Improving scalability in systems neuroscience

SUMMARY

Emerging technologies to acquire data at increasingly greater scales promise to transform discovery in systems neuroscience. However, current exponential growth in the scale of data acquisition is a double-edged sword. Scaling up data acquisition can speed up the cycle of discovery but can also misinterpret the results or possibly slow down the cycle because of challenges presented by the curse of high-dimensional data. Active, adaptive, closed-loop experimental paradigms use hardware and algorithms optimized to enable time-critical computation to provide feedback that interprets the observations and tests hypotheses to actively update the stimulus or stimulation parameters. In this perspective, we review important concepts of active and adaptive experiments and discuss how selectively constraining the dimensionality and optimizing strategies at different stages of discovery loop can help mitigate the curse of high-dimensional data. Active and adaptive closed-loop experimental paradigms can speed up discovery despite an exponentially increasing data scale, offering a road map to timely and iterative hypothesis revision and discovery in an era of exponential growth in neuroscience.

INTRODUCTION

Systems neuroscience faces the daunting challenge of understanding brain networks of complex and poorly understood topologies. Over the past decade, however, a technological revolution in neuroscience has enabled tremendous growth in the volume and quality of scientific data. Experimental tools allowing measurements of large-scale in vivo neuronal population activity at high resolution using multiple (e.g., electrical, optical, magnetic) modalities and across multiple brain regions are becoming wide-spread. Alongside advances in instrumentation, methods to efficiently preprocess, characterize, and fit models to large-scale neuroscientific data are also being developed (Stevenson and Kording, 2011; Paninski and Cunningham, 2018). How should we use large-scale neurotechnologies to understand brain network mechanisms? In this perspective we review the state of the art in neurotechnology through the lens of the curse of high-dimensional neural data analysis. The curse of high-dimensional data arises from the consequences of scaling data dimensionality (Vershynin, 2018; Wainwright, 2019) and leads to exponentially increasing computation time. We propose that a new generalization of closed-loop experiments, which we term active, adaptive closed-loop (AACL) experiments, will be important to successfully mitigating the scalability in neuroscience, especially for discovering brain network mechanisms.

Discovery is a process of obtaining new knowledge based on multiple steps of verification. In systems neuroscience, knowledge can be expressed in multiple forms, including understanding the animal’s behavior, the effectiveness of experimental stimuli, the regularity of the neural response, and the causal link between neural codes and behavior. Discovery can revise existing theories or hypotheses or even create a paradigm shift in research practice. The standard discovery cycle involves data acquisition, analysis, and interpretation to test hypotheses and update concepts, which is fundamental to scientific progress (conceive-acquire-analyze-test-revise; Figure 1A). However, the concept of a “loop” is underemphasized in discovery cycle for two important reasons: first, there is no nested internal loop containing feedback, and second, there is no strict time constraint between steps. Large-scale neuroscience presents a challenge for the cycle of discovery. Counterintuitively, the growth of neuroscience data (in dimensionality, volume, and size) can slow and even impede the cycle of discovery. High dimensionality of data can overwhelm analysis because of the bandwidth bottleneck, and efforts to address the bottleneck can effectively decrease statistical power. The loss of statistical power occurs because unless simplification (such as averaging or subsampling) is assumed, statistical estimates of model variables may become increasingly biased, which may lead to further misinterpretation of results. The alternative involves increasing the recording duration to account for the increased dimensionality (thereby increasing acquisition time) or processing all collected data sequentially without time constraints (thereby increasing analysis time), either of which can slow down hypothesis testing or revision and progress toward scientific goals. This is particularly disruptive in neuroscience because the nervous system is dynamic and plastic. We cannot revise and test important classes of hypotheses, such as specificity and causality, until we verify the steps to complete an iteration. For instance, if the tested hypothesis is that circuit A is responsible for behavior B, even if we observe that a neurostimulation (as a perturbation of A) disrupts a task (B), we still need to verify several important questions to fully test the hypothesis: first, whether neurons collected from A show significantly changed patterns that correlate with the behavior; second, what types of neurons and how many of them contribute to such changes; third, how the specific stimulation input (e.g., timing and intensity) causally alters the task behavior. Much like how we cannot step into the same river twice, by the time the original hypothesis has been tested, the context within which the results were obtained cannot be revisited. This substantially limits, and can potentially even make impossible, our ability to test alternative hypotheses.

Figure 1. Scaling in neural data acquisition.

(A) Cycle of knowledge discovery (conceive-acquire-analyze-test-revise). The acquire step consists of recording large-scale neuronal activity during behavior. The analyze step consists of data analysis and interpretation. The test and revise steps involve testing the hypothesis and revising it as necessary to close the cycle.

(B) Near or faster than exponential growth in the number of recorded neurons on the basis of in vitro or in vivo electrophysiology (up-to-year update from https://stevenson.lab.uconn.edu/scaling/). Recent neurotechnology development for simultaneous neuronal recordings suggest a further jump from exponential growth.

(C) Trade-off between sampling frequency and the number of recorded neurons on the basis of microscopy imaging (from Lecoq et al., 2019). On the basis of the new light beads microscopy (LBM) technique, ~1,000,000 neurons were recorded within ~5.4 × 6 × 0.5 mm³ volumes at ~2 Hz (Demas et al., 2021).

(D) Schematic of “scale-speed limit” for data acquisition and analysis steps. For a fixed scale, the pace of discovery is determined by slowest scale-speed factor.

(E) Discovery rate (DR) scaling with data acquisition and analysis: DR = (number of discoveries/time per discovery). The number of discoveries scales proportionally to the scale of data acquisition. The time per discovery scales inversely proportionally to the rate at which data can be analyzed. Scaling the rate of data analysis with the rate of data acquisition should lead to an accelerating DR (blue curve). However, the CoD effectively slows down the discovery rate, and DR scaling requires a correction; this correction to the DR can dominate. If the CoD correction scales faster than the rate of data analysis, the otherwise accelerating DR flattens (red curve). Closed-loop experimental designs can mitigate the CoD and restore an accelerating DR (green shaded area).

