Available Master and Bachelor thesis 2020/21


Interested in one of the projects? Send an email to stefano.nichele@oslomet.no.

Modelling and training of Artificial (Spiking) Neural Networks with variable spikes propagation velocity (Supervised by: Kristine, Ola, Stefano) - Available for master and bachelor thesis

This thesis is part of two ongoing research projects at the OsloMet Living Technology Lab (part of OsloMet AI Lab) financed by the Norwegian Research Council, Socrates and DeepCA. Both projects are in collaboration with the Sandvig Lab at NTNU, where networks of biological neurons are grown in a "petri dish" over micro-electrode arrays (MEA) which allow the recording of their neural activity and their stimulation through electrical signals.

In this thesis, a novel AI computer model which consists of a specific type of artificial neural network, i.e. spiking neural network (SNN), will be developed where the propagation speed of signals (spikes) in the network may vary over time as a result of the network activity. Such behavior has been observed experimentally in biological neurons, as reviewed in [1]. This is related to ongoing work on fabrication of specific axon spike tracking chips to be tested with biological neurons. Therefore, the results and models developed in this thesis project will be used in the future to reproduce and investigate the results obtained at the Sandvig Lab.

The developed SNN models will be tested with AI benchmarks, and properties / performances will be investigated. In order to improve the performances of the network models, a training algorithm has to be implemented. One possible choice for this type of model is evolutionary algorithms, which have been extensively used to train artificial neural network models (neuroevolution).

[1] Debanne, D. Information processing in the axon. Nat Rev Neurosci 5, 304–316 (2004). https://doi.org/10.1038/nrn1397

[2] R. Wang, C. Jin, A. McEwan and A. van Schaik, "A programmable axonal propagation delay circuit for time-delay spiking neural networks," 2011 IEEE International Symposium of Circuits and Systems (ISCAS), Rio de Janeiro, 2011, pp. 869-872, doi: 10.1109/ISCAS.2011.5937704.


Functional connectivity graph from in-vitro neural recordings with weight estimation (Supervised by: Ola, Kristine, Stefano) - Available for master thesis

This thesis is part of two ongoing research projects at the OsloMet Living Technology Lab (part of OsloMet AI Lab) financed by the Norwegian Research Council, Socrates and DeepCA. Both projects are in collaboration with the Sandvig Lab at NTNU, where networks of biological neurons are grown in a "petri dish" over micro-electrode arrays (MEAs) which allow the recording of their neural activity and their stimulation through electrical signals.

While the neuronal cultures studied in this project are composed of hundreds of thousands of neurons, the MEAs provides only 60 electrodes (or a few thousands for the new CMOS MEAs). It is therefore not possible to communicate with every single neuron and infer the exact neural network topology. However, by recording the activity on the electrodes, it is possible to construct a functional connectivity graph by looking at the correlations of activities on the different electrodes. The extracted network models represent the functional (behavioral) map of the neural culture. Once such network models are extracted from real data, it is possible to study the graphs with network theory measures. However, in this project the main goal is to be able to generate a weighted graph from its unweighted counterpart using machine learning algorithms, with the goal of optimizing the tradeoff between information transmission in the network and energy consumption. The results of the optimized network will be compared with weight estimates calculated from the recorded electrical activity using information theory estimation methods.

In addition, the strengths of the connections may change over time in response to network self-organization and growth and/or due to the appearance of a neural disease. Recordings from real biological networks are available over large periods of time for both healthy and perturbed (sick) cultures. Therefore, this project has direct medical implications, in particular to gain a better understanding of the effects of neural diseases on the functionality of the networks, such as their ability to learn.

[1] E. Bullmore and O. Sporns, “The economy of brain network organization,” Nature Reviews Neuroscience. 2012.

[2] S. B. Laughlin and T. J. Sejnowski, “Communication in neuronal networks,” Science. 2003.


Modelling the behavior of in-vitro biological neural networks from neural recordings with dynamical system models (cellular automata, random boolean networks, echo state networks and liquid state machines) using the EvoDynamic framework (Supervised by: Sidney, Kristine, Stefano) - Available for master thesis

This thesis is part of two ongoing research projects at the OsloMet Living Technology Lab (part of OsloMet AI Lab) financed by the Norwegian Research Council, Socrates and DeepCA. Both projects are in collaboration with the Sandvig Lab at NTNU, where networks of biological neurons are grown in a "petri dish" over micro-electrode arrays (MEA) which allow the recording of their neural activity and their stimulation through electrical signals.

