I am a Ph.D. student in his second year at Georgia State University where I work under the supervision of Andrey Shilnikov on real-time optimizing of mathematical neurons. I am attempting to incorporate machine learning and data science techniques to assist specifically electrophysiological neuroscientists in classifying their behavior and optimizing a mathematical neuron to recreate said behavior.
Data science and machine learning are hot topics that are sometimes confused with one another. Like computational neuroscience, data science is an interdisciplinary field that combines statistics, data analysis, and machine learning methods to understand phenomena. Machine learning, on the other hand, is the use of algorithms to perform a specific task relying on previously accumulated data patterns. I highly recommend any aspiring academic to look into these differences and areas as their importance will only grow with the continued aggregation of data between academics.
For my research, I have used logistic regressions to determine the correct parameter space for alpha synapses and iterative optimization algorithms to fit a system of differential equations that simulate the voltage behavior of biological neurons. Both of which I’ve presented on my poster at the 2019 DS SIAM conference. There is a number of mathematical neuron simulations that are made of a system of differential equations. The neuron model I use is from the Plant model that was modified by my lab cohort Deniz Alacam. Dr. Alacam modified the Plant model in a simple but powerful way, namely adding bifurcation parameters to the equilibrium potentials of two ionic currents. These bifurcation parameters allow us to change the bursting nature of the neuron, switching from endogenous bursting to tonic spiking to quiescent. However, I’d like to also speak about my future research revolving around bursting neurons.
In neuroscience, there are two kinds of bursting neurons, endogenous and network bursters. Endogenous bursters have periods of rapid firing that are followed by quiescent periods in isolation. In contrast, network bursters are neurons that are configured in a way that causes them to output a bursting pattern. To an observing electrophysiologist, the voltage output of both can look remarkably similar if not impossible to differentiate. This difficulty can at best result in the electrophysiologist undertaking time-consuming methods to determine the nature of the burster and at worst cause the electrophysiologist to assume its nature without further information. I am aiming and proposing to create a machine-learning algorithm to automate the burster nature determination in real-time.
Those familiar with machine learning will recognize this a classification problem. One machine learning technique that is used for classification is, ironically, artificial neural networks (ANN). The creation of ANNs is inspired by biological neural networks however they are not necessarily identical in their function. These artificial networks learn to perform a task by adjusting the output of the artificial network to the examples fed to the network. This method is also what inspired the initial iterative optimizing algorithm that I presented in my SIAM poster. For my ANN I will train it using previously classified and verified neural recordings of both endogeneous and network bursters.
As an advocate and enthusiast for machine learning, I would like to end with a word caution on machine learning. Specifically, the perceived objectivity of artificial intelligence. This is a dangerous assumption that could lead to the further systemic marginalization of the disenfranchised if these AIs are rolled out prematurely. Because we are removing human decision making, the general public that is unfamiliar with the pitfalls of AI will falsely view the decisions made by AI as objective and free of human error. However as IBM states on their research website “AI systems are only as good as the data we put into them.” Bad data create bad AI because bad data contain implicit racial, gender, or ideological biases. I would go further and state that any unanalyzed or controlled data are bad. I hate to put a damper on the intrepid journey of AI engineers and scientists but a little caution could refine and accelerate the AI advancement.