AACL experiments are different from open-loop or passive closed-loop experiments in that not only are strict time constraints imposed on every step (e.g., acquisition, analysis, stimulation), but also an optimization procedure is used in some or even all steps on the basis of active feedback (Figure 2A). AACL feedback signals can be iteratively used for many purposes: to optimize experimental stimuli and other experimental parameters (Walker et al., 2019; Ponce et al., 2019), to select neural channels for recording and/or telemetry (Choi et al., 2020), to perform decoder adaptation (Dangi et al., 2013), to optimize stimulation parameters (Tafazoli et al., 2020), and to optimize objective functions and other aspects of control policies (Bolus et al., 2018). In principle, each step of an AACL experiment may contain nested inner loops. In contrast, passive closed-loop experiments use a fixed policy. The experimental stimulus is predetermined and cannot adapt. The decoder is fixed. Control and stimulation parameters are predetermined.

Figure 2. Active, adaptive closed-loop (AACL) experimental paradigms provide solutions to curse-of-dimensionality problems.

(A) Schematic of AACL experimental paradigm. Active feedback iteratively updates or optimizes the sampling, resampling, or stimulation parameters at each stage of the loop (marked by three dashed arrows). Feedback may have different timescales and forms: experimental design (Lewi et al., 2009, 2011), adaptive stimulus optimization (Walker et al., 2019; Ponce et al., 2019), neurally defined stimulation (Berényi et al., 2012; Paz et al., 2013; Grosenick et al., 2015; Zhang et al., 2021), closed-loop decoder adaptation (Dangi et al., 2013), user-defined control or prosthetics (Carmena et al., 2003; Shanechi et al., 2016), or adaptive closed-loop neurostimulation (Tafazoli et al., 2020).

(B) Schematic of adaptive stimulus optimization on the basis of BMI with neural decoder adaptation. In the closed-loop, an end-to-end trained neural network model predicted thousands of neuronal responses to arbitrary new natural input and synthesized optimal stimuli: most exciting inputs (MEIs) (adapted with permission from Walker et al., 2019).

(C) Illustration of subsampling/resampling in the closed-loop acquire step (Choi et al., 2020). Recording channels can be selected from signal channels by optimizing the subset of signal channels selected to maximize the number of recorded neurons given the available neurons known when all signal channels are recorded. Optimization involves adaptively sampling the signal channels selected to maximize the yield of recorded neurons.

(D) A snapshot of the closed-loop analyze step for large-scale rat hippocampal recordings. A GPU-powered population-decoding system was developed for ultrafast reconstruction of spatial positions from rodent’s unsorted spatiotemporal spiking patterns, with real-time speed to decode rat’s run position (latency < 250 ms) or memory replays (latency < 20 ms). Furthermore, the approach enabled assessment of the statistical significance of online-decoded replay (adapted with permission from Hu et al., 2018).

(E) Illustration of the scaling between the acquire and analyze steps to accommodate real-time operation (about a fraction of a millisecond per spike). The GPU decoding strategy can scale up to thousands of channels (adapted with permission from Hu et al., 2018).

(F) Schematic of a BMI with neural decoder adaptation (Dangi et al., 2013). The adaptation design elements include the adaptation, timescale, selective parameter adaptation, smooth decoder updates, and intuitive decoder adaptation parameters.

(G) Schematic of a BMI with neural feedback (adapted with permission from Yang et al., 2021). Model-based closed-loop controllers can be designed and consist of a Kalman state estimator and a feedback controller. Brain network activity can be used as feedback and the model-based closed-loop controller identified the stimulation parameters to drive the internal brain state to a particular target.

(H) Schematic of adaptive closed-loop stimulation (Tafazoli et al., 2020). The system learns to use multi-site spatially patterned electrical stimulation to control the pattern of activity of a population of neurons.

In this perspective, we discuss how jointly scaling up data acquisition and data analysis in an active and adaptive manner can speed up the cycle and enable AACL experiments. We first review scalability in neurotechnology and instrumentation, highlighting how multiple trends increase the size, volume, and dimensionality of experimental observations. We then point to our main thesis: that scaling is a double-edged sword, in that it can speed up the cycle of discovery in systems neuroscience, but it involves defining and following a sequence of predetermined experimental steps. In considering the cycle of discovery, each step in an AACL experiment has a respective “scaling-speed limit.” The overall rate of discovery is limited by the slowest factor in each step (Figure 1D), which can be exacerbated by large data volumes and high dimensionality, which can overwhelm our capacity for analysis and interpretation. Consequently, the lack of scalability of data analytic tools introduces barriers to scientific discovery. Finally, we discuss the features and limitations of AACL experiments and review strategies to speed up data analysis.

SIZE, DEPTH, AND MULTIPLE SITES IN NEUROPHYSIOLOGICAL AND IMAGING RECORDINGS

Neurotechnologies involve a range of physical modalities spanning sound, light, electricity, and magnetism, as well as multimodal mechanisms such as optoacoustics/photoacoustics and magnetoacoustics (Marblestone et al., 2013; Gottschalk et al., 2019). Modern neural interfaces that can record and/or stimulate the nervous system are dramatically expanding the number of neural signal channels that can be monitored and manipulated. When the word “scale” is used, we refer to the dimensionality, size, or volume of neural signals, which should not be confused with the spatial or temporal granularity at which data are acquired.

Accessing brain tissue at single-cell resolution has traditionally involved implanting electrodes directly into the brain. Multielectrode array recording devices remain the gold-standard approach to recording in vivo electrophysiological cellular activity (Hong and Lieber, 2019). Growth in the number of simultaneously recordable signal channels has been driven by electrode fabrication, packaging, materials, and implementation. Neuron density, brain area size, and tissue displacement due to wiring and other physical device properties impose fundamental limits on the number of recordable neurons. New neural recording technologies that exploit nanoscale features and integrated electronics are significantly increasing the number of single cells that can be recorded concurrently in single or multiple sites of the brain. Two fundamental factors are paving the way toward large-scale neurophysiology. One factor is to increase the number of electrodes/channels through advanced packaging or new materials (Scholvin et al., 2016). To date, hundreds to thousands of electrodes have been implanted to record animal brains (Figure 1B; Berényi et al., 2014; Shobe et al., 2015; Jun et al., 2017; Chung et al., 2019; Steinmetz et al., 2018). The other factor is to use three-dimensional (3D) electrode array technology, by combining laminar electrode and two-dimensional (2D) electrode arrays, for recording layer-specific areas in brain circuits (Hoogerwerf and Wise, 1994; Rios et al., 2016).

