Neural Networks

Convolutional neural networks (CNNs) have become essential models for predicting neural activity and behavior in visual tasks. However, their ability to capture higher-level cognitive functions, such as numerosity discrimination, remains debated. Numerosity, the ability to perceive and estimate the number of items in a visual scene, is often proposed to rely on specialized number-detector units within CNNs, analogous to number-selective neurons observed in the brain. In this study, we use CORnet, a biologically inspired CNN architecture inspired by the organization of the primate visual system. To address a limitation of classical Representational Similarity Analysis (RSA)--its assumption that all units contribute equally--we apply pruning, a feature selection approach that identifies the units most relevant for explaining behavioral similarity structure. Our results show that number-detector units are not critical for population-level representations of numerosity, challenging their proposed role in previous studies.

Neurons in the cerebral cortex are organized topographically. In the primate visual cortex, neighboring neurons often respond to similar stimulus parameters, such as receptive field position, orientation, color, and spatial frequency. Preferred stimulus parameters change smoothly across the cortical surface. If such topographic organization plays an important role in computation, it is likely to emerge in artificial neural networks. In this study, a multistream convolutional neural network was constructed in which filters in the first convolutional layer were arranged in a two-dimensional filter matrix according to their output connections. The network was trained using supervised learning for image classification. Although adjacent filters in the filter matrix can develop any structure in principle, they acquire similar degrees of orientation and color selectivity. Moreover, they prefer similar orientations, hues, and spatial frequency. The similarity decreases with distance between filters in the matrix. Furthermore, neural-network model instances that have a strong relationship between filter distance and filter-property similarity performed better than those with a weak relationship. These results suggest that topographic organization emerges spontaneously in an artificial neural network and plays an important role in model performance, suggesting the importance of topographic organization for computations performed by artificial and biological neural networks.

AO_SCPLOWBSTRACTC_SCPLOWOne of the most compelling ideas for bridging neuroscience and artificial neural networks is the establishment of a framework based on three main components: network architecture, optimization mechanism, and loss (or objective) function to be minimized. While the first two components have been extensively explored, the definition of a loss or objective function in neuroscience has been addressed less thoroughly, often from perspectives such as predictive coding. In this work, we propose an elementary loss function grounded in the comparison of neuronal responses to two signals: an external one, used for learning, and an internal one, reflecting the acquired knowledge. The loss function is thus simply the basic difference between the two, which, in terms of logical signals, corresponds to a well-known non-linearly separable function: the XOR function. We illustrate with a computational example how a binarized image recognition algorithm can be straightforwardly implemented in an autoencoder, and we show how a neuronal motif organized around an inhibitory neuron could implement such XOR operation and provide a feedback signal that makes optimization possible.

Hippocampus is known to replay activity patterns to recall and process memories, which is often related to Hopfield-type attractor dynamics. Another line of theoretical studies suggests that hippocampal replay prioritizes replay of experiences to accelerate value learning for efficient decision making. It is unknown how hippocampal attractor dynamics perform prioritized memory sampling, and more broadly, how we can consistently relate dynamical (bottom-up) and functional (top-down) theories of hippocampal replay. In this paper, we propose an extended Hopfield-type attractor network model with momentum, kinetic energy, and conservation of the total energy, which is called momentum Hopfield model. We show that our model can be interpreted as CA3-CA1 network model with intrinsic oscillation, and such network model reproduces hippocampal replay in 1-D and 2-D spatial structures. We also prove that our model functionally works as Markov-chain Monte Carlo sampling in which recall frequencies of memory patterns can be arbitrarily biased. Using this property, we implemented prioritized experience replay using our model, which actually accelerated reinforcement learning for spatial navigation. Our model explains how dynamics of hippocampal circuits realize efficient memory sampling, providing a theoretical link between dynamics and functions of hippocampal replay.

Generalization is a fundamental criterion for evaluating learning effectiveness, a domain where biological intelligence excels yet artificial intelligence continues to face challenges. In biological learning and memory, the well-documented spacing effect shows that appropriately spaced intervals between learning trials can significantly improve behavioral performance. While multiple theories have been proposed to explain its underlying mechanisms, one compelling hypothesis is that spaced training promotes integration of input and innate variations, thereby enhancing generalization to novel but related scenarios. Here we examine this hypothesis by introducing a bio-inspired spacing effect into artificial neural networks, integrating input and innate variations across spaced intervals at the neuronal, synaptic, and network levels. These spaced ensemble strategies yield significant performance gains across various benchmark datasets and network architectures. Biological experiments on Drosophila further validate the complementary effect of appropriate variations and spaced intervals in improving generalization, which together reveal a convergent computational principle shared by biological learning and machine learning.

