Recurrent networks

…and formal methods, examining recurrent networks, attention mechanisms and state-space models. Work includes theoretical analysis, experimental algorithms and scientific publication. The advertised gross monthly salary is €2,300. The advert lists an October 2026 target start alongside a November application deadline; applicants should confirm start-date flexibility.

…exercises develop convolutional networks for classification, regression and image segmentation, with attention to preprocessing, augmentation, class imbalance, cross-validation and overfitting. Later sessions explore learning-rate and training-duration choices, recurrent networks and applications to omics data. Participants interpret model performance and discuss their own research problems. 2026-11-23–2026-11-27, 14:00–20:00 daily, Europe/Berlin (CET, UTC+01:00). O

Recurrent networks with balanced excitation-inhibition explain a wide range of neurophysiological observations, but can only implement a limited set of transformations on their input. On the other hand networks of firing-rate units with low-rank connectivity have universal computational capabilities, but do not work with spikes or generate noise self-consistently. Although empirical approaches to merge these two computational frameworks have been constructed, there is no established theory describing their unification. Here we develop such a theory. We study analytically and numerically networ

Behavioral Timescale Synaptic Plasticity (BTSP) is a type of synaptic plasticity in which large, Ca2+ plateau potentials in hippocampal pyramidal cells drive the formation of hippocampal fields such as place fields. BTSP differentiates itself from typical forms of plasticity by a) learning pre-post correlations on the timescales of seconds and b) learning in a one-to-few-shot manner. Though BTSP has been posited to be the primary driver of field formation in hippocampus, theoretical models that describe the population-level computations of hippocampus (such as the formation of "cognitive" or "

We consider two spiking neurons, with one neuron driving the other. From a statistical point of view, the relationship between the two neurons can be quantified by (i) the cross-correlation $C(t,t^{\prime})$ of the two spike trains, a measure of similarity of spontaneous fluctuations, and (ii) the linear response $K(t,t^{\prime})$ of the driven neuron's rate to a rate perturbation of the driving neuron. Connecting the two statistics, we analytically derive a simple cross-correlation–response relation (CRR), $C(t,t^{\prime})=K(t,t^{\prime})\lambda(t^{\prime})$, for an arbitrary model of the dr

Increasing the strength of synaptic connections (synaptic upscaling) within networks of randomly connected neurons transitions the internal dynamics from stationary to a chaotic regime [1, 2]. Synaptic upscaling in neural networks improves learning speed, accuracy, and robustness [3] and enhances the dimensionality of internal dynamics in response to external stimuli [2], by boosting chaotic activity. Nonetheless, the direct link between synaptic upscaling, the dimensionality of the network activity, and and enhanced computational capacity of neural networks for learning tasks remains unclear.

…In a recurrent networks, such variability implies quenched disorder in the synaptic weights, impeding the establishment of a continuous manifold of fixed points. If most positions in space are not any longer fixed points, is the popular notion of continuous attractors still relevant? What would be the neural dynamics underlying retrieval of spatial representations?We study the dynamics of recur

Biological neuronal networks exhibit hallmark features such as oscillatory dynamics, heterogeneity, modularity, and conduction delays. However, it has remained unclear to what extent these serve computational purposes. Inspired by physiological findings, we simulate recurrent networks of damped harmonic oscillators (DHO) in which one network node represents the aggregate activity of an underlying E-I circuit such as a cortical column (HORNs, see figure), at the same time reducing single-node oscillatory dynamics to its most basic form. We successively endow the networks with additional biologi

…functional roles, we simulated recurrent networks consisting of coupled damped harmonic oscillators (DHOs), so called HORNs. In HORNs, each DHO node represents the aggregate activity of a recurrently connected E-I circuit such as a cortical column. Mathematically, a DHO is given by a 2nd order ordinary differential equation with a single state variable x (amplitude) To model different intrinsic properties of an underlying E-I circuit's recur

Human brain function results from continuous activity flowing through specific cells, guided by specific synaptic arrangements. Therefore, determining the architecture and properties of functioning microcircuits is fundamental to understand the brain’s action. However, our current knowledge of human microcircuits is extremely limited, and relies heavily on assumptions from rodent research. Using tissue resected from 13 temporal lobe epilepsy patients, we perform multicellular patch-clamp based microcircuit analysis (n=80 recordings) to characterize synaptic connectivity in the human hippocampa

…and discusses mathematical models of recurrent networks.

