Neural Networks

Whole-brain connectomes are increasingly used as wiring diagrams for spiking simulations and as structural priors for artificial networks, usually after an ad hoc synapse-count threshold whose effect on validated function has never been measured. We define function-preserving sparsification - removing edges under a fixed budget while preserving validated behaviour and population responses - and study it in the leaky integrate-and-fire whole-brain model of Shiu et al. (2024). Six operators - weak-synapse pruning, uniform and weight-proportional random removal, degree-preserving rewiring, weight shuffling and a task-aware criterion - were compared at identical edge budgets across three validated circuits (sugar-evoked proboscis extension, its bitter suppression, grooming pathway specificity), silent and spontaneously active network states, and multiple seeds. Weak-synapse pruning preserves 79% of feeding output with 6% of edges, with a cliff at k* of approximately 9 synapses, whereas random, degree-preserving and weight-shuffled controls at the same budget abolish output, and weight-proportional random sampling survives only while it retains the strong-edge set: function resides in the conjunction of wiring placement and strength, not in degree statistics or synaptic mass. The cliff is task-dependent - grooming specificity collapses below k = 2 through disinhibition. In the active state failure changes character: output over-excites instead of vanishing, and bitter suppression is lost progressively from k = 5 and reverses at k = 31. A task-aware criterion preserves its training stimulus at 0.4% of edges but loses held-out inhibition first. These results provide a first benchmark for connectome sparsification and show that what must be kept depends on task and network state.

The brain frequently updates or controls its internal representations of sensory or motor variables over time. An example is transsaccadic updating of stimuli's retinotopic positions: When the eye moves in one direction, the positions are updated continuously in the opposite direction, and the process is controlled by the corollary discharge (CD) of the saccade motor command. We previously trained neural networks to perform such updating using the desired time courses of stimuli's positions at every time step. However, this may not be biologically plausible because, before a system learns to perform the task, it may not have access to complete time courses and the actual visual inputs are delayed and thus incorrect. Here we show that by adding the continuity equation to the loss, we can train neural networks with only stimuli's pre- and post-saccadic retinotopic positions (without specifying the intermediate transitions), and with heavily down-sampled time courses. Moreover, both the connectivity patterns and the CD control signal can be learned simultaneously. Since the continuity equation can be applied to waves of neural population activities that are distributed representations of variables, our work also suggests an advantage of distributed representations which are ubiquitous in the brain.

Large-scale Neuropixels recordings sample many brain areas, but no single experimental session records every area of interest. NeuroPaint, a transformer-based autoencoder, integrates such recordings by aligning area-specific latent dynamics across sessions and "inpainting" activity in unrecorded areas. Could a linear model, which would be computationally efficient and potentially more interpretable, suffice to achieve this? To address this question, we introduce Linear NeuroPaint, a linear variant of NeuroPaint, and evaluate it on two multi-area Neuropixels datasets. Linear NeuroPaint achieves strong predictive performance on recorded areas, however, it is less accurate than nonlinear models at cross-area inpainting, despite extensive hyperparameter optimization. Moreover, its parameters are not readily interpretable: incomplete cross-session alignment leads parameter-defined neuron clusters to exhibit heterogeneous response profiles. Together, these results show that linear transformations are insufficient for cross-session alignment and cross-area prediction. Nevertheless, Linear NeuroPaint provides a computationally efficient baseline for benchmarking multi-session neural models.

