If you would like to do an internship with me, here are a few possible projects. However, do not hesitate to formulate a project yourself!
Since we do mostly modelling, computer simulations and advanced data analysis (that means, bluntly said, computer programming and mathematics), some affinity and experience with math and programming would really help. If you’re motivated and willing to learn, you don’t have to be an expert of course, but especially for short internships it is better to be able to focus on the science than on learning how to program. Moreover, before starting any internship, it is good to think about the following questions:
- What do you expect from an internship, what do you expect to learn?
- What do you want to do next, and what (skills) do you need for that?
- What skills/techniques do you already possess?
- What topics would you like to work on, what are you interested in?
- What is a good work environment for you, what style of supervision works for you?
1. Network comparison and analysis
Recently, researchers have been able to record high-dimensional neuronal ensembles. However, how to deal with these datasets that contain high-dimensional data but low numbers of trials is still a source of much research. With individual differences between numbers of (recorded) neurons and network structure, it is impossible to obtain many trials. Also, with the recent advances in RNNs and DNNs, the question how different algorithms and network structures implement computational solutions to similar problems becomes relevant. Therefore, it is important to be able to compare networks: what are ‘similar’ networks, and which ones are ‘different’? Can we distinguish networks performing different tasks (for instance layers in cortex) based on their activity?
You will use existing data, either from our department, from online databases or simulated data. You will implement existing comparison methods (see literature) to see to what extent we can group and distinguish network on the basis of these comparison methods.
Examples of network analysis tools that could be used:
- Vinck M, Grossberger L, Battaglia F (2018) Unsupervised clustering of temporal patterns in high-dimensional neuronal ensembles using a novel dissimilarity measure. Bernstein Conference 2018. doi: 10.12751/nncn.bc2018.0233
- Williams, A. H., Kim, T. H., Wang, F., Vyas, S., Ryu, S. I., Shenoy, K. V., … Ganguli, S. (2018). Unsupervised Discovery of Demixed, Low-Dimensional Neural Dynamics across Multiple Timescales through Tensor Component Analysis. Neuron, 1–17. doi:10.1016/j.neuron.2018.05.015
- Niru Maheswaranathan, Williams, A., Golub, M., Ganguli, S., & David Sussillo. (2019). Extracting universal algorithmic principles from large populations of recurrent networks. In Cosyne Abstracts 2019, Lisbon, PT. (p. III-51).
- Hinne, M., Meijers, A., Bakker, R., Tiesinga, P. H. E., Morten, M., & Gerven, M. A. J. Van. (2017). The missing link : Predicting connectomes from noisy and partially observed tract tracing data. PLoS Computational Biology, 13(1), e1005374. https://doi.org/10.1371/journal.pcbi.1005374 (in collaboration with Bernhard Englitz and Max Hinne)
- van Elteren, C., & Quax, R. (2019). The dynamic importance of nodes is poorly predicted by static topological features. ArXiv Preprint, arXiv:1904.06654. (in collaboration with Rick Quax)
2. Single neuron comparison and analysis
Recently, we have published a large number of single-cell electrophysiological recordings (Lantyer et al. 2018), using current clamp and voltage clamp setups. In particular, the recordings include ‘frozen noise’ recordings (Zeldenrust et al. 2017), to estimate the information transfer in these neurons and the effects of different pharmacological agents (dopamine, serotonin, acethylcholin). The question is now how we can differentiate between neuron types and the effects of neuromodulators on these different neuron types.
You will use the ‘frozen noise’ data from our department (Lantyer et al. 2018) and fit neuron models to them, for instance using one of the methods mentioned below. Next, you will look for clusters in these data, to determine cell types and to determine the effects of pharmacological agents.
- Silva Lantyer, Angelica da, Niccolò Calcini, Ate Bijlsma, Koen Kole, Melanie Emmelkamp, Manon Peeters, Wim J. J. Scheenen, Fleur Zeldenrust, and Tansu Celikel. “A Databank for Intracellular Electrophysiological Mapping of the Adult Somatosensory Cortex.” GigaScience 7, no. 12 (December 1, 2018): 1–9. https://doi.org/10.1093/gigascience/giy147.
Frozen noise method
Zeldenrust, Fleur, Sicco de Knecht, Wytse J. Wadman, Sophie Denève, and Boris S. Gutkin. “Estimating the Information Extracted by a Single Spiking Neuron from a Continuous Input Time Series.” Frontiers in Computational Neuroscience 11 (2017): 49. https://doi.org/6.
