Measurement, manipulation and modeling of brain-wide neural population dynamics

Abstract

Neural recording technologies increasingly enable simultaneous measurement of neural activity from multiple brain areas. To gain insight into distributed neural computations, a commensurate advance in experimental and analytical methods is necessary. We discuss two opportunities towards this end: the manipulation and modeling of neural population dynamics.

Fig. 1. Overview of neural dynamics and manipulations.

a Linear dynamical system using neural recordings. Binned spiking activity, y(t), relates to a latent neural population state, x(t), that evolves according to linear dynamics, A, with inputs from other cortical areas, Bu(t). The neural population state is typically a low-dimensional trajectory, and the dynamics can be conceptualized as a flow field. b In many cases, the neural state can be represented as a low-dimensional trajectory in a subspace of the higher-dimensional recordings. c In this subspace, the neural state evolves according to neural dynamics, which define a flow field. In a LDS, the dynamics of A can be contractive, expansive, rotational, or a fixed point. Inputs may cause the flow field to exhibit more complex motifs, such as shown in this panel. Generally, dynamics and dimensionality reduction can also be modeled to be nonlinear. d By electrical or optogenetic stimulation, or applying perturbations to the sensory-behavior or force-behavior relationship, it is possible to make perturbations to the neural state, x(t). These perturbations can be “within-manifold”, which perturbs the neural state along its natural modes, or they can be “outside-manifold,” which perturbs the neural state along dimensions outside the plane spanned by x¹ and x². e The neural dynamics (flow field, A) can also be perturbed, for example by applying pharmacology or lesioning the circuit. The neural state changes as a result of the changing dynamics (highlighted in green). Panels a and b are modified by J. C. Kao and appeared in Pandarinath and colleagues 2018.

Fig. 2. Using neural networks to model multi-area computation.

a Multi-area, brain-wide electrical recording and stimulation of neural activity is rapidly becoming possible, and these data require new analyses and modeling to provide new scientific insights and theories of neural computation. High-density NeuroPixel electrodes are shown inserted in four locations, and many additional insertions are possible in both cortical and subcortical regions. Each insertion can access several different brain areas (M). Thus the size of the data is proportional to N × M. The proportionality constant depends on the type of neural recordings, with full broadband data including low-frequency local field potentials (LFPs) and action potentials (APs, or spikes) requiring the fastest sampling. For stimulation, it is possible to stimulate arbitrary waveforms on each electrode in each area and thus there are enormous combinatorial possibilities. Single and two photon optical imaging of genetically encoded calcium indicators and voltage indicators, and optogenetic neural modulation, are also widely used but are not shown for simplicity. b Multi-area RNNs can be trained to model each of these insertions. Visual areas may be modeled via convolutional neural networks^, or related artificial networks that incorporate recurrence. Areas like the prefrontal and motor cortex are typically modeled by RNNs^,. Regularizations may be employed so that activity within an RNN area resembles those recorded from each electrode array recordings.