Although electrophysiology can collect the neural activity of local brain areas with high temporal resolution, various optical imaging techniques are targeted at whole-brain recordings, focusing on network and circuit levels (Yang and Yuste, 2017). Optical imaging also enables chronic recordings of large-scale neuronal activity in an animal’s brain across days and weeks (Jercog et al., 2016; Kim et al., 2016; Huang et al., 2018; Pachitariu et al., 2017; Weisenburger and Vaziri, 2018). Recently, modern technologies have also rapidly improved the spatiotemporal resolution and sampling speed of optical imaging and microscopy (Rumyantsev et al., 2020; Wu et al., 2020). Ultimately, physical constraints will impose a limit on the effectiveness of optical imaging, as any imaging techniques encounter the trade-off among imaging speed, field of view, and depth.

For all neural interfaces that rely on electrophysiology or optical imaging, technological factors constrain the number of signal channels that can be recorded or controlled simultaneously (Marblestone et al., 2013; Kleinfeld et al., 2019). These constraints involve, for example, power and thermal dissipation for implanted wireless arrays (Zhou et al., 2019), sampling frequency or optical paths for microscopes (Figure 1C; Tsai et al., 2015; Sofroniew et al., 2016; Stirman et al., 2016; Lecoq et al., 2019), and wiring constraints for electrode arrays (Marblestone et al., 2013; Hong and Lieber, 2019; Raducanu et al., 2017). Constraints on simultaneous access lead to a selection problem involving how to use the available signal channels to optimally monitor and manipulate the neural population of interest (Saxena and Cunningham, 2019; Moreaux et al., 2020). If there were no constraints, one could simply measure from all signal channels, and there would be no selection problem. If there were too many constraints, there would be very few simultaneously accessible signal channels, obviating the problem of selection. For most modern neurotechnologies, however, the space of possible selections is combinatorial. For example, Neuropixels electrode arrays contain 960 electrodes (Steinmetz et al., 2018); however, only 384 recording channels can be acquired simultaneously. Subject to other constraints, there are 2.5¹⁴⁹ different possible selections for this array (Choi et al., 2020). Similarly, the two-photon random-access mesoscope (2p-RAM) has a cellular-resolution microscope with a 5 mm field of view that makes available up to 1 million neurons in the transgenic mouse expressing GCaMP in neurons (Sofroniew et al., 2016). However, adaptive optics strategies are necessary to flexibly and rapidly deliver light and make available neurons for simultaneous investigation. For instance, a system using custom optics and independently repositionable temporally multiplexed imaging beams could offer an expanded field of view (>9.5 mm²), allowing multi-site imaging of tens of thousands of neurons across multiple mouse cortical areas (Stirman et al., 2016). In each aforementioned case, modern instrumentation leads to a combinatorial explosion of possible selections.

CHRONIC EXPERIMENTS, TASK COMPLEXITY, NATURALISTIC BEHAVIOR

Increasingly, modern neurotechnologies are being deployed chronically in implanted systems (Schwarz et al., 2014; Tybrandt et al., 2018; Chicang et al., 2020). The main concerns of chronic electrophysiological recordings are the unit yield, longevity, stability, and quality of neural signals (Juavinett et al., 2019; Luo et al., 2020). In all recording devices, the interfaces between the nervous system and a synthetic sensor involve innovations in advanced materials (Chen et al., 2017). Advanced microelectrode technologies have been invented for recording interfaces to improve biocompatibility and stability (Fattahi et al., 2014), which enable us to sample activity of the same population of neurons repeatedly. In addition, wireless recording devices have become increasingly available for chronic data acquisition.

The challenge of neuroscience data analysis is further magnified by the complexity of behavior. New technologies allow complex, naturalistic, and unconstrained behaviors to be measured with increasing detail at the individual and group levels (Tseng et al., 2018). Some behaviors, such as navigation, can involve multiple animals in social interactions (Danjo et al., 2018) or in 3D spaces (Omer et al., 2018). Skeletal movements involve joint rotations with as many as 27 different joint angles for the primate arm and hand. Other task behaviors, such as motor learning, can last hours, days, and even weeks (Sandler, 2008). As the temporal duration increases, the task complexity also scales up.

Naturalistic behavior also introduces additional issues. To be considered naturalistic, a behavior should not depend on training to follow experimenter-defined instructions. In the absence of instructions, however, preferred behaviors will be acquired, and behavioral stereotypy can emerge; namely, subjects can choose to repeatedly make the same, potentially optimal, action sequences, such as “look-then-reach” when picking up a cup. To more completely study the underlying neural mechanisms, naturalistic behaviors may involve adaptively delivering instructions in an AACL experiment. In active sampling behaviors (such as sniffing or shifting gaze), subjects actively use attention and active sensing strategies to sample relevant cues for information seeking or decision making. Although animals can learn a sampling policy through attentional learning and reward maximization, it poses a challenge for experimenters to study neural correlates underlying such behaviors.

SUBSAMPLING AND RESAMPLING OF NEURAL SPACE

As the number of simultaneously recorded neurons from electrophysiology or calcium imaging becomes very large (e.g., 10,000–1,000,000), redundancy will arise. Given a specific recorded brain target, identification of a high- or low-dimensional neural code will vary according to the question of interest. For instance, the visual cortex may have a high-dimensional representation for visual signals but a low-dimensional representation for other nonvisual behavioral variables (Stringer et al., 2019a, 2019b). Random sampling is a widely used statistical strategy for estimating the properties of a large network or system. Supported by the law of large numbers and distribution invariance, subsampling assumes exchangeability and ergodicity of a stochastic dynamical system. In data acquisition, large-scale sampling of neural signals enables us to examine the resampling axis in order to assess neural dimensionality and coding sufficiency. For instance, a theoretical question regarding the neural code is “What is the dimensionality of odor space?” (Meister, 2015), or “What is the intrinsic multi-neuronal dimensionality or the complexity of dynamics that relates to the task behavior?” (Gao and Ganguli, 2015; Gao et al., 2017). Unlike traditional data-replacement resampling techniques, sequential neural resampling opens the door to measuring neuronal populations in an integrated manner to generate datasets that are sufficient to rigorously test hypotheses about brain functions. Additionally, researchers may test whether subsampling of neuronal populations can preserve the invariant structure of network structure or neural dynamics (Chen et al., 2014; Williamson et al., 2016; Gao et al., 2017; Liu et al., 2019).