Examples of tools that fit single neuron models to spike trains
- Brian 2 modelfitting toolbox
- Gonçalves, Pedro J., Jan-Matthis Lueckmann, Michael Deistler, Marcel Nonnenmacher, Kaan Öcal, Giacomo Bassetto, Chaitanya Chintaluri, et al. “Training Deep Neural Density Estimators to Identify Mechanistic Models of Neural Dynamics.” BioRxiv, 2019, 838383. http://dx.doi.org/10.1101/838383.
- Hesse, Janina, and Susanne Schreiber. “How to Correctly Quantify Neuronal Phase-Response Curves from Noisy Recordings.” ArXiv:1901.04399 [q-Bio], January 14, 2019. http://arxiv.org/abs/1901.04399.
- Ladenbauer, Josef, Sam McKenzie, Daniel Fine English, Olivier Hagens, and Srdjan Ostojic. “Inferring and Validating Mechanistic Models of Neural Microcircuits Based on Spike-Train Data.” Nature Communications 10, no. 1 (December 30, 2019): 4933. https://doi.org/10.1038/s41467-019-12572-0.
- Teeter, Corinne, Ramakrishnan Iyer, Vilas Menon, Nathan W. Gouwens, David Feng, Jim Berg, Aaron Szafer, et al. “Generalized Leaky Integrate-and-Fire Models Classify Multiple Neuron Types.” Nature Communications 9, no. 1 (2018): 1–15. https://doi.org/10.1038/s41467-017-02717-4.
- Shen, Jiamin, Ye Wang, and Lihong Cao. “Automatic Fitting of Neuron Parameters.” 2018 3rd IEEE International Conference on Cloud Computing and Big Data Analysis, ICCCBDA 2018, 2018, 558–62. https://doi.org/10.1109/ICCCBDA.2018.8386578.
- Pozzorini, Christian, Skander Mensi, Olivier Hagens, Richard Naud, Christof Koch, and Wulfram Gerstner. “Automated High-Throughput Characterization of Single Neurons by Means of Simplified Spiking Models.” PLoS Computational Biology 11, no. 6 (2015): e1004275. https://doi.org/10.1371/journal.pcbi.1004275.
3. Learning sensorimotor perception with reservoirs of spiking neural networks
Mice use their whiskers to navigate their environment. The sensory information from mechanoreceptors in the skin travels through brain stem and thalamus to the so-called ‘barrel cortex’, the part of cortex that processes information from the whiskers in rodents. Even though we can make biophysical simulation models on the one hand, and purely phenomenological models describing how to extract sensory information on the other, the bridge between these two is difficult to make. ‘FORCE’ learning is a powerful algorithm, with which recurrent neural networks can learn to perform tasks. It has been implemented in spiking neural networks by amongst others Thalmeier et al. and Nicola and Clopath (see below).
Using simultaneously recorded whisking and neural data from the department, and a spiking implementation of FORCE learning, you will make a biologically realistic implementation of a sensory perception task. Think about adapting the ‘reservoir’ in the following ways
- realistic numbers of excitatory and inhibitory neurons
- mimicking the effects of neuromodulators (see projects 2 and 8)
- ‘type 1’ and ‘type 2’ neurons
- Sussillo, D., & Abbott, L. F. (2009). Generating Coherent Patterns of Activity from Chaotic Neural Networks. Neuron, 63(4), 544–557. https://doi.org/10.1016/j.neuron.2009.07.018
- Nicola, W., & Clopath, C. (2017). Supervised learning in spiking neural networks with FORCE training. Nature Communications, 8(1), 1–15. https://doi.org/10.1038/s41467-017-01827-3
- Thalmeier, D., Uhlmann, M., Kappen, H. J., & Memmesheimer, R. M. (2016). Learning Universal Computations with Spikes. PLoS Computational Biology, 12(6), 1–29. https://doi.org/10.1371/journal.pcbi.1004895
4. Simulating sensorimotor perception using the efficient coding framework
Mice use their whiskers to navigate their environment. The sensory information from mechanoreceptors in the skin travels through brain stem and thalamus to the so-called ‘barrel cortex’, the part of cortex that processes information from the whiskers in rodents. Even though we can make biophysical simulation models on the one hand, and purely phenomenological models describing how to extract sensory information on the other, the bridge between these two is difficult to make. The efficient coding framework (Denève & Machens, 2016) is a powerful framework that describes how such computations could be performed.
Using simultaneously recorded whisking and neural data from the department, a kernel implementation of the efficient coding framework (Zeldenrust, Gutkin, Denève, 2019) and a Poisson implementation of the framework (Buxó, Rullán, Pillow, 2019) you will make a GLM-formulation of the efficient coding framework. Next, you will fit a GLM model to existing data (see project 1) and compare the fitted and the derived networks.