…extract computational mechanisms from recurrent networks trained on neuroscience tasks, we show that the collective dynamics and sub-population structure play fundamentally complementary roles. Although various tasks can be implemented in networks with fully random population structure, we found that flexible i

Current voltage-dependent plasticity models assume that any plastic event needs a postsynaptic action potential and use induction mechanisms that are insensitive to postsynaptic voltage fluctuations. However, experiments have demonstrated that inhibitory inputs can affect plasticity updates, likely through voltage-dependent effects. To address such findings, here we introduce the novel Voltage-Dependent Pathway model (VDP) where the induction variables track the membrane potential, increasing its sensitivity to fast fluctuations. We first confirm that this model produces spike-timing dependent

Committing to a decision tends to bias future behaviors, a phenomenon dubbed “confirmation bias”. Daily life requires us to make decisions on remembered past experiences; still, the understanding of confirmation bias in working memory remains sparse. Here, we show that the stimulus-specific drift dynamics should be incorporated into any viable account of confirmation bias in working memory. We crafted a paradigm where participants view a grating briefly, make a binary decision on it, and estimate its orientation, with varying delays between those epochs. We developed a dynamical model wherein

…natural questions: (i) How do recurrent circuits generate codes that are compatible with the convexity of receptive fields? (ii) How can we utilize the constraints imposed by the convex receptive

…the optimal operating point of the recurrent networks. Moreover, the data from the visual cortex of the monkeys during the attentional task indicate that they flexibly change the closeness to the criti

…Sequences arise in recurrent networks with asymmetric connectivity structures [1]. These models are of particular interest because they can be learned with biologically plausible temporal asymmetric Hebbian learning (TAH) rules, providing a foundation for modeling of sequential activity in neural circuits. In existing models, the time in

…that spontaneous synchronization in recurrent networks is the key mechanism for selective processin

Memories are known to reactivate during sleep. A recent modelling study (Fauth & van Rossum, eLife, 2019) could reproduce this based on self-reactivations of heavily interconnected cell assemblies and showcased its beneficial consequences for memories. However, to be maintained, the memories needed frequent reactivations such that the weights within the assemblies remain at a high level. In this work, we extend the model such that memories are maintained independently of reactivations. We suggest that long-term memories are mainly represented by the number of connections between assembly

Understanding the trade-off between plasticity and stability–how neural networks form new memories while avoiding runaway potentiation or fast overwriting of memories–is a fundamental problem in neuroscience. Previous theoretical studies that relied on non-plausible plasticity rules did not reproduce realistic synaptic weight distributions or neural activity statistics and predicted fast forgetting times. We revisit this problem by investigating the implications of a multiplicative calcium-based plasticity rule in a recurrent network. This rule was recently fitted to experiments done in physio

Understanding how neural circuits in the brain wire up during development is important for the implementation of numerous functions in adulthood, but also for the prevention and treatment of many neurological disorders that result from bad wiring. Before the onset of sensory experience, spontaneous activity in the developing brain organizes and refines circuits. Hence, alterations in spontaneous activity can lead to severe deficits in wiring. One such alteration occurs in the Fragile X mouse model, where a single genetic mutation reduces cortical inhibition and increases the number of neurons

We introduce Norse: An open-source library for gradient-based training of spiking neural networks. In contrast to neuron simulators which mainly target computational neuroscientists, our library seamlessly integrates with the existing PyTorch ecosystem using abstractions familiar to the machine learning community. This has immediate benefits in that it provides a familiar interface, hardware accelerator support and, most importantly, the ability to use gradient-based optimization. While many parallel efforts in this direction exist, Norse emphasizes flexibility and usability in three ways. Use

The Stabilized Supralinear Network is a model of recurrently connected excitatory (E) and inhibitory (I) neurons with a supralinear input-output relation. It can explain cortical computations such as response normalization and inhibitory stabilization. However, the network's connectivity is designed by hand, based on experimental measurements. How the recurrent synaptic weights can be learned from the sensory input statistics in a biologically plausible way is unknown. Earlier theoretical work on plasticity focused on single neurons and the balance of excitation and inhibition but did not cons

…For example, many networks can give rise to similar functional responses, and the same network can function differently depending on context. Whether certain patterns of synaptic connectivity are required to generate specific network-level computations is largely unknown. Here we introduce a geometric framework for identifying synaptic connections required by steady-state r