The hippocampus (HPC), prefrontal cortex (PFC), and thalamic nuclei, such as reuniens (Re), form a reciprocally connected circuit that plays a critical role in processing hippocampus-dependent memory. Accumulating evidence suggests that this triangular modular circuit is crucial for performing cognitive tasks that require context-dependent memory, which belong to a class of behavioral tasks difficult for animals to learn. Experiments are gradually revealing what behavioral information these brain regions represent, but how the triangular circuit gives rise to the observed divisions of labor remains unknown. It is also unclear whether the triangular modular circuit brings any advantage in solving such tasks. Here, we addressed these questions by constructing a prefrontal-thalamo-hippocampal circuit model comprising interconnected long-short-term memory (LSTM) units and training it on contextual memory-dependent spatial navigation tasks. Our model revealed the critical roles of the distinct brain modules. The HPC module encoded spatial information, whereas the PFC module represented the spatiotemporal task structure in a context-dependent manner. The Re module integrated task-relevant information to facilitate learning in the PFC and HPC modules, dynamically harmonizing these modules. The thalamic coordination of the other modules enhanced the systems robustness in learning to navigate complex environments. This division of labor between the HPC, PFC, and Re modules was not specified a priori but emerged through learning, showing an interesting coincidence with the task-related activities of the prefrontal-thalamo-hippocampal circuit. Our results demonstrate that the multi-modular network structure is crucial for robust processing of context-dependent memory.

Recurrent neural networks (RNNs) have proven to be highly successful in emulating human-like cognitive functions such as working memory. In recent years, RNNs are evolving to incorporate more biophysical realism to produce more plausible predictions on how cognitive tasks are solved in real neural circuits. However, there are major challenges in constructing and training networks with the complex and nonlinear properties of real neurons. A major component of the success of RNNs is that they share the same mathematical base as deep neural networks, permitting highly efficient optimization of model parameters using standard deep learning techniques. To do so, they use abstract representations of neurons which fail to capture the impact of cell-level biophysical and morphologic properties that may benefit network-level function. Expanding task-trained RNNs with biophysical properties such as dendrites and active ionic currents poses substantial challenges, as it moves these models away from the validated training regimes known to be highly effective for RNNs. To address this gap, we developed a biophysically detailed reservoir computing (BRC) framework with the goal of extracting mechanistic insights from biophysical neural models, and propose that these insights can be used to guide model choices that will work for specific categories of cognitive tasks. The BRC network was constructed with synaptically coupled excitatory and inhibitory cells, in which the excitatory cells include multicompartment biophysically active dendrites; motivated by empirical studies suggesting dendrites have desirable computational benefits (e.g. pattern classification and coincidence detection). We trained the BRC network to do a simplified working memory task where it had to maintain the representation of an extrinsic "cue" input. We studied the impact of extrinsic input time constants (fast AMPA vs slow NMDA) and location (dendrite vs soma) on the ability of a network to solve the task. Our results revealed that cue inputs through NMDA receptors are particularly efficient for solving the working memory task. Further, the properties of NMDA receptors are uniquely suited for cue inputs delivered at the dendrite, as networks trained with dendritic AMPA cue inputs failed to solve the task. Detailed examination of the cell and network dynamics that solve the task reveals distinct local network configurations and computing principles for the different types of extrinsic input. Overall, much like the body of mechanistic insights that have underpinned the success of training RNNs, this study lays the groundwork for applying the BRC framework to train biophysically detailed neural models to solve complex human-like cognitive tasks.

