in the human visual cortex and DNNs solving vision tasks.
In the previous decade, DNNs have emerged as the state-of-the-art models for predicting neural responses in the human and monkey visual cortex. Research has shown that training on a task related to a brain region’s function leads to better predictivity than a randomly initialized network. Based on this observation, we proposed that we can use DNNs trained on different computer vision tasks to identify functional mapping of the human visual cortex.
To validate our proposed idea, we first investigate a brain region occipital place area (OPA) using DNNs trained on scene parsing task and scene classification task. From the previous investigations about OPA’s functions, we knew that it encodes navigational affordances that require spatial information about the scene. Therefore, we hypothesized that OPA’s representation should be closer to a scene parsing model than a scene classification model as the scene parsing task explicitly requires spatial information about the scene. Our results showed that scene parsing models had representation closer to OPA than scene classification models thus validating our approach.
We then selected multiple DNNs performing a wide range of computer vision tasks ranging from low-level tasks such as edge detection, 3D tasks such as surface normals, and semantic tasks such as semantic segmentation. We compared the representations of these DNNs with all the regions in the visual cortex, thus revealing the functional representations of different regions of the visual cortex. Our results highly converged with previous investigations of these brain regions validating the feasibility of the proposed approach in finding functional representations of the human brain. Our results also provided new insights into underinvestigated brain regions that can serve as starting hypotheses and promote further investigation into those brain regions.
We applied the same approach to find representational insights about the DNNs. A DNN usually consists of multiple layers with each layer performing a computation leading to the final layer that performs prediction for a given task. Training on different tasks could lead to very different representations. Therefore, we first investigate at which stage does the representation in DNNs trained on different tasks starts to differ. We further investigate if the DNNs trained on similar tasks lead to similar representations and on dissimilar tasks lead to more dissimilar representations. We selected the same set of DNNs used in the previous work that were trained on the Taskonomy dataset on a diverse range of 2D, 3D and semantic tasks. Then, given a DNN trained on a particular task, we compared the representation of multiple layers to corresponding layers in other DNNs. From this analysis, we aimed to reveal where in the network architecture task-specific representation is prominent. We found that task specificity increases as we go deeper into the DNN architecture and similar tasks start to cluster in groups. We found that the grouping we found using representational similarity was highly correlated with grouping based on transfer learning thus creating an interesting application of the approach to model selection in transfer learning.
During previous works, several new measures were introduced to compare DNN representations. So, we identified the commonalities in different measures and unified different measures into a single framework referred to as duality diagram similarity. This work opens up new possibilities for similarity measures to understand DNN representations. While demonstrating a much higher correlation with transfer learning than previous state-of-the-art measures we extend it to understanding layer-wise representations of models trained on the Imagenet and Places dataset using different tasks and demonstrate its applicability to layer selection for transfer learning.
In all the previous works, we used the task-specific DNN representations to understand the representations in the human visual cortex and other DNNs. We were able to interpret our findings in terms of computer vision tasks such as edge detection, semantic segmentation, depth estimation, etc. however we were not able to map the representations to human interpretable concepts. Therefore in our most recent work, we developed a new method that associates individual artificial neurons with human interpretable concepts.
Overall, the works in this thesis revealed new insights into the representation of the visual cortex and DNNs...
This thesis presents a first-of-its-kind phenomenological framework that formally describes the development of acquired epilepsy and the role of the neuro-immune axis in this development. Formulated as a system of nonlinear differential equations, the model describes the interaction of processes such as neuroinflammation, blood- brain barrier disruption, neuronal death, circuit remodeling, and epileptic seizures. The model allows for the simulation of epilepsy development courses caused by a variety of neurological injuries. The simulation results are in agreement with ex- perimental findings from three distinct animal models of epileptogenesis. Simula- tions capture injury-specific temporal patterns of seizure occurrence, neuroinflam- mation, blood-brain barrier leakage, and progression of neuronal death. In addition, the model provides insights into phenomena related to epileptogenesis such as the emergence of paradoxically long time scales of disease development after injury, the dose-dependence of epileptogenesis features on injury severity, and the variability of clinical outcomes in subjects exposed to identical injury. Moreover, the developed framework allows for the simulation of therapeutic interventions, which provides insights into the injury-specificity of prominent intervention strategies. Thus, the model can be used as an in silico tool for the generation of testable predictions, which may aid pre-clinical research for the development of epilepsy treatments.
Navigating a complex environment is assumed to require stable cortical representations of environmental stimuli. Previous experimental studies, however, show substantial ongoing remodeling at the level of synaptic connections, even under behaviorally and environmentally stable conditions. It remains unclear, how these changes affect sensory representations on the level of neuronal populations during basal conditions and how learning influences these dynamics.
Our approach is a joint effort between the analysis of experimental data and theory. We analyze chronic neuronal population activity data – acquired by out collaborators in Mainz – to describe population activity dynamics during basal dynamics and during learning (fear conditioning). The data analysis is complemented by the analysis of a circuit model investigating the link between a neural network’s activity and changes in its underlying structure.