An important open problem is to generate and train models of the biological networks at different levels of abstraction. Models are very useful for studying (1) possible effects of changes in the network connectivity (it is not possible to test such changes directly on the biological networks without changing the actual networks) and (2) performances of the different biological network topologies for AI benchmarks and computing tasks.

The goal of this project is to generate abstract models of biological neural networks from recordings of neural data from MEAs using different dynamical systems models. Such models include abstract Cellular Automata / random Boolean network models, recurrent artificial neural networks such as echo state networks and liquid state machines. The models will be generated using the EvoDynamics framework [1], a software package that uses evolutionary algorithms to train models of dynamical systems.

[1] EvoDynamics: https://github.com/SocratesNFR/EvoDynamic/


Cellular automata with plasticity: cellular learning automata (Supervised by: Sidney, Anis, David, Stefano) - Available for master and bachelor thesis

This thesis is part of ongoing research projects at the OsloMet Living Technology Lab (part of OsloMet AI Lab) on neuro-inspired unconventional computing machines. The long term goal of this project is to build computing machines that go beyond the current von Neumann paradigm of computing, by taking inspiration from how the brain works. Cellular Automata (CA) are interesting models of cellular computing, where the actual information processing, transmission of information and storage are massively distributed and parallelized, and each component of the system interacts only locally with the closest neighbors. One such example of cellular automata is the Game of Life, which is proven to be computationally universal.

While CA can produce very complex computations, they lack one key aspect of (biological and artificial) neural networks, i.e., plasticity. Neuroplasticity is the ability of the brain to change (learn) over time. One class of automata that can learn over time is Learning Automata (LA), a special type of reinforcement learning automata.

This project aims at creating a new CA-based machine learning paradigm by combining CA and LA, i.e., Cellular Learning Automata (CLA). In this way, each cell in a CA may change over time its function (transition rule) based on the actual local activity of the system, providing a mechanism of plasticity in cellular automata (a kind of Hebbian learning for CA).


Growing Neural Cellular Automata Intelligence (Supervised by: Sidney, Tom, Stefano) - Available for master thesis

This thesis is part of ongoing research projects at the OsloMet Living Technology Lab (part of OsloMet AI Lab) on neuro-inspired unconventional computing machines. The long term goal of this project is to build intelligent machines that go beyond the current von Neumann paradigm of computing, by taking inspiration from how the brain works. Cellular Automata (CA) are interesting models of cellular computing, where the actual information processing, transmission of information and storage are massively distributed and parallelized, and each component of the system interacts only locally with the closest neighbors. One such example of cellular automata is the Game of Life, which is proven to be computationally universal. Another example, in this case of a continuous CA is Lenia. One recent work by Google is Growing Neural Cellular Automata, a continuous CA where neural networks are trained to control the growth process of given shapes, and to regenerate when damaged.

In this project, the goal is to extend such continuous CA worlds to map specific cells to be system inputs/outputs, instead of targeting given shapes. This is performed by training neural networks to achieve a certain input/output function, i.e., to create a physical (simulated in a CA) process that learns to solve the given task (a kind of cellular automata brain).

Mordvintsev, Alexander, et al. "Growing neural cellular automata." Distill 5.2 (2020): e23.

Chan, Bert Wang-Chak. "Lenia: Biology of Artificial Life."


The Science Garden: Visualization of Scientific Papers (Supervised by: Trym, Peder, Tom, Stefano) - Available for bachelor thesis

This thesis is hosted by the OsloMet Living Technology Lab (part of OsloMet AI Lab). Aim: To create a webpage that builds a graph based on the chains of citations and references of a given scientific paper or authors’ papers so that far reaching impacts and dependencies can be easily visualized.

Problem: The body of scientific work humanity has produced so far is immense and the dependencies (chains of references) and far reaching impacts are difficult to grasp when presented in their current format of a simple reference list or the “cited by” list provided by Google Scholar.

