Computational models in the age of large datasets

Abstract

Technological advances in experimental neuroscience are generating vast quantities of data, from the dynamics of single molecules to the structure and activity patterns of large networks of neurons. How do we make sense of these voluminous, complex, disparate and often incomplete data? How do we find general principles in the morass of detail? Computational models are invaluable and necessary in this task and yield insights that cannot otherwise be obtained. However, building and interpreting good computational models is a substantial challenge, especially so in the era of large datasets. Fitting detailed models to experimental data is difficult and often requires onerous assumptions, while more loosely constrained conceptual models that explore broad hypotheses and principles can yield more useful insights.

From Whole-Brain Data to Functional Circuit Models: The Zebrafish Optomotor Response

“…Such flexibility is common in multiparameter models when correlated parameter changes compensate to avoid error (Fisher et al, 2013; O’Leary et al, 2015). Individual connections could compensate in our model (Figure 7C, left), so we identified connectivity patterns that independently affect model accuracy (Figure 7C, right).…”

“…This will be common going forward, as neuroscientists typically characterize high-dimensional systems using a few stimuli and behaviors (Fisher et al, 2013; Gao and Ganguli, 2015; O’Leary et al, 2015). Nevertheless, a few combined parameters were critical to explain the behavior, thereby distilling the behaviorally relevant features of the population response.…”

Neural Population Dynamics during Reaching Are Better Explained by a Dynamical System than Representational Tuning

“…Furthermore, RNNs provide an ideal medium for more detailed study, as the ground truth of synaptic connectivity, plasticity, noise, trial-to-trial variability, and responses to unexpected perturbations are known and can be manipulated directly. However, “exploring an artificial model universe comes with its own risk” [41] and proper models must resist the temptation of explaining purely idiosyncratic properties, but rather those that are able to explain large amounts of variance in electrophysiological data. Our results also emphasize that explaining a large amount of variance in neural data in and of itself does not necessary lead to mechanistic insight [42], as the observation of rotational structure arose under multiple models, and future work is needed to determine the biological circuit mechanism underlying population level rotational structure.…”

Training deep neural density estimators to identify mechanistic models of neural dynamics

“…24 25 29 computational model that incorporates the mechanisms we believe to be at play, based on scientific knowledge, 30 intuition, and hypotheses about the components of a system and the laws governing their relationships. The goal of 31 such mechanistic models is to investigate whether a proposed mechanism can explain experimental data, uncover 32 details that may have been missed, inspire new experiments, and eventually provide insights into the inner workings 33 of an observed neural or behavioral phenomenon [1][2][3][4]. Examples for such a symbiotic relationship between model 34 and experiments range from the now classical work of Hodgkin and Huxley [5], to population models investigating 35 rules of connectivity, plasticity and network dynamics [6-10], network models of inter-area interactions [11,12], and 36 models of decision making [13, 14].A crucial step in building a model is adjusting its free parameters to be consistent with experimental observations.…”