Using chronic two-photon imaging data recorded in awake mouse auditory cortex, we reproduce previous findings that responses of neuronal populations to short complex sounds typically cluster into a near discrete set of possible responses. This means that different stimuli evoke basically the same response and are thus grouped together into one of a small set of possible response modes. The near discrete set of response modes can be utilized as a sensitive and robust means to detect and track changes in population activity over time. Doing so we find that sound representations are subject to a significant ongoing remodeling across the time span of days under basal conditions. Auditory cued fear conditioning introduces a bias into these ongoing dynamics, resulting in a differential generalization both on the level of neuronal populations and on the behavioral level. This means that sounds that are perceived similar to the conditioned stimulus (CS+) show an increased co-mapping to the same response mode the CS+ is mapped to. This differential generalization is also observed in animal behavior, where sounds similar to the CS+ result in the same freezing behavior as the CS+, whereas dissimilar sounds do not. These observations could provide a potential mechanism of stimulus generalization, which is one of the most common phenomena associated with post-traumatic stress disorder, on the level of neuronal populations.
To investigate how the aforementioned changes in neuronal population activity are linked to changes in the underlying synaptic connectivity, we devised a circuit model of excitatory and inhibitory neurons. We studied this firing rate model to investigate the effect of gradual changes in the network’s connectivity on its activity. Apart from an input dominated uni-stable regime (one response per stimulus independent of the network) and a network dominated uni-stable regime (one response per network independent of the stimulus), we also find a multi-stable regime for strong recurrent connectivity and a high ratio of inhibition to excitation. In this regime the model reproduces properties of neural population activity in mouse auditory cortex, including sparse activity, a broad distribution of firing rates, and clustering of stimuli into a near discrete set of response modes. This clustering in the multi-stable regime means that, not only can identical stimuli evoke different responses, depending on the network’s initial condition, but different stimuli can also evoke the same response.
Applying gradual drift to the network connectivity we find periods of stable responses, interrupted by abrupt transitions altering the stimulus response mapping. We study the mechanism underlying these transitions by analyzing changes in the fixed points of this network model, employing a method to numerically find all the fixed points of the system. We find that such abrupt transitions typically cannot be explained by the mere displacement of existing fixed points, but involve qualitative changes in the fixed point structure in the vicinity of the response trajectory. We conclude that gradual synaptic drift can lead to abrupt transitions in stimulus responses and that qualitative changes in the network’s fixed point topology underlie such transitions.
In summary we find that cortical networks display ongoing representational drift under basal conditions that is biased towards a differential generalization during fear conditioning. A circuit model is able to reproduce key characteristics of auditory cortex, including a clustering of stimulus responses into a near discrete set of response modes. Implementing synaptic drift into this model leads to periods of stable responses interrupted by abrupt transitions towards new responses.
Die vorgelegte Dissertation behandelt den Einfluss homöostatischer Adaption auf die Informationsverarbeitung und Lenrprozesse in neuronalen Systemen. Der Begriff Homöostase bezeichnet die Fähigkeit eines dynamischen Systems, bestimmte interne Variablen durch Regelmechanismen in einem dynamischen Gleichgewicht zu halten. Ein klassisches Beispiel neuronaler Homöostase ist die dynamische Skalierung synaptischer Gewichte, wodurch die Aktivität bzw. Feuerrate einzelner Neuronen im zeitlichen Mittel konstant bleibt. Bei den von uns betrachteten Modellen handelt es sich um eine duale Form der neuronalen Homöostase. Das bedeutet, dass für jedes Neuron zwei interne Parameter an eine intrinsische Variable wie die bereits erwähnte mittlere Aktivität oder das Membranpotential gekoppelt werden. Eine Besonderheit dieser dualen Adaption ist die Tatsache, dass dadurch nicht nur das zeitliche Mittel einer dynamischen Variable, sondern auch die zeitliche Varianz, also die stärke der Fluktuation um den Mittelwert, kontrolliert werden kann. In dieser Arbeit werden zwei neuronale Systeme betrachtet, in der dieser Aspekt zum Tragen kommt.
Das erste behandelte System ist ein sogennantes Echo State Netzwerk, welches unter die Kategorie der rekurrenten Netzwerke fällt. Rekurrente neuronale Netzwerke haben im Allgemeinen die Eigenschaft, dass eine Population von Neuronen synaptische Verbindungen besitzt, die auf die Population selbst projizieren, also rückkoppeln. Rekurrente Netzwerke können somit als autonome (falls keinerlei zusätzliche externe synaptische Verbindungen existieren) oder nicht-autonome dynamische Systeme betrachtet werden, die durch die genannte Rückkopplung komplexe dynamische Eigenschaften besitzen. Abhängig von der Struktur der rekurrenten synaptischen Verbindungen kann beispielsweise Information aus externem Input über einen längeren Zeitraum gespeichert werden. Ebenso können dynamische Fixpunkte oder auch periodische bzw. chaotische Aktivitätsmuster entstehen. Diese dynamische Vielseitigkeit findet sich auch in den im Gehirn omnipräsenten rekurrenten Netzwerken und dient hier z.B. der Verarbeitung sensorischer Information oder der Ausführung von motorischen Bewegungsmustern. Das von uns betrachtete Echo State Netzwerk zeichnet sich dadurch aus, dass rekurrente synaptische Verbindungen zufällig generiert werden und keiner synaptischen Plastizität unterliegen. Verändert werden im Zuge eines Lernprozesses nur Verbindungen, die von diesem sogenannten dynamischen Reservoir auf Output-Neuronen projizieren. Trotz der Tatsache, dass dies den Lernvorgang stark vereinfacht, ist die Fähigkeit des Reservoirs zur Verarbeitung zeitabhängiger Inputs stark von der statistischen Verteilung abhängig, die für die Generierung der rekurrenten Verbindungen verwendet wird. Insbesondere die Varianz bzw. die Skalierung der Gewichte ist hierbei von großer Bedeutung. Ein Maß für diese Skalierung ist der Spektralradius der rekurrenten Gewichtsmatrix.