- Buxó, Camille E Rullán, and Jonathan W Pillow. “Poisson Balanced Spiking Networks.” BioRxiv, 2019, 836601. http://dx.doi.org/10.1101/836601.
- Denève, Sophie, and Christian K. Machens. “Efficient Codes and Balanced Networks.” Nature Neuroscience 19, no. 3 (2016): 375–82. https://doi.org/10.1038/nn.4243.
- Zeldenrust, Fleur, Boris Gutkin, and Sophie Denève. “Efficient and Robust Coding in Heterogeneous Recurrent Networks.” BioRxiv, 2019. https://doi.org/10.1101/804864.
- Wärnberg, Emil, and Arvind Kumar. “Low Dimensional Activity in Spiking Neuronal Networks.” PLoS Computational Biology 15, no. 5 (2019): e1007074. https://doi.org/10.1371/journal. pcbi.1007074.
5. Comparing methods of single-cell information transfer
Measuring the mutual information between the input into a neuron and its output spike train, helps us in determining the information-processing capabilities of the brain. However, it is notoriously difficult to measure mutual information, as during in-vitro experiments cells do not stay alive long enough (approximately one hour) to measure the full input-output probability distribution. Undersampled distributions typically give rise to strong biases.
Several methods have been proposed to deal with the problem stated above. Using single neuron models, you will implement and compare these methods, and assess their robustness against short samples (low spike numbers) and their sensitivity to neuron types (i.e. type 1 and type 2 neurons).
Alternatively, you can improve the Zeldenrust et al. 2017 method to include injected conductance instead of current, using dynamic clamp (see Prinz et al. 2004) methods in simulations and/or experiments.
- Dettner, A., Münzberg, S. & Tchumatchenko, T., 2016. Temporal pairwise spike correlations fully capture single-neuron information. Nature Communications, 7, pp.1–11.
- Houghton, C., 2018. Calculating the Mutual Information Between Two Spike Trains. bioRxiv, p.423608.
- Prinz, A. A., Abbott, L. F., & Marder, E. (2004). The dynamic clamp comes of age. TRENDS in Neurosciences, 27(4), 218–224. Retrieved from http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&dopt=Citation&list_uids=15046881
- de Ruyter van Steveninck, R. R., & Bialek, W. (1988). Real-time performance of a movement-sensitive neuron in the blowfly visual system: coding and information transfer in short spike sequences. Proceedings of the Royal Society of London. Series B, 234(1277), 379–414. Retrieved from http://www.jstor.org/stable/36210
- Strong, S. P., Koberle, R., de Ruyter van Steveninck, R. R., & Bialek, W. (1998). Entropy and Information in Neural Spike Trains. Physical Review Letters, 80(1), 197–200.
- Zeldenrust, F., de Knecht, S., Wadman, W. J., Denève, S., & Gutkin, B. (2017). Estimating the Information Extracted by a Single Spiking Neuron from a Continuous Input Time Series. Frontiers in Computational Neuroscience, 11, 49. https://doi.org/6
6. Estimating the influence of top-down versus bottom-up modulation
The neural response in barrel cortex depends on the incoming bottom-up sensory input, but also in top-down input from other brain areas (higher-order layers and motor cortex). The trial-to-trial variability of the neural response therefore depends both on the type of coding that is used (rate versus temporal for instance), but also on the amount of feedback or top-down input to the network. In this project we will use both real and simulated data to estimate the amount of feedback input to barrel cortex.
Using data from an online dataset (or other) we will use machine learning protocols to determine how well we can estimate several whisker input parameters (pole position, maximum curvature, whisker angle/phase upon contact) from the neural network response. In other words, we will train a classifier on the neural network data
The online Peron/Svoboda data include calcium recordings in head-fixed animals freely whisking against a pole with different positions. Each whisker contact-event, is determined by the pole location, whisker angle, angular velocity and phase upon contact and the maximum curvature of the contact. Due to the large difference in temporal resolution between the neural recordings and the whisker recordings, we can extract only a single neural population vector for each contact event. In this project we will train a (deep learning?) neural network to see how well the input parameters can be decoded from the population vector. This will give us an estimate on the coding precision of barrel cortex.