Competitive spiking networks can form sensory representations through local synaptic plasticity and lateral inhibition. A canonical example is the Diehl-Cook architecture, in which Poisson-encoded sensory inputs drive excitatory neurons through spike-timing-dependent plasticity (STDP), while a separate inhibitory population imposes competition between excitatory units. In most implementations this competitive circuit is fixed, although inhibitory synapses are plastic in cortical circuits and contribute to excitation-inhibition balance. It therefore remains unclear under which conditions biologically motivated inhibitory plasticity can replace fixed lateral competition without disrupting representation learning. Here we study this question in a competitive STDP network trained on MNIST. We compare fixed inhibition with a Vogels-style inhibitory STDP rule, slow homeostatic inhibition, and a budget-constrained slow homeostatic rule that conserves the total inhibitory input received by each excitatory neuron. A fixed-inhibition operating-point map showed that stability depends on the dynamical regime: 56% of tested fixed-inhibition runs completed training, whereas stronger inhibition at higher input gain produced runaway activity. The reference regime g = 2.0, wIE = 10.0, selected for subsequent comparisons, combined complete stability across replicate initializations with high classification performance (84.3 {+/-} 0.6% after 30,000 training examples). In this regime, the budget-constrained rule was stable in all replicate runs when its inhibitory budget matched the fixed-inhibition reference (BI = 3990), reaching 83.5 {+/-} 0.6% accuracy with firing-rate and inhibitory-conductance statistics close to fixed inhibition, whereas a mismatched budget (BI = 7980) produced runaway activity in every replicate. Parameter maps showed that Vogels-derived iSTDP was broadly but not uniformly stable (90% runs completed), while slow homeostatic inhibition was even more broadly stable (96%). Budget-constrained stability depended strongly on the conserved budget, i.e., low and matched budgets were stable or nearly stable, intermediate budgets were mixed, and BI = 7980 failed in all tested runs. These results show that inhibitory plasticity in competitive STDP networks is not governed by a local homeostatic update alone. Stable plastic competition depends jointly on the operating regime and on an inhibitory budget that fixes the total strength of competition while allowing local inhibitory weights to adapt. Mechanistically, budget matching separates local redistribution of inhibitory weights from changes in total inhibitory drive, preventing plasticity from shifting the network away from a stable competitive state. This identifies inhibitory budget matching as a practical design principle for replacing fixed inhibition with plastic, homeostatically regulated competition in STDP-based spiking networks.

We present a network model that enables a navigating agent to shape the spatial and temporal structure of its paths through its environment. The model is inspired by the insect fan-shaped body. In it, a recurrent neural network receives an input that encodes an agent's velocity vector. Meanwhile, network activity determines the setpoint of a control loop that influences this velocity vector through motor outputs. The network thus guides future movement using memory of past movement. We illustrate how the network's dynamics determine the trajectory features it extracts to produce the setpoint. We then relate the joint dynamics of network activity and the agent's velocity vector to the structure of the resulting paths. This network can produce paths suitable for different forms of exploration. It can also stabilize movements initiated by lower-level systems. Our analyses illustrate how recurrent neural dynamics and memory can shape behavior when embedded in closed loops.

Emerging research on desynchronization reveals strong relations to memory formation and recollection with much of the evidence originating from studies of the hippocampus, a brain region central to memory processing. However, growing evidence emphasizing the role of desynchronization is accompanied by challenges, most notably the synchronization/desynchronization conundrum. The conundrum arises from an apparent contradiction between learning and Shannon's information theory: spike-timing-dependent plasticity (STDP) acquires information via synchronized populational activity whereas information theory suggests that information channel is encoded through higher variability. Both frameworks are well-founded. Learning is widely accepted as the mathematical abstraction of STDP while Shannon's information theory showcases its power ubiquitously across fields from communication systems to neuroscience. Here, we propose that learning and Shannon's information theory are not necessarily opposing principles. Rather, learning plays a critical role in the process of synchronization and desynchronization. To justify our claim, we conduct computational simulations based on a Hodgkin-Huxley neuronal model coupled with NMDA, AMPA, and GABA synapses with inputs of biologically realistic spiking data. Our results show that enabling STDP significantly influences synchronization and desynchronization dynamics. Specifically, we find that STDP and desynchronization form a regulatory loop, in which STDP regulates the level of synchrony, and synchrony regulates subsequent learning strength. Further study reveals STDP is capable of switching neurons from synchronization to desynchronization. This framework mitigates the synchronization/desynchronization conundrum by clarifying the relationship between desynchronization and learning, and offers a new perspective on how synchrony contributes to memory formation and recollection.