CURSE OF HIGH-DIMENSIONAL DATA ANALYSIS

The combination of task complexity, multimodality, and large-scale chronic experimental paradigms can quickly generate high-dimensional, structured neural and behavioral data whose analysis and interpretation can outpace computational capabilities. A statistical curse of dimensionality (CoD) arises to impede the discovery cycle within the analyze step.

The common theme of CoD problems is that when dimensionality increases, the volume of the space increases so rapidly that the available data become very sparse. For instance, to study d-dimensional behavioral variables, we design N experimental trials and record m neurons. If we increase d and m separately or jointly while keeping N unchanged, the insufficient sample size will make it difficult to relate a neural space R^m to a behavioral space R^d. In this case, in order to establish statistical significance, the number of samples (trial by duration) needed to support the result often grows exponentially with the dimensionalities d and m.

High-dimensional neural data impose a CoD across many statistical analyses. First, neural data analysis depends on second and higher order computations critical to understanding networks, such as functional connectivity. However, the number of trials and duration of trials needed for a reliable statistical estimate does not scale with data dimensionality. Statistical estimation of the covariance matrix in a principal-component analysis (PCA) can suffer strong bias and/or high variance when the sample size is insufficient given the data dimensionality (Box 1). Second, statistical estimation, by either a model-free or a model-based approach, can be ill posed when analyzing high-dimensional data. Although model-free approximations can have a small number of parameters, they may lack neuroscientific validity. In contrast, model-based approaches can involve many parameters but pose challenges for model fitting when the data are high dimensional. Therefore, incorporation of hypothesis-driven theories, priors, and constraints into the model may help solve ill-posed estimation problems. Dimensionality reduction techniques are important tools to tackle large-scale neural recordings on a single-trial basis (Box 2; Cunningham and Yu, 2014). Third, the complexity and long timescales of task behaviors will introduce plasticity or non-stationarity in neural recordings, posing additional estimation challenges.

Box 1. Correlation matrix estimation.

The correlation matrix is a central statistical measure central to principal-component analysis (PCA). As the dimensionality of observed variables, n, increases, the number of estimated parameters scales quadratically in n. A curse of dimensionality (CoD) arises when the sample size N is small in relation to the dimensionality, n, and the correlation structure of neuronal populations cannot be reliably estimated because of an insufficient number of experimental trials or duration in neuroscience experiments.

A simple model-free solution to the CoD imposes local proximity structure onto the correlations; namely, only neurons recorded from nearby electrodes are connected. Consequently, the number of parameters scales linearly with n. However, the spatial proximity assumption may not hold in practice. For instance, two hippocampal place cells recorded from distant electrodes may share overlapping place fields and hence correlated activity.

A model-based approach to the correlation matrix estimation is via partial correlation. Partial correlation is equivalent to conditional correlation when the random variables are multivariate normal distributed. If the observations are discrete (e.g., multinomial), the equivalence also holds when the conditional expectation of the random variables is linear (Baba et al., 2004). By using partial correlation, one can solve n linear regression problems, each of which involves n − 1 regressors and 1 predictor. Because the n regression problems can be solved independently, the computation can be scaled up using parallel computing. Therefore, by mitigating the CoD, partial correlation estimates may be not only more reliable, but also computationally efficient.

An alternative approach involves random projection or subsampling, in which one selects m variables (m ≪ n) and repeats the linear regression procedure with different subsets. Statistically, random projection-based correlation estimates assume sufficient sparsity in order to stably embed the subsets into a low-dimensional subspace. Therefore, the correlation estimates are robust with respect to varying number of neurons because of recording instability.

Box 2. Dimensionality reduction.

In systems neuroscience, dimensionality reduction methods are important to answering the neural dimensionality question: how many neurons are required to resolve the dynamics underlying a behavioral task? The answer may depend on the coding specificity of the stimulus or behavioral variables (Stringer et al., 2019a, 2019b). Knowing the answer can improve our understanding of the scaling property of neuronal population in both encoding and decoding (Williamson et al., 2016).

The choice of neural decoding methods also leads to dimensionality concerns. Linear decoding methods (such as factor analysis and Kalman filter) are commonly used because of their simplicity. In contrast, despite potentially better performance, many nonlinear decoding methods are less commonly used. One important reason is that nonlinear methods suffer a curse of dimensionality. For instance, nonlinear function estimation scales polynomially or even exponentially in terms of dimensionality. Moreover, fitting nonlinear functions requires parameter search in the presence of local minima which also scales with n. As a result, nonlinear methods often lack scalability.

Dimensionality reduction methods can help alleviate the computational curse (Cunningham and Yu, 2014). This has motivated the development of a variety of advanced nonlinear dimensionality reduction methods to examine neuronal population activities (Yu et al., 2009; Gao et al., 2016; Wu et al., 2017). However, nonlinear dimensionality reduction approaches are computationally expensive and depend on strong assumptions, such as the ability to conceptualize experimental measurement as a random projection of neural activity.

Adaptive subsampling provides a complementary approach to measurement by random projection and can address the CoD present for dimensionality reduction, which is critical in the context of closed-discovery-loop experiments. Because neurons exhibit log-normal firing rate distributions, applying dimensionality reduction methods to large numbers of neurons may not sufficiently capture the long-tail behavior. As a result, neuronal representations may be incompletely characterized. By allowing for better sampling in the tails, adaptive subsampling of neurons can provide a more complete picture.

Scaling data acquisition and analysis should accelerate the rate of discovery (Figure 1E). However, the curse of high-dimensional data exponentially increases the time necessary to obtain each discovery. As a result, the discovery rate may saturate as data acquisition and analysis increase in scale. The challenge is to maintain an increasing rate of discovery while increasing the scale of data acquisition and analysis. As we discuss below, AACL experiments may offer a solution.

