Bayesian inference of structured latent spaces from neural population activity with the orthogonal stochastic linear mixing model.

The brain produces diverse functions, from perceiving sounds to producing arm reaches, through the collective activity of populations of many neurons. Determining if and how the features of these exogenous variables (e.g., sound frequency, reach angle) are reflected in population neural activity is important for understanding how the brain operates. Often, high-dimensional neural population activity is confined to low-dimensional latent spaces. However, many current methods fail to extract latent spaces that are clearly structured by exogenous variables. This has contributed to a debate about whether or not brains should be thought of as dynamical systems or representational systems. Here, we developed a new latent process Bayesian regression framework, the orthogonal stochastic linear mixing model (OSLMM) which introduces an orthogonality constraint amongst time-varying mixture coefficients, and provide Markov chain Monte Carlo inference procedures. We demonstrate superior performance of OSLMM on latent trajectory recovery in synthetic experiments and show superior computational efficiency and prediction performance on several real-world benchmark data sets. We primarily focus on demonstrating the utility of OSLMM in two neural data sets: µECoG recordings from rat auditory cortex during presentation of pure tones and multi-single unit recordings form monkey motor cortex during complex arm reaching. We show that OSLMM achieves superior or comparable predictive accuracy of neural data and decoding of external variables (e.g., reach velocity). Most importantly, in both experimental contexts, we demonstrate that OSLMM latent trajectories directly reflect features of the sounds and reaches, demonstrating that neural dynamics are structured by neural representations. Together, these results demonstrate that OSLMM will be useful for the analysis of diverse, large-scale biological time-series datasets.

Data-driven distillation and precision prognosis in traumatic brain injury with interpretable machine learning.

Traumatic brain injury (TBI) affects how the brain functions in the short and long term. Resulting patient outcomes across physical, cognitive, and psychological domains are complex and often difficult to predict. Major challenges to developing personalized treatment for TBI include distilling large quantities of complex data and increasing the precision with which patient outcome prediction (prognoses) can be rendered. We developed and applied interpretable machine learning methods to TBI patient data. We show that complex data describing TBI patients' intake characteristics and outcome phenotypes can be distilled to smaller sets of clinically interpretable latent factors. We demonstrate that 19 clusters of TBI outcomes can be predicted from intake data, a ~ 6× improvement in precision over clinical standards. Finally, we show that 36% of the outcome variance across patients can be predicted. These results demonstrate the importance of interpretable machine learning applied to deeply characterized patients for data-driven distillation and precision prognosis.

Resolving Non-identifiability Mitigates Bias in Models of Neural Tuning and Functional Coupling.

In the brain, all neurons are driven by the activity of other neurons, some of which maybe simultaneously recorded, but most are not. As such, models of neuronal activity need to account for simultaneously recorded neurons and the influences of unmeasured neurons. This can be done through inclusion of model terms for observed external variables (e.g., tuning to stimuli) as well as terms for latent sources of variability. Determining the influence of groups of neurons on each other relative to other influences is important to understand brain functioning. The parameters of statistical models fit to data are commonly used to gain insight into the relative importance of those influences. Scientific interpretation of models hinge upon unbiased parameter estimates. However, evaluation of biased inference is rarely performed and sources of bias are poorly understood. Through extensive numerical study and analytic calculation, we show that common inference procedures and models are typically biased. We demonstrate that accurate parameter selection before estimation resolves model non-identifiability and mitigates bias. In diverse neurophysiology data sets, we found that contributions of coupling to other neurons are often overestimated while tuning to exogenous variables are underestimated in common methods. We explain heterogeneity in observed biases across data sets in terms of data statistics. Finally, counter to common intuition, we found that model non-identifiability contributes to bias, not variance, making it a particularly insidious form of statistical error. Together, our results identify the causes of statistical biases in common models of neural data, provide inference procedures to mitigate that bias, and reveal and explain the impact of those biases in diverse neural data sets.

DL-TODA: A Deep Learning Tool for Omics Data Analysis.