Large-scale biophysically-detailed models of the brain are instrumental in unraveling how cellular, synaptic, and circuit-level mechanisms contribute to neural dynamics. Such models are built around large populations of single-neuron models whose morphologies must preserve physiological variability to reproduce integrative properties and firing patterns observed experimentally. Experimental reconstruction of morphologies is a time-consuming and resource-intensive process, unlikely to supply them at the required scale. Several data-driven methods for morphology synthesis have been developed, which nonetheless rely on extensive datasets for parameter estimation. To address this limitation, we developed NeuWalk, to our knowledge the first public Python library for morphology synthesis based on branching-and-annihilating, biased-and-correlated random walks. It estimates branching and annihilation rates from Sholl intersections and, optionally, bifurcation counts, while also ensuring their replication in synthetic morphologies; user-specified combinations of biases, implementing somatofugality, self-avoidance, spatial competition, and branching-angle control, sculpt morphological patterns characteristic of each neuron class. We then validated the method on mitral and middle tufted cells of the olfactory bulb, semilunar and pyramidal cells of the anterior piriform cortex, and neocortical pyramidal neurons: means and variances of Sholl intersections, bifurcation counts, and total dendritic length were statistically indistinguishable between synthetic and natural morphologies, with the single exception of its variance in apical dendrites of piriform cortex pyramidal neurons. As Sholl intersections and bifurcation counts are commonly reported in the literature, and biases can be inferred from published images or qualitative descriptions, NeuWalk requires no immediate access to experimental reconstructions, striking a trade-off between data-driven parameterization and empiricism without sacrificing morphological accuracy.

The dentate gyrus' principal neuron, the granule cell (GC), plays a key role in pattern separation, transforming similar inputs into dissimilar outputs. We used a new measure based on information theory to evaluate this transformation in an experimentally derived in silico GC model with realistic ion channel distributions across its dendrites. We found that model performance aligned well with current-clamp recordings from mouse brain slices. However, it was also possible to find artificial parameter combinations in the model that functionally outperformed the recorded cells while matching their action potential shapes. Rather than maximising computational power, the biologically realistic GC model balanced energy efficiency against increased functionality in a Pareto optimal way: No possible parameter combination was found to improve one of the objectives without impairing the other. Interestingly, immature GCs adopted a different weighted trade-off with a stronger emphasis on energy efficiency than pattern separation, but they remained Pareto optimal over their putative developmental trajectory. Our results support the idea that evolution favours neurons that can satisfy functional and energetic demands simultaneously.

Hippocampal sharp-wave ripple (SWR) events occur during post-experience rest and sleep and play a significant role in memory consolidation. Recent ex vivo studies have found orexin-driven neuromodulatory effects on SWR parameters and proposed that they can influence memory consolidation. Additionally, behavioral studies demonstrate contradictory results regarding orexinergic effects on memory acquisition depending on task structure and intervention time. This inconsistency, along with the diffuse spread of orexin receptors (OXR) in the CA1 and CA3 hippocampus, challenges theoretical interpretation of orexinergic effects on memory. To reconcile this problem, we propose a rate-based model linking orexinergic tone-dependent SWR statistics to memory performance. The model combines an encoding module whose gain scales with the global orexinergic tone with a consolidation module driven by OX2R-specific replay dynamics. The model is based on a plasticity rule that considers SWR events as the unit of induction. Therefore, weight change requires sufficient replay content and depolarizing drive. We observed that, depending on the task structure, encoding-dominant learning benefits discrimination as the orexin drive increases. Consolidation-dominant learning shows a non-monotonic relationship with an internal optimum level of orexinergic tone. This interaction highlights that stronger truncation at higher receptor engagement shortens memory replay and reduces the efficacy of the consolidation process. The model predicts that the optimal regime depends on task complexity reflected in sequence length. Together, these results demonstrate how task demands determine whether increased orexinergic drive enhances or impairs effective consolidation and SWR-dependent replay in the hippocampus.

Online identification of nonlinear dynamical systems faces two challenges. Noisy, partial observations make the latent state difficult to estimate. Local attractors can also keep passive trajectories in regions that reveal little about the unknown dynamics. Both challenges commonly arise in closed-loop neural experiments where stimulation must be chosen from noisy spike counts, while neural activity may remain near a stable state unless stimulation drives it elsewhere. We address this problem by actively selecting inputs to collect informative observations during the experiment. We introduce PALDI (Planning-based Active Latent Dynamics Identification), which combines extended-Kalman state prediction, single-step Laplace state updates, blockwise Gaussian parameter updates, and finite-horizon planning. Its parameter expected-information-gain surrogate combines dynamics sensitivity with state Fisher information and discounts apparent parameter information when latent states are uncertain. We compare PALDI with nonadaptive inputs, a myopic policy, and two active baselines: the one-step D-optimal method FLEX and receding-horizon curiosity with uncertainty sampling. On a bistable Wilson-Cowan model of excitatory and inhibitory populations, PALDI approaches the true-model reference and remains strongest as the observation signal-to-noise ratio decreases or the readout geometry becomes biased. We then test identification in a spiking decision network where we use a reduced dynamics and observation model to approximate the data-generating process. PALDI reaches a median held-out prediction R2 of 0.90, close to a fitted reference, and its model gives the highest change-of-mind success among identification policies. Information-guided finite-horizon planning is a practical strategy for identifying neural circuit dynamics with closed-loop stimulation.