AACL EXPERIMENTAL PARADIGMS

Closed-loop experiments represent a paradigm shift from open-loop experiments. In closed-loop experiments, neural signals are processed to algorithmically generate feedback signals that are delivered to the subject according to a policy (Zrenner et al., 2016; Yang and Shanechi, 2016; Ciliberti et al., 2018; Srinivasan et al., 2018; Kane et al., 2020; Bolus et al., 2018; Walketal., 2019; Ponce et al., 2019; Tafazoli et al., 2020; see also reviews in Potter et al., 2014; El Hady, 2016). Traditionally, feedback in closed-loop experiments can take a variety of forms. If the purpose of a brain-machine interface (BMI) is to control an external actuator, feedback can be the delivery of stimulation to the nervous system; if the goal of a BMI is to control sensory feedback, feedback can be the timing of sensorimotor information. However, in all closed-loop BMIs, data acquisition is subject to the signal bandwidth constraint, and analysis and feedback are subject to the time constraint. The timescale of feedback is often at the order of milliseconds or seconds that can map from circuit functions to behavior. Here we argue that passive closed-loop experiments are still insufficient and inefficient. Specifically, we introduce AACL experiments, which generalize concepts familiar to traditional closed-loop experimental designs and include active feedback that is based on multiple stages of knowledge discovery. The terms “active” and “adaptive” are subtly different yet often exchangeable in the literature. In an “active” design, the experimenter can manipulate the instrumentation or experimental stimuli according to a random or predefined policy. Unlike passive feedback, which arises automatically regardless of the user’s intention, active feedback emphasizes the effort of seeking valuable information from the feedback signal and then iteratively optimizes the discovery process at various stages (e.g., sampling, resampling analysis, stimulation). In an “adaptive” design, the experimenter can modify the decoder or stimulation parameters on the basis of feedback or error-correction learning.

AACL experiments enable testing of hypotheses that cannot be tested by non-AACL experiments in two distinct ways. Some hypotheses can, in principle, be tested using both AACL and non-AACL experiments, but non-AACL experiments are sufficiently inefficient that, in practice, the hypothesis cannot be tested because of lack of time. For example, hypotheses that depend on neurostimulation efficacy, which requires estimating a map of responses to stimulation at different stimulation sites. Other examples of new knowledge acquired in practice by AACL experiments include neuron-stimulus sensitivity, maximal electrode channel unit yields, and system controllability. Other hypotheses cannot be tested by non-AACL experiments, even in principle, and require the use of AACL experiments, such as hypotheses that depend on learning, especially when learning occurs rapidly and when learning is irreversible. When both AACL and non-AACL experiments can be performed in principle, the nature of the knowledge gained is similar, except that AACL experiments produce knowledge at a faster rate because of their improved efficiency. When AACL experiments cannot be performed by non-AACL experiments, the nature of knowledge gained is distinct.

To use neurostimulation again as an example, traditional closed-loop stimulation is designed with an on/off stimulation approach, with predetermined stimulation parameters. In contrast, AACL experiments can actively seek feedback from neuronal firing and adjust the stimulation parameters or control policy to optimize the “natural” cost function (Bolus et al., 2018; Tafazoli et al., 2020). The cost function is defined by the difference between the observed neural responses and predicted neural responses, where the predictor can be a simple linear-nonlinear Poisson (LNP) model or an artificial neural network.

We propose that AACL experiments offer a natural solution to the scaling bottleneck and improve the scalability. In contrast to the standard “conceive-acquire-analyze-test-revise” paradigm, which does not impose strict time constraints on each step, AACL experiments collect and analyze neural data in a sequential manner, with time constraints, and test adaptive hypotheses with timely neurofeedback that accounts for neural plasticity during the course of learning and adaptation (Figure 2A). The active and adaptive strategies can be implemented, independently or jointly, throughout the acquisition, analysis, and feedback steps. The form of feedback may be diverse, in terms of stimulus optimization, experimental design, decoder adaptation, neurostimulation, and other user-defined feedback control. The discovered knowledge accumulates with the completion of each step. The discovery cycle continues until the experimental subject reaches the predefined experimental goal algorithmically according to the policy. Notably, certain stages of the AACL experiments accommodate many other names proposed in the literature as special cases, such as active experimental design, active stimulus selection, closed-loop feedback control, and closed-loop decoder adaptation (CLDA). AACL experiments therefore generalize the concept of closed-loop experiments across timescales for closing the loop and iterating the discovery cycle, as quickly as a fraction of a second, to chronic experimental preparations, as long as months.

The concept of adaptive experiments is not new in neuroscience. For instance, design of adaptive experiments is a long-established standard for psychometric testing, such as the use of QUEST procedure (Watson and Pelli, 1983). At slower feedback timescales, iterative closed-loop paradigms are already well established in various domains within systems neuroscience. Neuronal stimulus selectivity in the ventral visual pathway exists in a high-dimensional space of sensory stimuli. To assess ventral stream selectivity, Quiroga et al. (2005) systematically searched for responses of single neurons to stimuli. The limited time available for experiments required closing the loop in two stages of correlational studies. First, responses in a screening session were analyzed and then used to select target stimuli for the testing session. Although each stage was an open-loop experiment, closing the feedback loop across stages tamed the explosion of potential experiments and made possible a more focused investigation. Dramatic increases in the number of neural signal channels that can be monitored and manipulated mean that neuroscience investigations increasingly lie within a high-dimensional space of experimental designs. These capabilities are opening the door to new applications of closed-loop experimental paradigms to map networks as part of large-scale investigations of multiregional communication (Box 3). To follow a similar philosophy but with improved efficiency, an analog of AACL experiment is to identify sensory stimuli that optimize visual neuronal responses at a fast, sub-second timescale. Specifically, Walker et al. (2019) developed “inception loops,” a closed-loop paradigm combining in vivo recording from thousands of neurons with in silico nonlinear response modeling. The closed-loop model-based response prediction enabled them to generate synthetic yet optimal stimuli (Figure 2B). Therefore, designing adaptive closed-loop image synthesis systems to explore the single or population neuronal response properties represents a new paradigm in visual neuroscience (Ponce et al., 2019; Bashivan et al., 2019).

Box 3. Mapping networks in neuroscience.

Brain function and dysfunction are increasingly understood as being due to the expression of multiple overlapping network mechanisms. Network mechanisms of multiregional communication are most often inferred from the structure of correlations in neural activity. The availability of recordings from many signal channels has fueled progress. However, functionally connectivity analyses have been typically applied to signals that measure neuronal function indirectly and do not necessarily scale because of fundamental limits on signal resolution, such as blood-oxygen-level-dependent (BOLD) functional magnetic resonance imaging (fMRI) and widefield calcium imaging signals. Inferring network mechanisms from high-dimensional neuronal recordings is hampered by the CoD. Moreover, using correlations to interpret activity patterns as being due to interactions is subject to significant confounds. Correlations are sensitive to the confounding influence of common inputs from other brain regions, yielding network edges even when the receiver does not receive any input from the sender.