Head direction (HD) cells can fire selectively as a function of the animals head azimuth direction and form an internal compass for navigation. They were found in the mammalian limbic system including the dorsal presubiculum and entorhinal cortex. The underlying network updates its directional estimate in a self-organized fashion and can be recalibrated by external sensory cues. Although ring-attractor models were proposed to account for azimuth coding on horizontal planes, they cannot explain conjunctive azimuth-and-pitch tuning observed in HD cells of the presubiculum of flying bats navigating in three-dimensional (3-D) space. Based on the 3-D electrophysiological recordings, we developed a toroidal continuous attractor network that jointly encodes horizontal azimuth and vertical pitch angle of head direction. The model can reproduce the experimentally recorded tuning curves of individual HD cells, and accurately encode the 3-D dynamic head direction of bat by HD cell population. The model also simulate the influence of horizontal visual cue manipulation on the HD system in 3-D space and predicts how horizontal landmark rotation induces a sustained azimuthal offset that persists after the cue is removed, which is comparable to two-dimensional experimental findings. This research clarifies how conjunctive 3-D direction codes are generated and modified by vestibular input, visual information and recurrent connectivity. It also uncovers the computational principles and properties of the brains navigation functions in realistic 3-D environments and offers new theoretical reference for future studies. Author summaryA central function of the brains navigation system is to track head direction, which in terrestrial mammals is largely confined to a horizontal plane. However, for animals like bats that navigate freely in three-dimensional (3-D) space, the neural code for head direction has been found to be high-dimensional. While ring attractor network models elegantly explain horizontal azimuth coding, how the 3-D head direction code is formed and affected by sensory information remains unknown. Here, we develop a toroidal continuous attractor network model with vestibular and visual modules that explains the jointly encoding of azimuth and pitch angles, and the effect of visual cue. We introduce the toroidal topology to the continuous attractor network, and the high-dimensional angular velocity signal from vestibular system and visual input are used to update the population dynamics of the network. Our model is constrained by and reproduces 3-D electrophysiological data from flying bats, captures the unique tuning curves of individual head direction cells, and further achieves the accurate encoding of dynamic 3-D head movements by population activity. Evidences in the rodent have suggested visual cue manipulation recalibrates head direction coding horizontally. Our model further successfully simulates the influence of visual cue rotation in 3-D space, predicting a persistent, global offset in the azimuth encoding--an effect that endures after cue removal and aligns with 2-D rodent experimental phenomena. This work provides a mechanistic, network-level explanation for 3-D head direction encoding and reveals how multimodal cues are integrated to form the internal compass in a volumetric world.

Animals and humans use multiple behavioral strategies to perform tasks. However, neural implementations of multiple strategies remain elusive, as some studies propose distinct pathways, while others observe overlapping brain regions associated with strategies. We propose a hybrid deep reinforcement learning (H-DRL) method, in which one network model implements model-free and inference-based behaviors through synaptic plasticity and recurrent activity. H-DRL uses a single updating rule and switches the strategy according to task demands without an explicit arbitrator. H-DRL reproduced mixed strategies of humans in a two-step task. In the mouse perceptual decision-making task, H-DRL adapted the recurrent dynamics with rich learning when the task condition required inference-based behavior, while adopting model-free behavior with lazy learning for a simple condition. The activity of H-DRL units showed condition-dependent maintenance of previous events, consistent with orbitofrontal cortical activity in mice. Our model provides a unified view that one cortical network automatically determines strategies in use depending on task conditions.

Understanding neural dynamics is crucial for uncovering how the brain processes information and controls behavior. Linear dynamical systems (LDS) are widely used for modeling neural data due to their simplicity and effectiveness in capturing latent dynamics. However, LDS assumes a stable mapping from the latent states to neural activity, limiting its ability to capture representational drift--gradual changes in the brains representation of the external world. To address this, we introduce the Stiefel Manifold Dynamical System (SMDS), a new class of model designed to account for drift in neural representations across trials. In SMDS, emission matrices are constrained to be orthonormal and evolve smoothly over trials on the Stiefel manifold--the space of all orthonormal matrices--while the dynamics parameters are shared. This formulation allows SMDS to leverage data across trials while accounting for non-stationarity, thus capturing the underlying neural dynamics more accurately compared to an LDS. We apply SMDS to both simulated datasets and neural recordings across species. Our results consistently show that SMDS outperforms LDS in terms of log-likelihood and requires fewer latent dimensions to capture the same activity. Moreover, SMDS provides a powerful framework for quantifying and interpreting representational drift. It reveals a gradual drift over the course of minutes in the neural recordings and uncovers varying drift rates across dimensions, with slower drift in behaviorally and neurally significant dimensions.