How does the brain generate sequences of discrete elements, such as actions or words, that follow an abstract rule (a grammar)? Motivated by an experiment in which monkeys learned to draw multi-object images according to grammatical rules, we trained a recurrent neural network to generate sequences Am Bn: producing shape A m times, then B n times, or vice versa. Our findings are threefold. First, the network plans the sequence by moving to an initial state encoding m and n along orthogonal directions for upcoming execution. Second, the network repeats each shape by progressing along a shape-specific manifold. Third, transitions from Am to Bn follow parallel trajectories between manifolds across values of n, thus preserving n. Each of these is an experimentally testable prediction. The network generalizes to longer novel sequences by starting further along the same orthogonal directions. These results reveal a neural circuit implementation of a symbolic algorithm for generating rule-based action sequences.

Alterations in impulsivity characterize several neuropsychiatric disorders, however the neural processes that underlie an impulsive choice have not been identified. A recent body of literature suggests that there may be generalizable aspects of neural representations across animals. Whether this is also true for dynamic rather than just static (like geometrical) properties of neural representations, and specifically in the context of impulsive behaviors, is less clear. Therefore, the goal of this study is to characterize generalizable features of latent dynamics that disambiguate an impulsive from non-impulsive choice across many different data sets and animals. A deep learning-based latent factor model was constructed from neural recordings that were acquired in rat anterior cingulate cortex while performing a delay discounting task, which is commonly used to assess impulsivity. Prior work shows this brain region is critical for the performance of this task. We leveraged piecewise linear recurrent neural networks (PLRNNs) to perform a nonlinear latent factor analysis across animals and experimental sessions. Our PLRNN explained a substantially larger percent of variance in fewer dimensions than linear approaches. We cross-validated performance of the model on test data and following the disruption of recurrent connectivity or auto-regressive latent dynamics. These tests confirmed that our fitting procedure consistently embedded generalizable neuronal dynamics within the PLRNN. Further, we found that the latent dynamics exhibited strong rotational features and progressed faster for impulsive choices. In summary, our model captures properties of the latent dynamics that are consistent across multiple data sets and disambiguate impulsive choices.

Nonlinear neural networks develop structured internal representations, yet how their geometry is determined by the tasks being learned remains poorly understood. Here, we develop an analytical framework for piecewise-linear feedforward and recurrent networks that links learning, activation-region structure, and neural selectivity. We show that, at any stationary point of gradient-aligned learning, a nonlinear network decomposes into local linear regressions over its activation regions. Deviations from the corresponding least-squares solutions are jointly constrained by network weights shared across regions and vanish in the low-error regime, yielding an approximate piecewise least-squares decomposition of the task. This structure admits a decision-tree interpretation, which we recover numerically by fitting trees to predict network activation patterns from the input. We further derive how the regions in which neurons are active determine the task statistics they capture, thereby organizing neural selectivity into distinct subpopulations. The resulting predictions closely match the selectivity geometry observed in simulated networks and two empirical neural datasets. Finally, we show that neural baseline regulates activation pattern diversity, placing networks along a continuum between coarse, generalizing representations and fine-grained, expressive representations. Together, these results establish activation regions as a unifying framework for describing how nonlinear networks decompose task structure into local computations and how those computations can be revealed through neural activity.