Metagenomics is a technique for genome-wide profiling of microbiomes; this technique generates billions of DNA sequences called reads. Given the multiplication of metagenomic projects, computational tools are necessary to enable the efficient and accurate classification of metagenomic reads without needing to construct a reference database. The program DL-TODA presented here aims to classify metagenomic reads using a deep learning model trained on over 3000 bacterial species. A convolutional neural network architecture originally designed for computer vision was applied for the modeling of species-specific features. Using synthetic testing data simulated with 2454 genomes from 639 species, DL-TODA was shown to classify nearly 75% of the reads with high confidence. The classification accuracy of DL-TODA was over 0.98 at taxonomic ranks above the genus level, making it comparable with Kraken2 and Centrifuge, two state-of-the-art taxonomic classification tools. DL-TODA also achieved an accuracy of 0.97 at the species level, which is higher than 0.93 by Kraken2 and 0.85 by Centrifuge on the same test set. Application of DL-TODA to the human oral and cropland soil metagenomes further demonstrated its use in analyzing microbiomes from diverse environments. Compared to Centrifuge and Kraken2, DL-TODA predicted distinct relative abundance rankings and is less biased toward a single taxon.

BigNeuron: a resource to benchmark and predict performance of algorithms for automated tracing of neurons in light microscopy datasets.

BigNeuron is an open community bench-testing platform with the goal of setting open standards for accurate and fast automatic neuron tracing. We gathered a diverse set of image volumes across several species that is representative of the data obtained in many neuroscience laboratories interested in neuron tracing. Here, we report generated gold standard manual annotations for a subset of the available imaging datasets and quantified tracing quality for 35 automatic tracing algorithms. The goal of generating such a hand-curated diverse dataset is to advance the development of tracing algorithms and enable generalizable benchmarking. Together with image quality features, we pooled the data in an interactive web application that enables users and developers to perform principal component analysis, t-distributed stochastic neighbor embedding, correlation and clustering, visualization of imaging and tracing data, and benchmarking of automatic tracing algorithms in user-defined data subsets. The image quality metrics explain most of the variance in the data, followed by neuromorphological features related to neuron size. We observed that diverse algorithms can provide complementary information to obtain accurate results and developed a method to iteratively combine methods and generate consensus reconstructions. The consensus trees obtained provide estimates of the neuron structure ground truth that typically outperform single algorithms in noisy datasets. However, specific algorithms may outperform the consensus tree strategy in specific imaging conditions. Finally, to aid users in predicting the most accurate automatic tracing results without manual annotations for comparison, we used support vector machine regression to predict reconstruction quality given an image volume and a set of automatic tracings.

The Neurodata Without Borders ecosystem for neurophysiological data science.

The neurophysiology of cells and tissues are monitored electrophysiologically and optically in diverse experiments and species, ranging from flies to humans. Understanding the brain requires integration of data across this diversity, and thus these data must be findable, accessible, interoperable, and reusable (FAIR). This requires a standard language for data and metadata that can coevolve with neuroscience. We describe design and implementation principles for a language for neurophysiology data. Our open-source software (Neurodata Without Borders, NWB) defines and modularizes the interdependent, yet separable, components of a data language. We demonstrate NWB's impact through unified description of neurophysiology data across diverse modalities and species. NWB exists in an ecosystem, which includes data management, analysis, visualization, and archive tools. Thus, the NWB data language enables reproduction, interchange, and reuse of diverse neurophysiology data. More broadly, the design principles of NWB are generally applicable to enhance discovery across biology through data FAIRness.

The brain is an immensely complex organ which regulates many of the behaviors that animals need to survive. To understand how the brain works, scientists monitor and record brain activity under different conditions using a variety of experimental techniques. These neurophysiological studies are often conducted on multiple types of cells in the brain as well as a variety of species, ranging from mice to flies, or even frogs and worms. Such a range of approaches provides us with highly informative, complementary 'views' of the brain. However, to form a complete, coherent picture of how the brain works, scientists need to be able to integrate all the data from these different experiments. For this to happen effectively, neurophysiology data need to meet certain criteria: namely, they must be findable, accessible, interoperable, and re-usable (or FAIR for short). However, the sheer diversity of neurophysiology experiments impedes the 'FAIR'-ness of the information obtained from them. To overcome this problem, researchers need a standardized way to communicate their experiments and share their results ­ in other words, a 'standard language' to describe neurophysiology data. Rübel, Tritt, Ly, Dichter, Ghosh et al. therefore set out to create such a language that was not only FAIR, but could also co-evolve with neurophysiology research. First, they produced a computer software program (called Neurodata Without Borders, or NWB for short) which generated and defined the different components of the new standard language. Then, other tools for data management were created to expand the NWB platform using the standardized language. This included data analysis and visualization methods, as well as an 'archive' to store and access data. Testing the new language and associated tools showed that they indeed allowed researchers to access, analyze, and share information from many different types of experiments, in organisms ranging from flies to humans. The NWB software is open-source, meaning that anyone can obtain a copy and make changes to it. Thus, NWB and its associated resources provide the basis for a collaborative, community-based system for sharing neurophysiology data. Rübel et al. hope that NWB will inspire similar developments across other fields of biology that share similar levels of complexity with neurophysiology.