Recent work maps large-scale brain networks and studies the mechanisms of multiregional communication by recording neural responses while delivering low-amplitude stimulation pulses in a causal network analysis (Qiao et al., 2020). Taking a causal sampling approach offers important advantages. Causal responses cannot be due to common input. Delivering isolated low-amplitude microstimulation pulses also offers the opportunity to more directly probe network excitability while avoiding the confounding effects and network responses. Inferences from large-amplitude stimulation pulses or pulse trains may recruit network responses that do not reflect direct functional interactions between the stimulation and recording sites (Lozano et al., 2019). Large amplitude pulses and pulse trains can effectively change the interaction instead of measuring the interaction.

Because mapping networks using a causal network analysis allows a selective targeting of neurons and neural circuits for investigation on the basis of their role in the network, we may be able to mitigate the curse of dimensionality associated with scaling up data acquisition and analysis without constraints.

Closed-loop BMIs can not only learn optimizing sensory stimuli but also learn active sensing strategies (Richardson et al., 2019). Specifically, experimental manipulation of task-relevant sensory feedback, provided by intracortical microstimulation (ICMS) that encodes egocentric bearing to the hidden goal direction, can reveal distinct motor strategy adaptation to match novel sensor properties for goal-directed behavior. Additionally, BMIs seek to deliver either neural feedback by stimulating neural activity (SENSE-STIMULATE) or user feedback through an external interface that the user controls (SENSE-CONTROL). In neural feedback BMIs, subjects do not need to be aware of the operation of the interface. The BMI seeks to disrupt ongoing network excitation or inhibition (e.g., seizure control or optogenetic control; Berényi et al., 2012; Paz et al., 2013; Grosenick et al., 2015) and/or shape neural plasticity (e.g., mood regulation; Zhang et al., 2021; Shanechi, 2019). In contrast, user feedback BMIs (e.g., visual and motor prostheses) depend on how the user learns to use the interface (Carmena et al., 2003; Koralek et al., 2012; Shenoy and Carmena, 2014). Another example of AACL experiment is CLDA used in BMI systems, which can accelerate learning and improve performance by iteratively updating a BMI decoder’s parameter (Dangi et al., 2013; Figure 2F). In these cases, volition, awareness, and agency play important roles as the subject controls the relevant patterns of neural activity decoded by the BMI. In principle, neural feedback-based and user feedback-based BMIs can be combined. For example, BMIs based on feedback that the user controls could also feature neural feedback protocols designed to recruit brain plasticity and enhance learning (Shenoy and Carmena, 2014).

BMIs offer clinical opportunities as neuroprosthetic devices (Collinger et al., 2013; Ajiboye et al., 2017). Additionally, BMIs provide a novel experimental platform for performing adaptive perturbations and causal circuit manipulations. One successful AACL application is to use an adaptive closed-loop patterned stimulation strategy (Tafazoli et al., 2020), which learns to use multi-site electrical stimulations to control the pattern of a population of neurons. Additionally, BMIs can help reveal important circuit mechanisms and are particularly useful when studying learned behaviors and sensorimotor control (Jarosiewicz et al., 2008; Koralek et al., 2012; Sadtler et al., 2014; Golub et al., 2016). By making explicit the system inputs and outputs, BMIs allow us to resolve the neural computations that drive learning and test how network structure influences learning (Orsborn and Pesaran, 2017). As BMIs require low-latency feedback, they can also be used with causal circuit manipulations to stimulate or inactivate in a state-dependent manner. State dependence allows manipulations to be sensitive to the dynamic properties of brain processes (Qiao et al., 2020) and is an essential component of closed-loop feedback control algorithms (Shanechi et al., 2016; Srinivasan et al., 2018; Yang et al., 2021). Therefore, BMIs can enable us to conditionally test specific causal functional roles for neural circuits or their plausible links to behaviors.

FEATURES AND LIMITATIONS OF AACL EXPERIMENTS

Discovery does not rely on closed-loop experiments per se; neither do closed-loop experiments automatically lead to discovery in neuroscience. However, AACL experiments can provide timely feedback and update new hypotheses iteratively during the course of discovery process.

High-dimensional capabilities enabled by modern neurotechnologies present not only opportunities in establishing the links between neuronal activity and behavior but also challenges and paradigm shifts in neural data analysis and interpretation. Traditional neuroscience paradigms based on spike sorting and tuning curve estimation will inevitably fail to capture the complexity and dynamics of naturalistic behaviors because the behaviors occupy high-dimensional spaces. AACL experiments offer opportunities to perform “active” experimental designs that algorithmically select experimental parameters from a high-dimensional space of configurations. In traditional “passive” experimental designs, each step of the cycle has a predetermined policy. For instance, the acquire step uses the fixed stimulus configurations, whereas in the analyze step, the stimulation configuration or control strategy is fixed. Active experimental designs feature adaptive selection strategies that optimize each step in a closed-loop using real-time neurofeedback. For instance, animal training can be optimized (Bak et al., 2016), experimental stimulus design can be optimized in a sequential manner (Lewi et al., 2009, 2011); and feedback control or neurostimulation can be optimized on the fly (Cunningham et al., 2011; Swann et al., 2018). As a result, we can efficiently test hypotheses sequentially and potentially even in parallel.

The challenge presented by high-dimensional experimental configurations is particularly acute in the case of neurostimulation experiments. Unlike neural recordings, which can be performed at multiple sites simultaneously, neurostimulation experiments can be performed at only one area by choosing “when,” “how,” and “where” to stimulate. The resulting spatiotemporal patterns of stimulation occupy a particularly high-dimensional configuration space, which cannot necessarily be probed simultaneously. In general, only a relatively small number of configurations can be tested in a single experimental session. As the nervous system is adaptive and plastic, with constantly changing neural responses, we cannot necessarily rely on comparing stimulation responses to different configurations in different sessions. Novel AACL experimental designs will be critical for progress.