Decision-making requires accumulation of evidence over time, yet exactly how neural circuits implement this computation remains debated. Classical models posit that evidence accumulation is realized by ramping neural activity, whereas recent experimental findings suggest an alternative mechanism in which neural activity remains largely silent before abruptly transitioning into an active decision state. This mechanism leaves a fundamental question unanswered: how is evidence accumulated during the neural activity-silent period? To address this question, we trained excitatory-inhibitory recurrent neural networks (RNNs) with short-term synaptic plasticity (STP) on a reaction-time decision task. We found that after training, the networks learned to leverage STP to make decisions in an activity-silent manner. Choice-tuning analysis revealed a clear functional organization of the network, and connection-pruning analysis uncovered the coupling between choice-selective excitatory and inhibitory neural groups. Guided by the architecture of trained RNNs, we constructed a minimal neural circuit model, in which evidence accumulates in facilitated synaptic efficacy that progressively reshapes the circuit's phase portrait, eventually driving the network state across the stability boundary and triggering a rapid transition to a choice state. We further compared activity before a choice between the theoretical STP and Static models to examine the potential energy benefit of activity-silent decision-making. Together, our results suggest STP as a mechanism for evidence accumulation in activity-silent decision-making and offer insights into the design of energy-efficient artificial neural networks.

A central goal of neuroscience is to establish how similar (or different) brain networks are across individuals, development, psychiatric conditions, and even between biological and artificial brains. Whenever one such comparison is made, it relies on the assumption that we have a measure that produces an accurate and plausible similarity score. Yet, the field lacks one such measure, and the wide variety of existing measures often produces conflicting results. To address this problem, we perform a systematic comparison of 16 established similarity measures. First, we show that different measures consistently disagree on which brain networks are most similar. Second, we rank all measures according to multiple criteria such as biological plausibility, computational efficiency, and sensitivity. We show that DeltaCon is the most accurate measure in terms of parameter recovery, and that it ranks among the fastest and most noise-tolerant. For this reason, we conclude that DeltaCon is the best general-purpose similarity measure. However, no measure ranks best on more than two of the five criteria we assessed, and different measures should be preferred depending on the exact research question. Measure selection can now be made on evidence and should be a crucial decision in any study that compares brain networks.

From a statistical point of view, clustering spike time-windows from multi-electrode arrays (MEAs) recordings is a challenging high-dimensional, unsupervised clustering task where stationarity, commonly assumed in standard time-series analysis, is often violated and for which a gold standard is currently unavailable. Here we aim at providing practical guidance on how to cluster this type of data by systematically comparing 108 clustering pipelines differing along three main dimensions: (i) the data feature space; (ii) the metrics used to quantify distance between data points; and (iii) the clustering algorithm. For our benchmark we used both labeled synthetic data, mimicking four physiologically-inspired classes of spike time-windows, and a set of real MEAs recordings from a two-dimensional in-vitro neural culture. The performance of the competing pipelines was evaluated in terms of balanced accuracy and computational time when analyzing the synthetic datasets, while the Silhouette score was used for the real dataset, where no ground-truth is available. Overall, our analysis shows that the best combination is formed by k-means with Euclidean distance applied after Principal Component Analysis (PCA) of the spike time-windows. Conversely, hierarchical clustering showed the highest computational burden, while Independent Component Analysis and kernel PCA provided less effective noise suppression.

Recurrent cortical circuits are hypothesized to perform Bayesian posterior sampling through their stochastic dynamics. Yet identifying which sampling algorithm a circuit intrinsically implements, as opposed to mapping manually designing algorithms onto circuits, requires rigorous theoretical analysis of the biological circuit dynamics, and remains a fundamental open challenge because of the analytical intractability of nonlinear recurrent dynamics. Here, we tackle this challenge through comprehensive theoretical analysis of analytically tractable nonlinear recurrent circuit dynamics used in neuroscience research-- the continuous attractor network (CAN). Surprisingly, we find that the circuit, without any non-Gaussian components in its connectivity or internal variability, intrinsically emerges heavy-tailed subspaces dynamics described as scale-mixture Langevin posterior sampling. The mechanism emerges from the circuits own nonlinear dynamics: divisive normalization renders the overall population activation Gamma-distributed which multiplicatively modulates the time constant of the Langevin sampling in the stimulus feature subspace. Eventually, this scale-mixture structure produces a heavy-tailed (Student-t) sampling step size distribution, significantly accelerating the sampling speed. Moreover, the scale-mixture Langevin sampling approaches a Levy-like regime in the weak input limit. Our results analytically identify an intrinsic circuit algorithm embedded in subspaces of nonlinear recurrent dynamics, and establish a minimal circuit mechanism for endogenous heavy-tail statistics without any explicit non-Gaussian components.