Columnar Localization and Laminar Origin of Cortical Surface Electrical Potentials.

Electrocorticography (ECoG) methodologically bridges basic neuroscience and understanding of human brains in health and disease. However, the localization of ECoG signals across the surface of the brain and the spatial distribution of their generating neuronal sources are poorly understood. To address this gap, we recorded from rat auditory cortex using customized µECoG, and simulated cortical surface electrical potentials with a full-scale, biophysically detailed cortical column model. Experimentally, µECoG-derived auditory representations were tonotopically organized and signals were anisotropically localized to less than or equal to ±200 µm, that is, a single cortical column. Biophysical simulations reproduce experimental findings and indicate that neurons in cortical layers V and VI contribute ∼85% of evoked high-gamma signal recorded at the surface. Cell number and synchrony were the primary biophysical properties determining laminar contributions to evoked µECoG signals, whereas distance was only a minimal factor. Thus, evoked µECoG signals primarily originate from neurons in the infragranular layers of a single cortical column.SIGNIFICANCE STATEMENT ECoG methodologically bridges basic neuroscience and understanding of human brains in health and disease. However, the localization of ECoG signals across the surface of the brain and the spatial distribution of their generating neuronal sources are poorly understood. We investigated the localization and origins of sensory-evoked ECoG responses. We experimentally found that ECoG responses were anisotropically localized to a cortical column. Biophysically detailed simulations revealed that neurons in layers V and VI were the primary sources of evoked ECoG responses. These results indicate that evoked ECoG high-gamma responses are primarily generated by the population spike rate of pyramidal neurons in layers V and VI of single cortical columns and highlight the possibility of understanding how microscopic sources produce mesoscale signals.

NeuroGPU: Accelerating multi-compartment, biophysically detailed neuron simulations on GPUs.

BACKGROUND: The membrane potential of individual neurons depends on a large number of interacting biophysical processes operating on spatial-temporal scales spanning several orders of magnitude. The multi-scale nature of these processes dictates that accurate prediction of membrane potentials in specific neurons requires the utilization of detailed simulations. Unfortunately, constraining parameters within biologically detailed neuron models can be difficult, leading to poor model fits. This obstacle can be overcome partially by numerical optimization or detailed exploration of parameter space. However, these processes, which currently rely on central processing unit (CPU) computation, often incur orders of magnitude increases in computing time for marginal improvements in model behavior. As a result, model quality is often compromised to accommodate compute resources. NEW METHOD: Here, we present a simulation environment, NeuroGPU, that takes advantage of the inherent parallelized structure of the graphics processing unit (GPU) to accelerate neuronal simulation. RESULTS & COMPARISON WITH EXISTING METHODS: NeuroGPU can simulate most biologically detailed models 10-200 times faster than NEURON simulation running on a single core and 5 times faster than GPU simulators (CoreNEURON). NeuroGPU is designed for model parameter tuning and best performs when the GPU is fully utilized by running multiple (> 100) instances of the same model with different parameters. When using multiple GPUs, NeuroGPU can reach to a speed-up of 800 fold compared to single core simulations, especially when simulating the same model morphology with different parameters. We demonstrate the power of NeuoGPU through large-scale parameter exploration to reveal the response landscape of a neuron. Finally, we accelerate numerical optimization of biophysically detailed neuron models to achieve highly accurate fitting of models to simulation and experimental data. CONCLUSIONS: Thus, NeuroGPU is the fastest available platform that enables rapid simulation of multi-compartment, biophysically detailed neuron models on commonly used computing systems accessible by many scientists.

The impact of reducing signal acquisition specifications on neuronal spike sorting.

Measuring electrical potentials in the extracellular space of the brain is a popular technique because it can detect action potentials from putative individual neurons. Electrophysiology is undergoing a transformation where the number of recording channels, and thus number of neurons detected, is growing at a dramatic rate. This rapid scaling is paving the way for both new discoveries and commercial applications; however, as the number of channels increases there will be an increasing need to make these systems more power efficient. One area ripe for optimization are the signal acquisition specifications needed to detect and sort action potentials (i.e., "spikes") to putative single neuron sources. In this work, we take existing recordings collected using Intan hardware and modify them in a way that corresponds to reduced recording performance. The accuracy of these degraded recordings to spike sort using MountainSort4 is evaluated by comparing against expert labels. We show that despite reducing signal specifications by a factor of 2 or more, spike sorting accuracy does not change substantially. Specifically, reducing both sample rate and bit depth from 30 kHz and 16 bits to 12 kHz and 12 bits resulted in a 3% drop in spike sorting accuracy. Our results suggest that current neural acquisition systems are over-specified. These results may inform the design of next generation neural acquisition systems enabling higher channel count systems.