This issue exists at two levels, the impact and the dependencies of a publication.

Impact: "Cited by" lists only tracks a single layer of impact, but any given publication may be the stepping stone for another publication that spawns a dozen other publications and so on. Knowing this far reaching impact of any given publication is in itself interesting, but is also important when tracking the impact of a retraction, or the importance of producing replications of a given paper as the more other papers relies on a specific finding, the more important the quality of the work becomes.

Dependencies: A similar issue arises for the dependencies. If a paper used as reference is retracted this may warrant revision of the publications that reference it. However, such retractions may lie further back than the first layer of references making detecting such issues difficult. A visual representation of these dependencies will give authors a quick and easy method of assessing the quality of the foundation of a given paper.

Project outline: In this project the task will be to design a webpage that inputs a publication and generates a graph showing multiple levels of the chains of citations originating from the input paper as well as the chains of references the paper depends on.

Desired functionality: - Each node should be a clickable hyperlink which takes the user to the paper represented by the node - Hovering over the node/clicking it should bring up it’s abstract - The system should detect if a paper has been retracted and the graph displayed on the webpage should display this in a visual fashion within the graph. - It should also be possible to input authors which should generate a "garden" of graphs where each publication the author has produced is the seed node for a graph.

Visual considerations: - It would be desirable if the visual format of the graph mimicked plants in some sense o Retractions could then be represented as unhealthy branches or similar. o Dependencies as roots o "cited by" as trunks and branches - The Y axis of the webpage should correspond to chronological time such that each publication node is at a height corresponding to its publication date - The creation of the graph should be shown visually as the webpage builds the graph from the seed node(s). As if it is a growing plant.

https://retractionwatch.com/

https://pypi.org/project/scholarly/

https://www.connectedpapers.com/


Deep hybrid (artificial) reservoirs: combinations of cellular automata, random boolean networks, echo state networks, and liquid state machines (Supervised by: Trym, Tom, Sidney, Stefano, Claudio) - Available for master thesis

This thesis is part of an ongoing research project at the OsloMet Living Technology Lab (part of OsloMet AI Lab) financed by the Norwegian Research Council, DeepCA. The DeepCA project aims at exploring novel AI methods for recurrent neural networks both by taking inspiration from how biological neural networks work and utilizing them as components in hybrid systems within a reservoir computing paradigm.The project is conducted in collaboration with the Sandvig Lab at NTNU, where networks of biological neurons are grown in a "petri dish" over micro-electrode arrays (MEA) which allow the recording of their neural activity and their stimulation through electrical signals.

Reservoir computing is a growing field within artificial intelligence where a non-linear problem is projected into an untrained recurrent neural network (or any other multi-dimensional untrained medium) before a single trained output layer is utilized. Echo state networks and liquid state machines are modern approaches to reservoir computing using different types of artificial neuron models and take inspiration from the brain. Cellular Automata (CA) are on the other hand a grid of cells, each cell can be in a number of states, which in discrete steps change state based on local interactions in a structured way. Random Boolean networks (RBN) are similar to Cellular automata except the connections between cells are not structured but randomly initialised. CAs have been used as a plausible substrate for reservoir computing.

Recent advances in reservoir computing have been showing promising results by stacking reservoir layers [1], as is typically done in Deep Learning. This study did however only use a single type of reservoir, different types may show different dynamics and computations and may thus impart additive benefits in problem solving. This partly takes inspiration from the fact that different brain areas have different structures which may be optimized for different tasks.

In this project, deep hybrid reservoirs will be tested, where each reservoir layer is composed by a different type of architecture. Architectures of interest are Cellular Automata, random Boolean networks, echo state networks and liquid state machines as described above.

[1] Gallicchio et al. (2017). Deep reservoir computing: a critical experimental analysis https://www.sciencedirect.com/science/article/abs/pii/S0925231217307567


Comparison and benchmarking of cellular automata reservoir vs. random boolean networks reservoir (Supervised by: Tom, Stefano) - Available for master and bachelor

This thesis is part of an ongoing research project at the OsloMet Living Technology Lab (part of OsloMet AI Lab) financed by the Norwegian Research Council, DeepCA. The project is in collaboration with the Sandvig Lab at NTNU, where networks of biological neurons are grown in a "petri dish" over micro-electrode arrays (MEA) which allow the recording of their neural activity and their stimulation through electrical signals, i.e. training.