An important approach features AACL experiments with active designs to guide neuronal subsampling and resampling (Figure 2C). A central issue in these experimental designs is whether the properties of the repeatedly sampled populations reflect properties of the underlying distribution. Closed-loop acquisition is like an active search in the space of neural activity to maximize the signal-to-noise ratio. Unlike active sensing in behavior that reflects the animal’s behavioral policy, neuronal subsampling is guided by the experimenter’s policy, subject to physical, time, and bandwidth constraints. For instance, we can design an algorithm that optimizes the joint electrode selections for all recording channels according to the experimenter’s policy in order to maximize the isolation quality of detected neurons (Choi et al., 2020). Analyzing resampled populations is very effective when performing dimensionality reduction. Subsampling m neurons from a population of n neurons can be viewed as a random projection from an n-dimensional manifold (Ganguli and Sompolinsky, 2012; Gao and Ganguli, 2015); in this sense, resampling can be viewed as multiple random projections of n neurons. The Johnson-Lindenstrauss lemma states that random projections preserve the pairwise distances of high-dimensional data (Bingham and Mannila, 2001). As a result, properties of the underlying distribution that depend on pairwise distances, such as in dimensionality reduction techniques, are preserved by resampled populations. Modes estimated from resampled populations may share other distributional properties with the underlying population. Notably, neural activity often follows a log-dynamic law (Buzsáki and Mizuseki, 2014), and linear combinations of subsampled lognormal distributed neural responses can also be approximated by a lognormal distribution (Asmussen and Rojas-Nandayapa, 2008).

Causality is the holy grail in all systems neuroscience inquiries. It is important to distinguish between correlation and causation in closed-loop experiments. Correlational dependencies describe associations of measurements that experiments do not control, whereas causal dependencies link a dependent variable to an experimentally controlled variable (Jazayeri and Afraz, 2017). The key concept in causal inference is randomization, such as a random external stimulus or random perturbation (e.g., microstimulation or optogenetic stimulation). The relationship between every dependent variable and the randomized variable is causal, whereas the relationship between non-randomized variables and behavior remains correlational. As brain activity is high dimensional, correlations within massively under-sampled neuronal recordings cannot fully reveal circuit mechanisms. Although closed-loop experiments can contain both correlation and causation components, they can be distinguished from open-loop perturbation experiments in timing and specificity, thereby narrowing the search space of neural code-behavior relationship, that is, mitigating the dimensionality bottleneck.

It is also important to point out the limitations of closed-loop perturbation for causal dissection of circuit and behavior. First, the brain is complex, and many brain areas can engage in even a simple task or spontaneous behavior (Stringer et al., 2019). Therefore, even large-scale neural recordings can only provide a small window of the brain activity, and our target system is partially observable. The presence of latent variables can bring additional degree of complexity to precisely control the variable of interest (either the neural activity or behavior); the induced unexpected network-level side effects will complicate the data interpretation. Second, the brain is nonlinear and plastic; a control strategy that works in a certain condition will not necessarily generalize well to other tasks or behavioral states. Third, the behavior may also be complex (although the dimensionality of behavior is much smaller than the neural dimensionality), and each axis of behavioral space may be jointly or independently controlled by neural correlates. Fourth, electrical or optical stimulation may create undesirable lasting side effects (e.g., heat, toxicity, cell death, change in excitation-inhibition balance) that further bring uncertainties to specific brain functions under study. Finally, even the most sophisticated neural stimulation technologies available today suffer the limitation that they artificially activate or suppress neural activity. Such manipulations may highjack the system or drive the neuronal network into “unnatural” regimes. Many BMI technologies based on non-specific and unnatural perturbations may face the challenge of result interpretation, as traditional non-adaptive neurostimulations may cause varying degrees of side effects on behavior or even generate “false discoveries.” To mitigate this concern, recent work performing causal network analyses emphasizes the use of minimal perturbations to deliver single pulses (Qiao et al., 2020). Other work on multiregional network system identification shows that pairs of sites that share correlated neural activity patterns also tend to share stimulation responses (Yang et al., 2021), indicating that causal manipulations and recordings can be mutually informative and constrain network inferences. While realizing the fundamental limit of causality test in the brain, trial-and-error active manipulations remain the most important source of evidence that a brain area supports one aspect of behavior. Development of next-generation active and adaptive BMIs that deliver “naturalistic” patterned neurostimulations and incorporate sufficient control experiments would help alleviate the serious issue of false discovery. One type of new closed-loop feedback for neurostimulation, for instance, could be the output of neuronal firing patterns or the local network connectivity (Vlachos et al., 2016; Bolus et al., 2018; Tafazoli et al., 2020).

SPEEDING UP NEURAL DATA ANALYSIS

Another important issue of scalability in knowledge discovery is the speed. Even if the dimensionality of data remains constant, the increasing amount of data still creates an analysis bottleneck for knowledge discovery. In data analysis and interpretation, we aim to avoid exponential complexity or computation latency with respect to the number of neurons, seeking a linear or sub-linear order of complexity. In developing efficient analytic tools, computation speed and scalability are the key considerations.

First, closed-loop BMIs impose low-latency constraints in all steps of experiments. The computational overhead jointly depends on data size, CPU architecture, memory, and bandwidth. Overall, the computation latency is composed of two parts: total cost = fixed cost + scaling cost, where the first term is independent of the scaling, and the second term increases with the scale of data. Therefore, scaling up data acquisition can impose a great challenge in speed for computation because of limited resources in memory, bandwidth, and computing power. To accommodate scalable ultrafast neural data analysis, modern computing resources and dedicated hardware can help meet these resource requirements. According to the estimate of the doom of Moore’s law, the computing power of a single CPU will similarly reach its physical limit around 2022. In contrast, high-performance computing devices based on graphics processing units (GPUs) and field-programmable gate arrays (FPGAs) have become widely adopted for data analysis (Hu et al., 2018; Giovannucci et al., 2018).