Critical Point-Finding Methods Reveal Gradient-Flat Regions of Deep Network Losses.

Despite the fact that the loss functions of deep neural networks are highly nonconvex, gradient-based optimization algorithms converge to approximately the same performance from many random initial points. One thread of work has focused on explaining this phenomenon by numerically characterizing the local curvature near critical points of the loss function, where the gradients are near zero. Such studies have reported that neural network losses enjoy a no-bad-local-minima property, in disagreement with more recent theoretical results. We report here that the methods used to find these putative critical points suffer from a bad local minima problem of their own: they often converge to or pass through regions where the gradient norm has a stationary point. We call these gradient-flat regions, since they arise when the gradient is approximately in the kernel of the Hessian, such that the loss is locally approximately linear, or flat, in the direction of the gradient. We describe how the presence of these regions necessitates care in both interpreting past results that claimed to find critical points of neural network losses and in designing second-order methods for optimizing neural networks.

Plant single-cell solutions for energy and the environment.

Progress in sequencing, microfluidics, and analysis strategies has revolutionized the granularity at which multicellular organisms can be studied. In particular, single-cell transcriptomics has led to fundamental new insights into animal biology, such as the discovery of new cell types and cell type-specific disease processes. However, the application of single-cell approaches to plants, fungi, algae, or bacteria (environmental organisms) has been far more limited, largely due to the challenges posed by polysaccharide walls surrounding these species' cells. In this perspective, we discuss opportunities afforded by single-cell technologies for energy and environmental science and grand challenges that must be tackled to apply these approaches to plants, fungi and algae. We highlight the need to develop better and more comprehensive single-cell technologies, analysis and visualization tools, and tissue preparation methods. We advocate for the creation of a centralized, open-access database to house plant single-cell data. Finally, we consider how such efforts should balance the need for deep characterization of select model species while still capturing the diversity in the plant kingdom. Investments into the development of methods, their application to relevant species, and the creation of resources to support data dissemination will enable groundbreaking insights to propel energy and environmental science forward.

Improved inference in coupling, encoding, and decoding models and its consequence for neuroscientific interpretation.

BACKGROUND: A central goal of systems neuroscience is to understand the relationships amongst constituent units in neural populations, and their modulation by external factors, using high-dimensional and stochastic neural recordings. Parametric statistical models (e.g., coupling, encoding, and decoding models), play an instrumental role in accomplishing this goal. However, extracting conclusions from a parametric model requires that it is fit using an inference algorithm capable of selecting the correct parameters and properly estimating their values. Traditional approaches to parameter inference have been shown to suffer from failures in both selection and estimation. The recent development of algorithms that ameliorate these deficiencies raises the question of whether past work relying on such inference procedures have produced inaccurate systems neuroscience models, thereby impairing their interpretation. NEW METHOD: We used algorithms based on Union of Intersections, a statistical inference framework based on stability principles, capable of improved selection and estimation. COMPARISON: We fit functional coupling, encoding, and decoding models across a battery of neural datasets using both UoI and baseline inference procedures (e.g., ℓ1-penalized GLMs), and compared the structure of their fitted parameters. RESULTS: Across recording modality, brain region, and task, we found that UoI inferred models with increased sparsity, improved stability, and qualitatively different parameter distributions, while maintaining predictive performance. We obtained highly sparse functional coupling networks with substantially different community structure, more parsimonious encoding models, and decoding models that relied on fewer single-units. CONCLUSIONS: Together, these results demonstrate that improved parameter inference, achieved via UoI, reshapes interpretation in diverse neuroscience contexts.

Robust, automated sleep scoring by a compact neural network with distributional shift correction.