The primate retina is strikingly non-uniform, with receptor density falling off sharply from fovea to periphery. Every eye-movement brings a previously low-acuity peripheral area to high-acuity foveal processing. We hypothesize that this process provides a natural self-supervision signal, enabling representation learning by predicting fine details in the periphery. We implement this in a SimMIM-style vision transformer, replacing its blank masks (ViT-Blank) with a uniform blur mask (ViT-Blur) and further with retinal masks in which a foveal patch is randomly selected, and all other patches are progressively blurred based on eccentricity from the foveal patch (ViT-Retina). ViTs were pre-trained to reconstruct the full-resolution image from the corrupted input and then fine-tuned on clean images for classification. ViT-Blur and ViT-Retina outperform ViT-Blank on the pre-training reconstruction task while controlling for retained information from the input across corruption types. Critically, they also outperform ViT-Blank on downstream classification, showing that they learn representations that generalize better. The ViT-Blur and ViT-Retina are also more robust in impoverishment conditions such as higher mask ratios, fewer pre-training epochs, smaller training sets, and test time image corruptions. Using additive band-limited noise, we discovered that ViT-Blank relies on higher spatial frequencies compared to ViT-Blur and ViT-Retina, and we show this to be the likely cause of poorer generalization: the centroid of a model's spatial frequency tuning is negatively correlated with its classification accuracy ($r = -0.73$). We further show that a closed-form linear-ridge encoder derived from the same reconstruction objective reproduces this tuning ($r = 0.88$). Overall, our experiments show the promise of a biologically-inspired self-supervision objective for learning robust visual representations.

Neuroscientific learning theories propose that the brain acquires knowledge by constructing internal world models. Supervised learning, the dominant approach in deep neural networks (DNN), relies on external category labels, making it difficult to reconcile with biological learning. There is an increasing trend towards more biologically valid approaches, such as predictive (minimize future surprise) or contrastive (minimize response to expected, maximize to unexpected inputs) objectives, but these approaches typically rely on pretrained DNN models that vary widely in architecture, size, and hyperparameters, making direct comparisons difficult. Here, we isolate the effect of learning by comparing small, identical networks trained with predictive, contrastive, and supervised learning objectives, as well as local (layer-restricted) vs. global (full backpropagation) learning. We show that brain representations after statistical learning are better modeled by a predictive local target than a supervised or contrastive target, and that during learning, the brain attenuates category-specific representations while retaining predictive ones. We additionally show that predictive objectives explain brain variance that standard supervised DNNs do not, and that this variance is tied to predictive rather than stimulus or mismatch processing. These findings demonstrate a controlled approach for testing the algorithmic basis of learning and identify prediction as a core learning mechanism.

Modern recording technologies enable simultaneous measurement of high-dimensional neural activity and rich behavioral variables, creating both an opportunity and a modeling challenge: identifying the shared latent structure that mediates the brain-behavior relationship and the underlying dynamics through which it evolves across behavioral states. Existing approaches focus on different parts of this challenge: shared representation learning methods identify neural-behavioral subspaces but typically do not model dynamical structure, while dynamical models capture temporal variation and discrete state transitions in neural activity, but do not explicitly disentangle shared neural-behavioral structure from modality-specific variability. We introduce Switching Shared Latent Dynamics (SSLD), a unified framework that learns a latent representation shared between neural activity and behavior and endows this shared space with nonlinear switching recurrent dynamics to reflect the hypothesis that behaviorally relevant neural representations exhibit structured dynamical changes across behavioral states. Complementary private latent variables capture modality-specific variability, providing a window into neural activity that does not directly interact with behavior. We evaluate SSLD on a simulated dataset and four diverse experimental datasets spanning species, brain regions, and recording modalities: monkey motor and premotor cortex during reaching, somatosensory cortex during a bump task, widefield calcium imaging across mouse dorsal cortex during (a) self-initiated decision-making and (b) spontaneous movements. Across all four datasets, SSLD accurately reconstructs neural and behavioral signals, recovers discrete states in the dynamics that align with experimentally defined behavioral epochs, and isolates behaviorally-relevant neural information in the shared latent while preserving private neural variability. Ablations confirm that shared representation learning, behavioral supervision, and switching dynamics each contribute to performance. SSLD offers an interpretable approach to modeling shared neural-behavioral dynamics that complements ongoing efforts toward foundation models for neuroscience.