Presently, we have a biophysical model of L4 and L2/3 of barrel cortex. This model receives thalamic spike trains as input, which is generated by filter neurons responding to whisker angle and curvature. The model does not include top-down or feedback networks, but each neuron can receive user-defined external input, mimicking this feedback input. So trial-to-trial variability in a given simulated network is the result of
- Spike timing precision of simulated thalamic spike trains
- Initial conditions
- Synaptic failures
- External input (mimicking feedback/top-down modulation)
We can vary each of these four sources of trial-to-trial variability and estimate their influence on the quality of the input representation by the network, using the methods of the decoding subproject. Given the results of this decoding subproject, we can estimate the strength of the influence of each of these four sources of variability (i.e. how they should be set so that the simulated network has a similar performance as the online data set) in real networks.
7. Estimating (un)certainty from neural recordings
In the gap-crossing task (Celikel&Sakmann, 2007), freely moving rodents locate a stationary tactile target in darkness. They only cross from one platform to the next, when they are certain enough of the position of the target. Recently, van Bergen and Jehee (2018) have developed a method to decode uncertainty from neural recordings.
In this project, you will use data from the department obtained during the gap-crossing task. You will adapt the method from van Bergen and Jehee (2018), which was initially developed for fMRI data, so that it can be used to
van Bergen, R. S., & Jehee, J. F. M. (2018). Modeling correlated noise is necessary to decode uncertainty. NeuroImage, 180(August 2017), 78–87. https://doi.org/10.1016/j.neuroimage.2017.08.015
Celikel, T., & Sakmann, B. (2007). Sensory integration across space and in time for decision making in the somatosensory system of rodents. Proceedings of the National Academy of Sciences, 104(4), 1395–1400. https://doi.org/10.1073/pnas.0610267104
8. Biophysical modelling of the effects of neuromodulators (dopamine, acethylcholine, serotonin) and potassium channels
Recent findings in our Department suggest that several pharmacological agents (dopamine, serotonin, acethylcholin and a molecule that controls cell-cycle) are critically involved in the modulation of cortical excitability. Whole cell current clamp studies suggest that these modulators perform their action via changing the spike thresholds in inhibitory and excitatory neurons; voltage clamp studies show that this change in spike threshold is associated with voltage gated sodium and potassium channels (in particular the Kv1.1 and Kv3.1 channel, prominent voltage gated potassium channels).
This project will involve biophysical modelling in one of three levels:
- a compartmental model neuron
- a microcircuit (connected inhibitory and excitatory cell)
- a network model
to determine the effects of one of these neuromodulators on the firing properties of single pyramidal cells and interneurons, and the subsequent effects on the microcircuit or network level. The student will use our existing model of cortical pyramidal neurons and interneurons to
- on a single neuron level: determine the mechanistic causes of the observed changes in excitability, in particular the effects of the Kv1.1 density distribution on the firing properties.
- investigate the microcircuit effects of the observed changes in excitability of inhibitory and excitatory cells
- investigate the network effects of the observed changes in excitability of inhibitory and excitatory cells
Ideally, the results of this modelling study will be corroborated by in vitro experiments. The intended endpoint of this internship is a manuscript suitable for publication in a peer-reviewed journal.
- Stimberg, Marcel, Romain Brette, and Dan FM Goodman. “Brian 2, an Intuitive and Efficient Neural Simulator.” ELife 8 (2019): 1–41. https://doi.org/10.7554/elife.47314.
9. Cell recruitment with external stimulation
In standard population excitability procedures, neurons are stimulated extracellularly. A baseline is defined by measuring the EPSP amplitude in postsynaptic neurons, and determining to current that is needed to evoke an EPSP with a certain amplitude. Next, the amplitude of the current injection is increased by steps relative to this baseline. This experiment is then repeated following certain procedures or pharmacological treatments, to infer network effects. However, if connections (pre to post) or excitability of neurons change, the baseline might not be maintained over conditions. So how should changes in EPSP amplitude/slope be interpreted?
In this project, you will model how cells get recruited as a function of the external stimulation amplitude, i.e. you will make/use realistic neuron models (cell locations according to this model) and model how current spreads into a sphere. You will then model how cells get recruited. Finally, you will model how EPSPs in neighbouring columns are influenced by excitability and connectivity changes.
- Joucla, S., Glière, A., & Yvert, B. (2014). Current approaches to model extracellular electrical neural microstimulation. Frontiers in Computational Neuroscience, 8(February), 1–12. doi:10.3389/fncom.2014.00013
- Overstreet, C. K., Klein, J. D., & Helms Tillery, S. I. (2013). Computational modeling of direct neuronal recruitment during intracortical microstimulation in somatosensory cortex. Journal of Neural Engineering, 10(6). doi:10.1088/1741-2560/10/6/066016