Convolutional neural networks (CNNs) are a class of artificial neural networks (ANNs). Since the early 2010s, they have been widely adopted as models of primate vision and classifiers of neuroimaging data, becoming relevant for a wealth of neuroscientific fields. However, the majority of neuroscience researchers come from soft-science backgrounds (like medicine, biology, or psychology) and do not have enough quantitative skills to understand the inner workings of A/CNNs. To avoid undesirable black boxes, neuroscientists should acquire some rudiments of computational neuroscience and machine learning. However, most researchers do not have the time nor the resources to make big learning investments, and self-study materials are hardly tailored to people with little mathematical background. This paper aims to fill this gap by providing a concise but accurate introduction to CNNs and their use in neuroscience -- using the minimum required mathematics, neuroscientific analogies, and Python code examples. A companion Jupyter Notebook guides readers through code examples, translating theory into practice and providing visual outputs. The paper is organised in three sections: The Concepts, The Implementation, and The Biological Plausibility of A/CNNs. The three sections are largely independent, so readers can either go through the entire paper or select a section of interest.

As social beings, humans make decisions partly based on social interaction. Observing the behavior of others can lead to learning from and about them, potentially increasing trust and prompting trust-based behavioral changes. Observation-based decision making involves different neural structures. The orbitofrontal cortex (OFC) and lateral prefrontal cortex (LPFC) are known as neural structures mainly involved in processing emotional and cognitive decision values, respectively, while the anterior cingulate cortex (ACC) plays a pivotal role as a social hub, integrating the afferent expectancy signals from OFC and LPFC. This paper presents a neurocomputational model of the interplay between observational learning and trust, as well as their role in individual decision-making. Our model elucidates and predicts the emotional and rational behavioral changes of an individual influenced by observing the action-outcome association of an alleged expert. We have modeled the neurodynamics of three cortical structures (OFC, LPFC, and ACC) and their interactions, where the neural oscillatory properties, modeled with Dynamic Bayesian Probability, represent the observers attitude towards the expert and the decision options. As an example of an everyday behavioral situation related to climate change, we use the choice of transportation between home and work. The EEG-like simulation outputs from our model represent the presumed brain activity of an individual making such a choice, assuming the decision-maker is exposed to social information.

Human decisions under uncertainty are shaped by experience, but the computations that translate expectation and experience into choice remain debated in neural and cognitive science. Prior studies highlight reinforcement learning (RL) as a unifying framework, yet it is unclear whether human behavior under risk is better captured by symmetric updating from outcomes or by asymmetric learning that weights reward and loss differently. This work examines which learning strategies better explain trial-by-trial choices given contextual uncertainty and manipulations of outcome distributions. Our results show that a Risk Sensitive (RS) model with asymmetric learning rates best explains human behavior in our novel decision-making task. Fitting candidate models to individual trial histories yielded value signals that predicted both choice and response time. These results highlight that RS model, as an asymmetric learning provides a concise and identifiable account of behavior in decision-making under risk tasks.

Inhibitory synapses can control a neurons firing rate and also control supralinear dendritic integration. It is not known how inhibitory synapses can learn to perform these functions using only signals available locally at the synaptic site. We study an inhibitory plasticity rule based on the Bienenstock-Cooper-Munro theory in multicompartment models of striatal projection neurons, and show that it can perform these two functions. The rule uses local voltage-gated calcium concentration in the dendrites to regulate inhibitory synaptic strength. We show that, for rate-coded inputs, the rule can achieve precise control of neuronal firing rate after changes in excitatory input rate or excitatory synaptic strength. Additionally, for sparsely-coded inputs that activate localized synaptic clusters in dendrites, the rule can either allow or inhibit the supralinear dendritic response evoked by the clustered excitatory synapses, or equalize the dendritic response arising from different clusters. Finally, we demonstrate the use of learning to inhibit supralinear dendritic integration for solving the nonlinear feature binding problem (NFBP), in tandem with a simple excitatory plasticity rule. We conclude by discussing why the collateral inhibitory synapses between striatal projection neurons could contribute to solving the NFBP with this plasticity rule. Author summaryNeurons are the main cells in the nervous system that process information. They receive signals from the bodys senses--both external and internal--and use them to guide actions such as muscle movement and the regulation of bodily functions. A neuron becomes active when incoming signals excite it strongly enough. But for neurons to work timely, precisely, and reliably, their activity needs to be shaped, modified and controlled. This is done by inhibition, which comes from specialized inhibitory neurons. In this article we study how inhibition can learn to do two of its most basic roles in the nervous system. The first is to help neurons stay responsive across a wide range of input strengths--from very weak to very strong stimulation. For example, neurons in the retina allow vision both in dim starlight and in bright sunlight, even though these conditions differ in brightness by a trillion-fold. Inhibition contributes to handling this huge range by preventing overstimulation of the neurons in bright light. The second role of inhibition is to control strong, local excitations that occur on specific dendritic branches of a neuron. These local excitations can suddenly push a neuron into activity, and inhibition controls whether such excitations are allowed or suppressed. We use a learning mechanism that is already known to exist for excitatory synapses, but here we apply it to inhibition to explore what it could achieve. The results show that if inhibitory synapses used this same learning rule, they could support the two fundamental roles of inhibition in the nervous system described above.