Studying the biology of sleep requires the accurate assessment of the state of experimental subjects, and manual analysis of relevant data is a major bottleneck. Recently, deep learning applied to electroencephalogram and electromyogram data has shown great promise as a sleep scoring method, approaching the limits of inter-rater reliability. As with any machine learning algorithm, the inputs to a sleep scoring classifier are typically standardized in order to remove distributional shift caused by variability in the signal collection process. However, in scientific data, experimental manipulations introduce variability that should not be removed. For example, in sleep scoring, the fraction of time spent in each arousal state can vary between control and experimental subjects. We introduce a standardization method, mixture z-scoring, that preserves this crucial form of distributional shift. Using both a simulated experiment and mouse in vivo data, we demonstrate that a common standardization method used by state-of-the-art sleep scoring algorithms introduces systematic bias, but that mixture z-scoring does not. We present a free, open-source user interface that uses a compact neural network and mixture z-scoring to allow for rapid sleep scoring with accuracy that compares well to contemporary methods. This work provides a set of computational tools for the robust automation of sleep scoring.

Deep learning as a tool for neural data analysis: Speech classification and cross-frequency coupling in human sensorimotor cortex.

A fundamental challenge in neuroscience is to understand what structure in the world is represented in spatially distributed patterns of neural activity from multiple single-trial measurements. This is often accomplished by learning a simple, linear transformations between neural features and features of the sensory stimuli or motor task. While successful in some early sensory processing areas, linear mappings are unlikely to be ideal tools for elucidating nonlinear, hierarchical representations of higher-order brain areas during complex tasks, such as the production of speech by humans. Here, we apply deep networks to predict produced speech syllables from a dataset of high gamma cortical surface electric potentials recorded from human sensorimotor cortex. We find that deep networks had higher decoding prediction accuracy compared to baseline models. Having established that deep networks extract more task relevant information from neural data sets relative to linear models (i.e., higher predictive accuracy), we next sought to demonstrate their utility as a data analysis tool for neuroscience. We first show that deep network's confusions revealed hierarchical latent structure in the neural data, which recapitulated the underlying articulatory nature of speech motor control. We next broadened the frequency features beyond high-gamma and identified a novel high-gamma-to-beta coupling during speech production. Finally, we used deep networks to compare task-relevant information in different neural frequency bands, and found that the high-gamma band contains the vast majority of information relevant for the speech prediction task, with little-to-no additional contribution from lower-frequency amplitudes. Together, these results demonstrate the utility of deep networks as a data analysis tool for basic and applied neuroscience.

Sparse, Predictive, and Interpretable Functional Connectomics with UoILasso.

Network formation from neural activity is a foundational problem in systems neuroscience. Functional networks, after downstream analysis, can provide key insights into the nature of neurobiological structure and computation. The validity of such insights hinges on accurate selection and estimation of the edges connecting nodes. However, commonly used statistical inference procedures generally fail to identify the correct features, and further introduce consequential bias in the estimates. To address these issues, we developed Union of Intersections (UoI), a flexible, modular, and scalable framework for enhanced statistical feature selection and estimation. Methods based on UoI perform feature selection and feature estimation through intersection and union operations, respectively. In the context of linear regression (specifically UoILasso), we summarize extensive numerical investigation on synthetic data to demonstrate tight control of false-positives and false-negatives in feature selection with low-bias and low-variance estimates of selected parameters, while maintaining high-quality prediction accuracy. We demonstrate, with UoILasso, the extraction of sparse, predictive, and interpretable functional networks from human electrocorticography recordings during speech production and the inference of parsimonious coupling models from nonhuman primate single-unit recordings during reaching tasks. Our results establish that UoILasso generates interpretable and predictive functional connectivity networks.

Laminar origin of evoked ECoG high-gamma activity.

High-gamma (Hγ) activity from electrocorticography (ECoG) is a common-used signal for understanding the human brain, but its interpretation is impeded by a lack of spatial localization. To address this, we developed a novel recording approach to simultaneously record µECoG cortical surface electrical potentials (CSEPs) and laminar multiunit activity (MUA). We demonstrate that stimulus evoked CSEPs carry a multi-modal frequency response, peaking in the Hγ range. Laminar MUA responses exhibited similar tuning to CSEP Hγ directly over the intracortical recording site, suggesting a functional relationship. We fit CSEP Hγ to the simultaneously-recorded laminar MUA using a state-of-the-art sparse multi-linear regression model to identify laminar contributions to cortical surface Hγ. Our results indicate that CSEP Hγ recorded by ECoG reflects spiking activity from neurons in layer 3. These results provide initial insight into localizing the sources of CSEPs, which will guide clinical and BMI device decisions.

Human Sensorimotor Cortex Control of Directly Measured Vocal Tract Movements during Vowel Production.