Reservoir computing is a growing field within artificial intelligence where a non-linear problem is projected into an untrained recurrent neural network (or any other multi-dimensional untrained medium) before a single trained output layer is utilized. Cellular Automata (CA) is a grid of cells, each cell can be in a number of states, which in discrete steps change state based on local interactions in a structured way. Random Boolean networks (RBN) are similar to Cellular automata except the connections between cells are not structured but randomly initialised. CAs have been used as a plausible substrate for reservoir computing.

This thesis will consist of reading and reviewing the state of the art, implementing and experimenting with random Boolean networks as a reservoir and comparing the results to the state of the art cellular automata [1,2]. One possible algorithm for automatically searching suitable CA / RBN models is evolutionary algorithms.

[1] Nichele, S., Molund, A. (2017). Deep learning with cellular automaton-based reservoir computing.

[2] Babson, N., Teuscher, C. (2019). Reservoir Computing with Complex Cellular Automata.


Investigation and visualization of trajectories and attractors from data recordings of in-vitro biological neural networks at criticality (Supervised by: Stefano, Kristine) - Available for master and bachelor thesis

This thesis is part of two ongoing research projects at the OsloMet Living Technology Lab (part of OsloMet AI Lab) financed by the Norwegian Research Council, Socrates and DeepCA. Both projects are in collaboration with the Sandvig Lab at NTNU, where networks of biological neurons are grown in a "petri dish" over micro-electrode arrays (MEA) which allow the recording of their neural activity and their stimulation through electrical signals.

Biological neural networks (including our brain) tend to self-organize into a critical behavior activity which is neither synchronized nor random, i.e., critical. At criticality, the spiking behavior of the network is highly complex. Data recording of spiking activity from several biological neural networks in critical, sub-critical, and supercritical states have been collected from MEAs, for both healthy and perturbed (sick) neurons.

In this project, the goal is to identify and test ways of detecting and visualizing repetition of complex patterns (neural states) which the network traverses while it computes (trajectories) and if there exist repetition of states in which the network has been before (attractors). This will give insight into how biological neural networks process information and the effects of neural diseases on such processing. This project is part of a larger effort to design artificial intelligence that is more similar to how biological neural networks process information and learn.

[1] J. M. Beggs, “The criticality hypothesis: How local cortical networks might optimize information processing,” Philos. Trans. R. Soc. A Math. Phys. Eng. Sci., vol. 366, no. 1864, pp. 329–343, 2008.

[2] J. M. Beggs and D. Plenz, “Neuronal avalanches are diverse and precise activity patterns that are stable for many hours in cortical slice cultures,” J. Neurosci., vol. 24, no. 22, pp. 5216–5229, 2004.

[3] C. Haldeman and J. M. Beggs, “Critical branching captures activity in living neural networks and maximizes the number of metastable states,” Phys. Rev. Lett., 2005.


Cortical column inspired reservoir computing (Supervised by: Stefano, Trym) - Available for master thesis

This thesis is part of two ongoing research projects at the OsloMet Living Technology Lab (part of OsloMet AI Lab) financed by the Norwegian Research Council, Socrates and DeepCA. Both projects are in collaboration with the Sandvig Lab at NTNU, where networks of biological neurons are grown in a "petri dish" over micro-electrode arrays (MEA) which allow the recording of their neural activity and their stimulation through electrical signals.

In this project, the goal is to build a novel neural network model that is inspired by how the cortical neurons are organized. For an introduction on the cortical circuitry see [1].