Animals exposed to pairings of a neutral stimulus with reward acquire a conditioned response to the neutral stimulus. A prominent hypothesis, formalized in the Temporal Difference (TD) learning algorithm, is that animals learn to predict the future reward associated with the neutral stimulus ("value"). Though the TD algorithm does not explicitly specify what drives conditioned responding, a typical assumption is that it reflects the animals estimate of value. In TD learning, value estimates are updated using reward prediction error (RPE, the discrepancy between observed and predicted reward), and are thought to be signaled by the phasic activity of midbrain dopamine neurons. This hypothesis posits that dopamines effects on conditioned responding are mediated entirely by its effects on learning. However, recent experimental and theoretical evidence suggests that dopamine may play a more direct role in modulating conditioned responding. We use a combination of data analysis and computational modeling to probe the relationship between dopamine and conditioned responding. Our results suggest that dopamine directly modulates conditioned responding, in addition to its role in learning. These findings can be captured by a model in which dopamine RPE acts both indirectly (via learning) and directly on conditioned responding.

The primate retina dissects visual scenes into multiple retinocortical streams. The most numerous retinal ganglion cell (GC) types, midget and parasol cells, are further divided into ON and OFF subtypes. These four GC populations have anatomical and physiological asymmetries, which are reflected in the spike trains received by downstream circuits. Computational models of the visual cortex, however, rarely take GC signal processing into account. We have built a macaque retina simulator with the aim of providing biologically plausible spike trains for downstream visual cortex simulations. The simulator is based on realistic sampling density and receptive field size as a function of eccentricity, as well as on two distinct spatial and three temporal receptive field models. Starting from data from literature and earlier receptive field measurements, we synthetize distributions for receptive field parameters, from which the synthetic units are sampled. The models are restricted for monocular and monochromatic stimuli and follow data from the temporal hemiretina which is more isotropic. We show that the model patches conform to anatomical data not used in the reconstruction process and characterize the responses with respect to spatial and temporal contrast sensitivity functions. This simulator allows starting from a stimulus video and provides biologically plausible spike trains for the distinct unit types. This supports development of thalamocortical primate model systems of vision. In addition, it can provide a reference for more biophysical retina models. The independent parameters are housed in text files supporting reparameterization for particular macaque data or other primate species. Author summaryVisual environment provides a rich source of information, and the visual system structure and function has been studied for decades in many species, including humans. The most complex data in mammalian species are processed in the cerebral cortex, but to date we are still missing a functioning model of cortical computations. While the earlier anatomical and physiological data describe many details of the visual system, to understand the functional logic we need to numerically simulate the complex interactions within this system. To pave the way for simulating visual cortex computations, we have developed a functioning model for macaque retina. The neuroinformatics comprises a review and re-digitized existing retina data from literature, as well as statistics of earlier macaque receptive field data. Finally, we provide software which brings the collected neuroinformatics to life and allows researchers to convert visual input into biologically feasible spike trains for simulation experiments of visual cortex.

Human behavior differs substantially across individuals. While artificial neural networks (ANNs) are regarded as promising models of human perception, they are often assumed to lack such individual differences. Here, we demonstrate that multiple instances of the same ANN architecture exhibit substantial individual differences in behavior that mimic those observed in humans. We trained and tested 60 ANN instances from three architectures on a digit recognition task and found notable individual differences in overall accuracy, confidence, and response time (RT). Critically, these individual differences in ANN instances mapped consistently onto the individual differences produced by 60 humans performing the same task, with the mapping strength often approaching the human-to-human benchmark across all three behavioral metrics (accuracy, confidence, RT). The mapping generalized even across behavioral metrics: an ANN instance that aligned with an individual human on accuracy also aligned with the same individual on confidence and RT. These findings generalized to a more complex, 10-choice blurry object recognition task, though the human-ANN mapping was generally less robust than the human-human benchmark. Overall, these findings open the possibility of using ANN ensembles as computational proxies for probing the mechanisms underlying human variability.

