Towards a deep learning model of information encoding and decoding of in vitro neuronal cultures responses to electrical stimulation
-
1
Universidad Politécnica de Cartagena, Electronics, Computer Technology and Projects, Spain
-
2
Universidad Miguel Hernández de Elche, Bioengineering Institute, Spain
-
3
Universidad Nacional de Educación a Distancia (UNED), Artificial Intelligence, Spain
Motivation
In order to advance in the understanding of neuronal cultures as an alternative for in silico computers that could lead to a new generation of massivelly parallel biological neuroprocessors, and the possibility of creating artificial bioinspired systems able to efficiently mimic the biological ones, we investigate the capacity of convolutional and recurrent neural networks in modeling the behavior of in vitro neuronal cultures cultures in response to electrical stimulation -encoding problem-, along with the reconstruction of the stimuli -decoding problem-, comparing its performance with other machine learning and traditional modeling and decoding techniques.
In this work, we train machine learning systems that mimic the biological ones in order to encode and process information, and we visualize the feature extractors learned to obtain information that could reveal inner dynamics of the system. In addition, we discuss about the future role between artificial and biological computation and its possible applications.
Material and Methods
The spiking activity from embryonic cortical cultures seeded over a multi-electrode array(MEA) was recorded during spontaneous activity and following electrical stimulation, in order to analyze its global dynamical response.
The properties of the spontaneous activity of the cultures before the stimulation experiments is measured. We will apply the stimulus over a subset of electrodes to find out which of them trigger consistent neural responses over the neural circuit, to create input-output recordings that will be used for the modeling. After filtering raw data of spikes, the activity in electrodes after stimulation is compared with the basal activity (without stimulations) to detect which electrodes are sensitive to stimulations, and to discard the inactive ones.
The sequence of time-stamps of the spikes obtained in the preprocessing, from the ones recorded in an active electrodes during the 150 ms immediately after a stimulation, are extracted from the recording.
After this, the spike trains will be smoothed by means of a Gaussian filter which will transform the discrete spikes into spiking probabilities that will be predicted by model.
Two approaches will be considered: single trial fitting and several-trial PSTH fitting, both for modeling and stimulus reconstructing tasks.
Several machine learning and deep learning techniques –including convolutional neural networks and LSTM units-, along with traditional ones are applied to both the encoding problem, this is, mapping the input –electrical impulses- to the output –the culture’s recorded firing rates-, and the decoding problem: reconstructing the input from the culture responses, and a systematic search of hyperparameters such as optimization algorithms and error functions will be compared to find the best fitting settings.
Visualization and comparison of the features learned by the models in both the encoding and decoding tasks will helps us to analyze the most important characteristics of the information to be processed and could help us reveal the inner dynamics of the biological processor.
Results
The dataset consisted of 200 repetitions patterns from combinations of electrical pulses from 4 different electrodes and its combinations as the input to the biological system to be modeled, and to the populations responses treated as spike trains and transformed into dynamic spiking probabilities as the output, with time resolution of 10 ms. For the encoding problem, which consist on spike probabilities prediction –output- given the electrical stimulus –input-, several machine learning and traditional modeling techniques used in neuroscience, as Support Vector Machines and Generalized Linear Models has been applied and compared to Deep Learning models (Convolutional and Recurrent Neural Networks). A similar approach is adopted to address the decoding problem (reconstruction of the electrical patterns from the neuronal activity of the biological neuronal population). A posterior analysis of the features learned for both the encoding and decoding problem is performed.
Computational models of the biological processing are compared in terms of performance metrics –such as distance measures between the predicted spiking probabilities and the real ones and correlation coefficient-, computation time and number of parameters.
Discussion and Conclusion
Machine Learning-Deep Learning models are able to mimic the neural culture information processing, often outperforming traditional modeling techniques, both in the encoding and decoding tasks, and the results suggests that this difference could be higher when larger and more complex datasets become available. These models not only can reveal interesting information about the information codification in biological neuroprocessors, but also they can be developed and deployed efficiently in terms of memory consumption and computation time as in silico imitators of the neural populations –this is true in particular for deep learning models that can be accelerated by high performance parallel processors such as FPGAs and ASICs- , allowing this way to deploy “frozen” models of complex biological processors in applications such as bioinspired robotic control or neural population simulation for education in neuroscience.
References
-W. Pillow, L. Paninski, V. J. Uzzell, E. P. Simoncelli, and E. Chichilnisky. Prediction and decoding of retinal ganglion cell responses with a probabilistic spiking model. The Journal ofNeuroscience, 25(47):1100311013, 2005.
- J. I. Glaser, R. H. Chowdhury, M. G. Perich, L. E. Miller, and K. P. Kording, Machine learning for neural decoding, arXiv 1708.00909, 2017.
- A. Díaz-Tahoces, A. Martínez-Álvarez, A. Moll, L. Humphreys, J. Bolea, E. Fernández, Towards the reconstruction of moving images by populations of retinal ganglion cells, in Artificial Computation in Biology and Medicine, (Elche, Spain, 2015), pp. 220227.
- N. Parthasarathy et al., Neural Networks for Efficient Bayesian Decoding of Natural Images from Retinal Neurons, bioRxiv, Nov. 2017, p. 153759.
- E. Fernández, J. M. Ferrández, J. Ammermuller and R. Normann, Population coding in spike trains of simultaneously recorded retinal ganglion cells, Brain Research 887(1) (2000) 222-229.
-N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever and R. Salakhutdinov, Dropout: a simple way to prevent neural networks from overfitting, Journal of Machine Learning Research 15 (2014) 1929-1958.
Keywords:
deep learning,
neural networks,
neural coding and decoding,
MEA (MultiElectrode Arrays),
machine learning applied to neuroscience,
in vitro culture,
Convolutional Neural Networks,
recurrent neural networks
Conference:
MEA Meeting 2018 | 11th International Meeting on Substrate Integrated Microelectrode Arrays, Reutlingen, Germany, 4 Jul - 6 Jul, 2018.
Presentation Type:
Poster Presentation
Topic:
Neural Networks
Citation:
Lozano
A,
Calvo
MV,
Alegre-Cortés
J,
Álvarez-Sánchez
J,
De La Paz López
F,
Ferrandez
JM and
Fernandez
E
(2019). Towards a deep learning model of information encoding and decoding of in vitro neuronal cultures responses to electrical stimulation.
Conference Abstract:
MEA Meeting 2018 | 11th International Meeting on Substrate Integrated Microelectrode Arrays.
doi: 10.3389/conf.fncel.2018.38.00112
Copyright:
The abstracts in this collection have not been subject to any Frontiers peer review or checks, and are not endorsed by Frontiers.
They are made available through the Frontiers publishing platform as a service to conference organizers and presenters.
The copyright in the individual abstracts is owned by the author of each abstract or his/her employer unless otherwise stated.
Each abstract, as well as the collection of abstracts, are published under a Creative Commons CC-BY 4.0 (attribution) licence (https://creativecommons.org/licenses/by/4.0/) and may thus be reproduced, translated, adapted and be the subject of derivative works provided the authors and Frontiers are attributed.
For Frontiers’ terms and conditions please see https://www.frontiersin.org/legal/terms-and-conditions.
Received:
18 Mar 2018;
Published Online:
17 Jan 2019.
*
Correspondence:
PhD. Antonio Lozano, Universidad Politécnica de Cartagena, Electronics, Computer Technology and Projects, Cartagena, Spain, Spain, amlo.upct@gmail.com