In particular, the neocortex is organized in deep (layered) and parallel elements of about 100 neurons called cortical minicolumns. Macrocolums are agglomerates of several minicolumns (around 10.000 neurons, i.e., wide). Such structures have a certain degree of lateral interactions and recurrencies (sparsely connected recurrent neural networks). By taking inspiration from the cortical columns organization in our brain, we want to create a novel reservoir computing model which uses spiking neurons, with local (hebbian) learning rules, organized in minicolumns. Furthermore, one particular aspect that may be tested in such networks is their performances at criticality which according to the critical brain hypothesis may enhance the brain’s capacity

[1] https://www.insightsfromthebrain.com/ page 66-69


Reproducing stochastic cellular automata with quantum gates (Supervised by: Sidney, Stefano, Pedro, Sergiy) - Available for master thesis

Cellular automata and quantum computing are two important topics in modern computer science. Cellular automata (CA) are classical discrete models for reproducing complex processes of extended systems of coupled elements, based in simple rules which map the present state of each element and its direct neighbours into the next state of each element. Quantum computing is a field in its infancy aiming at solving the limitations of classical computation.

As common digital gates are the elementary components of classical digital circuits in a computer chip, quantum gates are elementary components of a quantum computer and their output, after measurement, are on- and off-state with a specific probability. Consequently, they may be good candidates for mimicking so-called stochastic cellular automata, discrete models which define the cellular automaton rules with an associated probability.

The aim of this project is to generate specific rules of stochastic cellular automata, using the minimal set of five quantum gates which is known to generate any quantum circuit. To build the quantum circuit associated to such specific stochastic rules, we will use evolutionary machine learning approaches based in mutation and selection of several configurations of quantum gates until the overall similarity with the given rule is attained. This project is part of ongoing projects at the OsloMet AI Lab, crossing different of its sub-divisions, namely the Living Technology Lab, the Quantum Computing Lab and the Stochastic Lab.


Entangling qubits through Machine Learning (Supervised by: Sergiy, Pedro, Sidney, Stefano) - Available for master thesis

Entanglement is a striking type of quantum correlations, quoted by Einstein as “spooky action at a distance”, which is the most valuable resource of quantum computations. To prepare a set of quantum bits, so-called qubits, in a highly entangled state is fundamental for quantum computations. It is like to load a car with fuel before a long race. Yet, a preparation of such states by using quantum gates, that are the building blocks of quantum circuits, is a highly nontrivial task which demands a knowledge of several aspects of quantum dynamics and a significant effort and time consuming analysis.

The idea of this project is to consider a possibility to delegate this cumbersome task to neural networks. Neural networks were found to be extremely good in solving practical classical problems, such as pattern recognitions, traffic control and big data sorting. The project will answer the question whether they are also good for solving quantum problems. In particular, it may provide us with a framework that retrieves a highly entangled state from a pre-given number of qubits. Being a machine learning algorithm it will learn from solving some of the simplest problems in this scope, using only a few qubits, and consequently be prepared for heavier tasks with more qubits. This project is part of ongoing projects at the OsloMet AI Lab, crossing different of its sub-divisions, namely the Living Technology Lab, the Quantum Computing Lab and the Stochastic Lab.


Artificial Life and Artificial Intelligence Art (Supervised by: Stefano, Kristin, Boel) - Available for master

The Faculty of Technology, Art and Design at OsloMet is a unique environment, with strong collaborations across technology, art and design. There is enormous interest in this renewed convergence of art and technology around the globe, with new institutions founded, public initiatives functioning increasingly professionally, a plethora of projects, events, and a considerable number of publications. The picture is one of a booming field.

The FELT project - Futures of Living Technologies [1] engages in the interrelations and intersections that occur between human beings, living environments and machines, relations on the edge of how we experience aliveness today. This might evoke a sense of the uncanny and a fear of being dominated by the machine, but also reveals a world of possibilities of becoming, creation of new forms and behaviors.

This research studies functions of living systems such as intelligence, evolution, reasoning and learning and provides a framework of state-of-the-art scientific research that is made available to artists. We aim at merging artistic strategies from bioart and techno-ecologies with contemporary perspectives on sensory experience and materiality in artistic production and research. Bioart explores the principles of phenomena associated with living systems. The blending of computer technology and robotics with biology is moving into the realm of constructing synthetic organisms and biological programming. Artistic responses vary from dystopian visions of total control too playful sci-fi utopian visions.