Memory consolidation is the process by which temporary, malleable memories are transformed into more stable, longer-lasting forms. On a coarse anatomical scale, consolidation redistributes memories in the brain, but it remains poorly understood how these changes manifest themselves on the finer, cellular scale of neuronal engrams and how they relate to the cognitive level. In this study, we developed a phenomenological model of engram dynamics under systems consolidation. The model describes consolidation as a brain-wide phenomenon, where memories deterministically follow a trajectory through a space of patterns distributed among brain regions. It captures a broad range of features of memory consolidation, including selective consolidation, semantization, and power-law forgetting. In the model, consolidation is accompanied by population-level changes in neuronal representations that resemble the widely observed phenomenon of representational drift. When only a subset of neurons is observed, the deterministic dynamics of the model can appear stochastic, and a readout of task features deteriorates over time even when a stable readout exists for the full system. Our model offers a dynamical systems perspective on memory consolidation as a distributed process, moving beyond the classic region-centered view, and provides a functional interpretation of drift as a means of redistributing engrams for improved memory retention.

Previously, electrophysiological differences between subpopulations of midbrain dopamine (DA) neurons were identified based on projection targets, including distinct responses to hyperpolarization and in the regularity of pacemaking. Here we explored single-compartment models of three subpopulations of DA neurons, projecting to medial shell of the nucleus accumbens (VTA-mNAcc), dorsomedial striatum (SNc-DMS) or dorsolateral striatum (SNc-DLS). We reduced the dimensionality to a phase plane consisting of membrane potential and one slow variable, either total slow potassium conductance or Kv4 channel inactivation. Nullclines are curves on which the rate of change of each variable is zero, given the value of the other variable. The voltage nullclines had three branches: upper spiking, unstable middle, and lower quiescent branch. Recruitment of Kv4 channels by the more prominent after-hyperpolarizing potential (AHP) in the DA-DMS and DA-DLS models channels stabilized pacemaking by creating a restorative moving fixed point along the quiescent branch. The slow inactivation of KV4 channels dominated and regularized the dynamics during the interspike interval; a dominant slow process may be a general mechanisn for stable regular pacemaking in a frequency range between 1-10 Hz. In contrast, the smaller AHP in VTA-mNAcc models prevented recruitment of this Kv4-mediated moving fixed point, which increased the sensitivity to synaptic inputs. On rebound from hyperpolarization the ability to produce robust ramps reverses between the DA neurons: now VTA-mNAcc projecting DA models fully recruited Kv4 channels and produced stable ramp-like pauses, whereas SNc-DLS projecting cells recruited significant regenerative inward CaV3 channels that overwhelmed Kv4 channels and produced rebound bursts. Author SummaryMidbraim dopamine (DA) neurons in the mammalian midbrain are linked to motivation, control of voluntary movement initiation, and reward-based learning. Their dysfunction is implicated in major disorders like Parkinsons disease, schizophrenia or substance use disorders. Firing patters like bursts or pauses in most DA subpopulations are thought to signal better or worse than expected outcomes. Here we use dynamic systems analysis to capture how functional diversity of DA neurons of their intrinsic properties results in differences of synaptic input integration leading to the generation of burst and pause patterns of electrical activity.

Latent Factor Analysis via Dynamical Systems (LFADS) is a powerful variational autoencoder for inferring neural population dynamics from spike train data. However, LFADS suffers from pos-terior collapse, where the learned posterior collapses to the prior, eliminating meaningful latent representations. Current solutions require computationally expensive Population-Based Training (PBT) to dynamically tune regularization hyperparameters. Here, we demonstrate that Large Lan-guage Model (LLM)-based program evolution can discover regularization schedules that prevent posterior collapse without PBT. Using FunSearch, an evolutionary algorithm that uses LLMs to generate and refine Python functions, we evolved adaptive regularization schedules that respond to training dynamics. Our best evolved schedule prevents posterior collapse across all tested conditions, maintaining KL divergence 6.5x higher than baseline schedules at 50 epochs (n = 10 seeds each, p < 0.001) and stable above 0.09 through 500 epochs across three Neural Latents Benchmark datasets, while preserving reconstruction quality. This work represents the first application of LLM-based program synthesis to variational autoencoder hyperparameter scheduling, offering a computationally efficient alternative to population-based optimization.