In the acquire step, open-source low-latency hardware (e.g., Open Ephys 2.0) has managed to minimize high-speed sampling delay with microsecond latency. Automated and scalable hardware-empowered spike sorting can accommodate real-time processing for large-scale data acquisition (Pachitariu et al., 2016; Chung et al., 2017; Jun et al., 2017; Yger et al., 2018). In the analyze step, computational tasks can be operated in real time for ultrafast decoding, detection, and control. Using the rodent hippocampus as an example, hippocampal replays during sharp-wave ripples are known to contribute to memory consolidation, planning, and future decision making (Buzsáki, 2015). Closed-loop perturbation experiments that are designed to investigate the contribution of these replay events may narrow the search space of relationship between neural code and behavior or eliminate alternative competing hypotheses (Girardeau et al., 2009; Fernández-Ruiz et al., 2019). Therefore, it is important to develop scalable methods that enable real-time decoding and assessment of these hippocampal replay contents to match the complexity of neural data, in the form of large-scale unsorted ensemble spikes (Ciliberti et al., 2018; Hu et al., 2018; Figures 2D and 2E), high-density field potential recordings (Cao et al., 2020; Frey et al., 2019), or large-scale calcium imaging (Tu et al., 2020). To take the primate motor cortex as another example, closed-loop BMIs have provided mechanistic insight into learning, plasticity, and functional reorganization (Jarosiewicz et al., 2008; Sadtler et al., 2014; Shenoy and Carmena, 2014). Development of scalable methods for decoding arm or hand movement or assessing neural population dynamics can greatly advance the research field in motor control (Trautmann et al., 2019; Sussillo et al., 2016). The key component of BMIs is the feedback, as a form of neurostimulations (Berényi et al., 2012; Paz et al., 2013; Grosenick et al., 2015; Zhang et al., 2021), user-defined feedback control (Figure 2G; Carmena et al., 2003; Dangi et al., 2013; Shanechi et al., 2016), or the prediction error of neural responses (Figure 2H; Tafazoli et al., 2020), which can be further used to perturb the circuit or causally change the behavior. Finally, the time window of closed-loop feedback is critical, as it allows interaction with neurons and circuits differently. Sub-millisecond feedback stimulation may prevent recurrent inhibition, but the same setup with an order of second delay may affect the system in a completely different manner. These uncertainties of mechanistic inquiry grow in time, especially when the casual chain between the cause (stimulation) and effect is long. Therefore, the timing imposes a strict low-latency constraint on closed-loop BMIs (Müller et al., 2013; Kane et al., 2020).

Second, scalable data-intensive computation demands fast and efficient computing strategies. Even though real-time operation may not be always required, offline processing of high-throughput, high-dimensional neural data can still be prohibitive, which may include neural-behavior mapping (Vogelstein et al., 2014), large-scale model fitting, data visualization, and computer simulations. For instance, structural data are fundamentally high dimensional, including 2D images, 3D volumes, and four-dimensional (4D) and five-dimensional (5D) hypervolumes for multi-spectral data. Large-scale neural circuit mapping may require both structural and functional data (Shi et al., 2015). High-performance computing is required to analyze high-resolution, high-throughput neuroanatomy and neuroimaging data. Behavioral data can be also high dimensional, especially when they are captured via high-fidelity video recordings. Data-intensive, automated image segmentation and 3D morphological reconstruction have been empowered by powerful deep learning methods for behavioral video or imaging analyses (Mathis et al., 2018; Pereira et al., 2019; Zhou et al., 2018; Arac et al., 2019). Another source of high dimensionality arises from multimodal measurements, such as concurrent electroencephalographic (EEG)/magnetoencephalographic (MEG) source localization (Antelis and Minguez, 2013). Furthermore, large-scale biologically inspired neuronal network modeling and computer simulations may leverage high-performance GPU or FPGA computing (Hoang et al., 2013; Sripad et al., 2018). Finally, distributed data analytics platform and computing infrastructure can help achieve fast and scalable data analysis of massive size (Freeman et al., 2014; Freeman 2015).

Third, artificial intelligence (AI) and machine learning can help accelerate the pace of neuroscience discoveries (Marblestone et al., 2016; Richards et al., 2019; Cichy and Kaiser, 2019) and scaling up of innovation (Kittur et al., 2019). On one hand, AI and deep learning can help or automate complex and large-scale neural data analyses to uncover the patterns in brain activities. For instance, in neural encoding, deep learning can help link complex patterns of neural activity and/or cortical anatomy to complex behavior (Minderer et al., 2019; Pandarinath et al., 2018), as well as control neuronal spiking or internal brain states (Bashivan et al., 2019). On the other hand, neuroscience can drive AI forward for knowledge discovery; neuroscience-inspired AI has achieved professional human-level intelligence for playing chess and computer games (Silver et al., 2016; Hassabis et al., 2017). AlphaGo, motivated from deep reinforcement learning, discovered a remarkable level of Go knowledge through a self-taught training process. Therefore, brain-inspired deep learning architectures can not only provide a new computational framework for brain information processing (Kriegeskorte, 2015; Banino et al., 2018) but also generate new insight in systems neuroscience and provide rapid theoretical and experimental progress (Richards et al., 2019).

CONCLUDING THOUGHTS: SCALING TO THE HUMAN BRAIN

Currently, our understandings of brain mechanisms in animal models and in the human brain are separated by a divide. This is due partly to the additional ethical, safety and efficacy, and financial constraints that govern the development of neurotechnologies for use in humans. Nevertheless, neurotechnologies are increasingly making possible studies of the human brain. The vast scale and complexity of the human brain inevitably means that understanding how to jointly scale data acquisition and data analysis will play an essential role in progress. To date, high-density biocompatible and stretchable electrode grids can record spikes and local field potentials (LFPs) at the surface of human brain (Khodagholy et al., 2015; Tybrandt et al., 2018). Scaling up data acquisition via high-density interfaces may further improve the spatiotemporal resolution of human brain mechanisms (Robinson et al., 2017; Matsushita et al., 2018; Escabí et al., 2014; Even-Chen et al., 2020; Sohrabpour et al., 2020). Concurrent multimodal and multi-site recordings, neuroimaging, and neurostimulation will also drive progress (Chang, 2015; Krook-Magnuson et al., 2015; Swann et al., 2018). Basic brain mechanisms will play a role in our understanding of the diseased brain. In translational or therapeutic applications, closed-loop human BMI systems have been widely adopted for restoring or enhancing sensory, motor, or cognitive brain functions, as well as delivering anesthesia drug (Shanechi, 2019; Moses et al., 2019; Gilja et al., 2015; Liberman et al., 2013; Yang and Shanechi, 2016). As a result, AACL experimental or adaptive BMI paradigms may have significant impacts on human brain science. The development of stable, secure, real-time brain-cloud interfaces similar to current mobile voice- and image-based interfaces will be critical to updating model-based inferences on the basis of new observations (Martins et al., 2019).

In summary, knowledge discovery in systems neuroscience is being transformed by advances in neurotechnology. Fundamentally, the scale of data acquisition and speed of data analysis jointly determine the rate of hypothesis testing or revision and ultimately the rate of discovery. The peril arises from how scaling up data acquisition slows down data analysis. AACL experiments offer a solution to improve scalability for knowledge discovery. Achieving this vision requires the coordination of scalable computation and active and adaptive experimental designs in real-time systems and interfaces. Ultimately, the successful scaling of knowledge discovery is essential to understand the complex brain mechanisms supporting cognition and behavior in health and disease.