In vorangegangenen theoretischen Arbeiten wurde gezeigt, dass für das betrachtete System ein Spektralradius nahe unterhalb des kritischen Wertes von 1 zu einer guten Performance führt. Oberhalb dieses Wertes kommt es im autonomen Fall zu chaotischem dynamischen Verhalten, welches sich negativ auf die Informationsverarbeitung auswirkt. Der von uns eingeführte und als Flow Control bezeichnete duale Adaptionsmechanismus zielt nun darauf ab, über eine Skalierung der synaptischen Gewichte den Spektralradius auf den gewünschten Zielwert zu regulieren. Essentiell ist hierbei, dass die verwendete Adaptionsdynamik im Sinne der biologischen Plausibilität nur auf lokale Größen zurückgreift. Dies geschieht im Falle von Flow Control über eine Regulation der im Membranpotential der Zelle auftretenden Fluktuationen. Bei der Evaluierung der Effektivität von Flow Control zeigte sich, dass der Spektralradius sehr präzise kontrolliert werden kann, falls die Aktivitäten der Neuronen in der rekurrenten Population nur schwach korreliert sind. Korrelationen können beispielsweise durch einen zwischen den Neuronen stark synchronisierten externen Input induziert werden, der sich dementsprechend negativ auf die Präzision des Adaptionsmechanismus auswirkt.
Beim Testen des Netzwerks in einem Lernszenario wirkte sich dieser Effekt aber nicht negativ auf die Performance aus: Die optimale Performance wurde unabhängig von der stärke des korrelierten Inputs für einen Spektralradius erreicht, der leicht unter dem kritischen Wert von 1 lag. Dies führt uns zu der Schlussfolgerung, dass Flow Control unabhängig von der Stärke der externen Stimulation in der Lage ist, rekurrente Netze in einen für die Informationsverarbeitung optimalen Arbeitsbereich einzuregeln.
Bei dem zweiten betrachteten Modell handelt es sich um ein Neuronenmodell mit zwei Kompartimenten, welche der spezifischen Anatomie von Pyramidenneuronen im Kortex nachempfunden ist. Während ein basales Kompartiment synaptischen Input zusammenfasst, der in Dendriten nahe des Zellkerns auftritt, repräsentiert das zweite apikale Kompartiment die im Kortex anzutreffende komplexe dendritische Baumstruktur. In früheren Experimenten konnte gezeigt werden, dass eine zeitlich korrelierte Stimulation sowohl im basalen als auch apikalen Kompartiment eine deutlich höhere neuronale Aktivität hervorrufen kann als durch Stimulation nur einer der beiden Kompartimente möglich ist. In unserem Modell können wir zeigen, dass dieser Effekt der Koinzidenz-Detektion es erlaubt, den Input im apikalen Kompartiment als Lernsignal für synaptische Plastizität im basalen Kompartiment zu nutzen. Duale Homöostase kommt auch hier zum Tragen, da diese in beiden Kompartimenten sicherstellt, dass sich der synaptische Input hinsichtlich des zeitlichen Mittels und der Varianz in einem für den Lernprozess benötigten Bereich befindet. Anhand eines Lernszenarios, das aus einer linearen binären Klassifikation besteht, können wir zeigen, dass sich das beschriebene Framework für biologisch plausibles überwachtes Lernen eignet.
Die beiden betrachteten Modelle zeigen beispielhaft die Relevanz dualer Homöostase im Hinblick auf zwei Aspekte. Das ist zum einen die Regulation rekurrenter neuronaler Netze in einen dynamischen Zustand, der für Informationsverarbeitung optimal ist. Der Effekt der Adaption zeigt sich hier also im Verhalten des Netzwerks als Ganzes. Zum anderen kann duale Homöostase, wie im zweiten Modell gezeigt, auch für Plastizitäts- und Lernprozesse auf der Ebene einzelner Neuronen von Bedeutung sein. Während neuronale Homöostase im klassischen Sinn darauf beschränkt ist, Teile des Systems möglichst präzise auf einen gewünschten Mittelwert zu regulieren, konnten wir Anhand der diskutierten Modelle also darlegen, dass eine Kontrolle des Ausmaßes von Fluktuationen ebenfalls Einfluss auf die Funktionalität neuronaler Systeme haben kann.