Istanbul Technical University


Computational Models

Basal ganglia - cortex - thalamus loop, which is involved in various cognitive processes, plays a significant role in neurological diseases and behavioral disorders. Modeling this loop and the structures involved are one of the research subtopics of SIMMAG.

a- Voluntary Action Selection

In our group, the voluntary action mechanism is modeled with two different approaches. In the first group of studies. the basal ganglia, cortex, thalamus loop is modeled with mass model approach where the behavior of different units are represented as the average behavior of neuronal populations [B1, B9,B11,B20, B21, B23, E1,E4, F1.11, F 1.14, F 1.16, F1.19, F1.29, Berat Denizdurduran Master Tezi]. Other approach is to model this loop as spiking neural networks [A1.1, B6, C1, E7, E8, E9, E10, F1.3, F1.6, F1.7]. We also conducted studies focusing on the relation between these two different approaches [D1, F2.12].
In mass models describing the basal ganglia, cortex, thalamus loop, the abnormal movement patterns seen in Parkinson’s and Huntington’s can be replicated by varying the dopamine neuromodulator parameter. The bifurcation analyses, phase portraits, and domain of attraction graphs are in use to investigate the behaviors obtained via the model.
The effect of the dopamine is also investigated in the studies where the loop is modeled as spiking neural networks. Since these models are relatively complex and high dimensional dynamical systems compared to the mass models, the behavior of the models are investigated via frequency analysis instead of dynamical system analysis methods. Via these models, it is shown that some of the findings in experimental neuroscience can also be validated with computational models.

b- Reinforcement learning

There is evidence for the resemblance of the reward-modulation in the brain for action selection and reinforcement learning. We work on temporal difference learning, a learning method for reinforcement learning, to modulate the communication between basal-ganglia and cortex with a reward signal received from the environment. This gives rise to learning of action selection depending on the environmental inputs within the basal-ganglia-thalamus-cortex loop modeled with mass models. Additionally, we study reward-modulated STDP mechanisms for learning in decision making problems using spiking neural network models of basal-ganglia-thalamus-cortex loop, which could be tested on robots.

c- Cortical Column Models

Biophysical models of cells that make up cortical columns, considered as the basic units of sensory and motor information processes, were created and group behaviors were investigated [Yusuf Kuyumcu YL tezi, Sadeem Kbah YL tezi, E3, E14, B8]. In addition, Jansen cortical column models were used to imitate alpha-wavelength brain activities and the behaviors observed during epilepsy [E6, Ahmet Bacaksız'ın YL tezi]. There is another study of our group that takes thalamus and cortical columns into account together via the mass models related to Alzheimer’s disease [A1.4, C3]. The mass model is analyzed with bifurcation graphs and the effect of neural connections on the alpha frequency observed in Alzheimer’s is investigated.

d- Addiction

The models -created with dynamical systems approach- which comprise different levels ranging from molecular basis to systems level are used to explain addiction. With the point of view that addiction is the malfunctioning of decision making and reward-based learning systems, the models are built on thalamo-cortico-basal ganglia loop. Specifically, our studies focused on nicotine addiction so far [A.1.5, B 13, B16, F 1.8, F1.10, F 1.12, F1.15, F1.17, F2.20, F2.22, F2.25].

e- Nucleus Accumbens

Even though Nucleus Accumbens is involved in reward, it is generally overlooked. Besides, Striatum (the main input structure of basal ganglia) and its relation to motor output are studied excessively in computational neuroscience. Nucleus Accumbens receives dopamine from the ventral tegmental area. Therefore, it is an important component in reinforcement learning, and we modeled it more detaily. [A1.2, F1.2].

We built biologically plausible models of the basal nucleus. We considered the effects of neurotransmitters (especially dopamin), the formation of synchronization in neural structures [E11, E12, E13], and effects of glial cells on learning processes in our models. We especially emphasized on the models of cell-type specific behaviors of the basal nucleus. These models are based on simple dynamic systems and we analyzed them using bifurcation theory [B3, F1.4, F 1.11, F1.13, F 1.19]. We also studied the relations between experimental results with our computational models [F1.1, F1.5].
On the other hand, we expand the Hodgkin-Huxley model to reflect the cell-type specific behaviors [B10]. We also emphasized that biophysical systems can be analyzed with different approaches [C2].

We prefer to implement artificial neural network (ANN) based learning models in Python and MATLAB from scratch, i.e. without using any AI libraries. We do the hyperparameter tuning by considering the state-space behaviors of the network and by using the network-specific datasets. Trained networks are tested on test-sets using proper statistical methods. In addition, we investigate the similarities between artificial networks and biological networks to be able to build more biologically plausible artificial networks.
Usually, we work on:
1. Multilayer perceptrons (MLP) which forms a basis for deep learning
2. Self-organizing maps that is a type of competitive learning
3. Adaptive resonance theory which is a model that is based on classical conditioning in psychology
4. Elman network which a model for natural language processing
5. Hopfield network which can also be considered as a model for human memory
6. Neocognitron model which is inspired by the visual system of the brain and forms a basis for convolutional neural networks.

Besides various studies for applying these models for different problem cases, we also developed novel architectures [A1.6, A1.8, B12, B17, B18, B22, C4, E15, E19, F1.20, F1.21].
Moreover, we also worked for modeling the cognitive processes which benefits the ANN’s [A1.9, A1.11, B25, F1.18, F1.22, F1.24, F1.25, F1.26,F1.27, F1.28, F 2.27, F2.28, F2.29, F2.30].

Hardware Implementations

Robots can be thought of as a tool to complete the brain-body-environment relation for the brain-inspired decision-making models. We utilize a mobile robot and a humanoid robot to test our basal-ganglia-thalamus-cortex models developed for action selection in mass model and spiking neural network levels [F2.21, F2.26,Emeç Erçelik Yüksek Lisans tezi, bitirme ödevi, F1.9, B11, B7]. The computational models receive environmental signals (images) as inputs from the sensors of robots to make an action decision and to actuate the selected behavior through robots' joints. In this way, we study the compatibility of our models to the experimental data as well as the capabilities of input encoding for the computational models.

Neurons physically take place in the human brain, and they can function in synchrony. We can implement the models describing the cognitive processes which are executed via the relations between neural structures on integrated circuits (IC). These circuits are called neuromorphic circuits in general. The equations implemented on an IC also work synchronously. Neuromorphic circuit design requires the design of the special structures at the transistor level. It is also possible to model a network consisting of numerous neurons real-time on an FPGA. However, if the model is costly in terms of operations, this can require multiple FPGAs working in parallel. In order to avoid this, simplified models can be used on a single FPGA. In our lab, mass models and Izhikevich spiking neuron models related to the cortex-basal ganglia- thalamus are implemented on FPGA [E1, E2, F2.13].