During speech production, we make vocal tract movements with remarkable precision and speed. Our understanding of how the human brain achieves such proficient control is limited, in part due to the challenge of simultaneously acquiring high-resolution neural recordings and detailed vocal tract measurements. To overcome this challenge, we combined ultrasound and video monitoring of the supralaryngeal articulators (lips, jaw, and tongue) with electrocorticographic recordings from the cortical surface of 4 subjects (3 female, 1 male) to investigate how neural activity in the ventral sensory-motor cortex (vSMC) relates to measured articulator movement kinematics (position, speed, velocity, acceleration) during the production of English vowels. We found that high-gamma activity at many individual vSMC electrodes strongly encoded the kinematics of one or more articulators, but less so for vowel formants and vowel identity. Neural population decoding methods further revealed the structure of kinematic features that distinguish vowels. Encoding of articulator kinematics was sparsely distributed across time and primarily occurred during the time of vowel onset and offset. In contrast, encoding was low during the steady-state portion of the vowel, despite sustained neural activity at some electrodes. Significant representations were found for all kinematic parameters, but speed was the most robust. These findings enabled by direct vocal tract monitoring demonstrate novel insights into the representation of articulatory kinematic parameters encoded in the vSMC during speech production.SIGNIFICANCE STATEMENT Speaking requires precise control and coordination of the vocal tract articulators (lips, jaw, and tongue). Despite the impressive proficiency with which humans move these articulators during speech production, our understanding of how the brain achieves such control is rudimentary, in part because the movements themselves are difficult to observe. By simultaneously measuring speech movements and the neural activity that gives rise to them, we demonstrate how neural activity in sensorimotor cortex produces complex, coordinated movements of the vocal tract.

Sparse coding of ECoG signals identifies interpretable components for speech control in human sensorimotor cortex.

The concept of sparsity has proven useful to understanding elementary neural computations in sensory systems. However, the role of sparsity in motor regions is poorly understood. Here, we investigated the functional properties of sparse structure in neural activity collected with high-density electrocorticography (ECoG) from speech sensorimotor cortex (vSMC) in neurosurgical patients. Using independent components analysis (ICA), we found individual components corresponding to individual major oral articulators (i.e., Coronal Tongue, Dorsal Tongue, Lips), which were selectively activated during utterances that engaged that articulator on single trials. Some of the components corresponded to spatially sparse activations. Components with similar properties were also extracted using convolutional sparse coding (CSC), and required less data pre-processing. Finally, individual utterances could be accurately decoded from vSMC ECoG recordings using linear classifiers trained on the high-dimensional sparse codes generated by CSC. Together, these results suggest that sparse coding may be an important framework and tool for understanding sensory-motor activity generating complex behaviors, and may be useful for brain-machine interfaces.

Timing during transitions in Bengalese finch song: implications for motor sequencing.

To investigate mechanisms of action sequencing, we examined the relationship between timing and sequencing of syllables in Bengalese finch song. An individual's song comprises acoustically distinct syllables organized into probabilistic sequences: a given syllable potentially can transition to several different syllables (divergence points), and several different syllables can transition to a given syllable (convergence points). In agreement with previous studies, we found that more probable transitions at divergence points occur with shorter intersyllable gaps. One intuition for this relationship is that selection between syllables reflects a competitive branching process, in which stronger links to one syllable lead to both higher probabilities and shorter latencies for transitions to that syllable vs. competing alternatives. However, we found that simulations of competitive race models result in overlapping winning-time distributions for competing outcomes and fail to replicate the strong negative correlation between probability and gap duration found in song data. Further investigation of song structure revealed strong positive correlation between gap durations for transitions that share a common convergent point. Such transitions are not related by a common competitive process, but instead reflect a common terminal syllable. In contrast to gap durations, transition probabilities were not correlated at convergence points. Together, our data suggest that syllable selection happens early during the gap, with gap timing determined chiefly by the latency to syllable initiation. This may result from a process in which probabilistic sequencing is first stabilized, followed by a shortening of the latency to syllables that are sung more often.NEW & NOTEWORTHY Bengalese finch songs consist of probabilistic sequences of syllables. Previous studies revealed a strong negative correlation between transition probability and the duration of intersyllable gaps. We show here that the negative correlation is inconsistent with previous suggestions that timing at syllable transitions is governed by a race between competing alternatives. Rather, the data suggest that syllable selection happens early during the gap, with gap timing determined chiefly by the latency to syllable initiation.