[1] https://sites.google.com/view/feltproject/home

Inspiration and examples:
http://artaward2018.alifelab.org/#winner
http://www.georgekhut.com/portfolio/behind-your-eyes-between-your-ears-liveworks/
http://www.studioovervelde.com/projects/edge-of-chaos/
https://jonmccormack.info/artworks/
http://www.mariacastellanos.net/


The living universe: evolving self-replicating hypergraph universes (Supervised by: Stefano) - Avalable for master and bachelor thesis

This thesis is part of ongoing research projects at the OsloMet Living Technology Lab (part of OsloMet AI Lab). The long term goal of this project is to build a radically new paradigm of computation. Cellular Automata (CA) are very simple computing universes, where the actual information processing, transmission of information and storage are massively distributed and parallelized, and each component of the system interacts only locally with the closest neighbors. CA can, with very simple rules, give rise to incredible complexity. One such example of cellular automata is the Game of Life, which is proven to be computationally universal.

Recently, Stephen Wolfram has hypothesized that our universe may be also governed by such simple rules, and has suggested unifying the fundamental theory of physics with such simple computations [1]. In particular, the proposed model of such a computational universe is a special type of multidimensional graphs, called hypergraphs (graphs in which edges may connect more than two nodes) that are generated from a simple initial condition with rewriting rules. While the rules may be very simple, some of them when applied multiple times give rise to very complex graphs (while some produce very simple graphs). The focus of this project is on finding rules that create complex graphs (computational universes) that can self-replicate, i.e., evolving universes that can develop and generate new universes as living systems do.

[1] https://writings.stephenwolfram.com/2020/04/finally-we-may-have-a-path-to-the-fundamental-theory-of-physics-and-its-beautiful/


Artificial neural networks synthetic neurobiology - analysis of artificial neural networks with tools inspired by computational neuroscience (Supervised by: Gustavo, Stefano) - Avalable for master thesis

Recent advances in deep neural networks have resulted in impressive results in several complex tasks. At the same time, recent advances in neuroscience have resulted in novel data analysis methods for the understanding of biological neural networks, brains, and animal/human behavior.

However, such methods applied to biological neural networks have not been yet applied to understand the behavior of artificial neural networks. One of the reasons may be the lack of scientific exchange between the two scientific communities. Another motivation is the abstraction of artificial neural networks that do not necessarily recapitulate the functioning of biological neurons.

In this project, state-of-the-art biologically-inspired neural network models will be analyzed using data analysis methods traditionally used in neuroscience research. The results of this project have high potential to impact the development of the field of explainable, trustworthy and sustainable AI.

Reading material: Deep neuroethology of a virtual rodent. Merel et al. ICLR 2020. https://arxiv.org/abs/1911.09451


Open project on in vitro neural data analysis (Supervised by: Kristine, Stefano) - Avalable for master thesis

This thesis is part of two ongoing research projects at the OsloMet Living Technology Lab (part of OsloMet AI Lab) financed by the Norwegian Research Council, Socrates and DeepCA. Both projects are in collaboration with the Sandvig Lab at NTNU, where networks of biological neurons are grown in a “petri dish” over microelectrode arrays (MEAs), which allow the recording and electrical stimulation of neural activity.

In this thesis, MEA data from biological neuronal networks will be analyzed to answer targeted research questions. Depending on the interests of the student and the available data, the thesis work may take on one of the following different approaches:

1. A methodological approach. This approach would involve taking sample data and developing an analytical tool that could be used by others in future research. This tool would allow the user to extract relevant features from the data, such as the network connectivity or features related to dynamic state of the network.

2. A modeling approach. This approach would involve constructing data-driven models that emulate features of the data. For example, models could be constructed using the connectivity extracted from the data, and their behavior could be compared with similar models with a different connectivity.

3. An experimental approach. This approach would involve understanding and attempting to answer the neuroscience-related questions at the root of the experimental design. For example, the data may be collected from diseased and healthy populations of neurons with an aim to see differences in their behavior, and the task would then be to find a way to classify the networks as diseased or healthy.

The availability of this thesis depends on the status of ongoing work and the availability of data. Please get in touch with the supervisors for details on the specific data and possible approaches currently available for this thesis.


Your own project (Supervised by: Stefano) - Available for bachelor and master thesis

Do you have a cool idea for your own project with research questions that fit with the scope of the Living Technology Lab?

Get in touch with Stefano!