Barcelona Computational, Cognitive and Systems Neuroscience (BARCCSYN) 2026

We recently proposed that the beta rhythm (15–30 Hz) provides a key aspect of routing information through the brain, namely, the formation of flexible, transient neural ensembles. We tested this hypothesis using spike and LFP recordings in non-human primates and MEG/EEG in healthy human participants performing a categorical decision-making task. We found that beta-band frequency shifts in frontal cortex signal categorical decision outcomes, with activity in these bands predicting the behavioral response. These results further substantiate the idea that beta provides the scaPolding for the formation of neural ensembles. We argue that beta frequency shifts arise from changes in connectivity between weakly coupled oscillators and that, more than a spectral fingerprint, they reflect an active mechanism to (re)-activate behaviorally relevant communication channels in the brain.

To understand adaptive behavior, we must understand how decisions are computed and updated through experience. For example, when facing a potential threat, animals may initially rely on instinctive responses, such as escape, yet with experience they can suppress these reactions as the environment is reinterpreted. In this talk, I will show how these transformations arise from interactions between cortical and subcortical circuits that dynamically reshape decision variables in aversive contexts.
Using an ethologically grounded paradigm in which mice escape from an overhead looming stimulus that mimics an aerial predator, I will show that repeated exposure leads to suppression of this instinctive response. We find that higher visual areas instruct this learning through projections to the inhibitory ventrolateral geniculate nucleus (vLGN). Neural population recordings reveal that specific vLGN neurons increase their activity across learning and are necessary for the expression of escape suppression, indicating that cortical inputs instruct learning, while plasticity within vLGN implements the learned behavioral change.
I will then extend this framework to show how vLGN integrates inputs from retrosplenial cortex and hypothalamic circuits to regulate decisions between safety and exploration. These pathways carry distinct signals related to prior experience and internal state, and their coordinated activity shapes when animals decide to leave safety. ogether, these results reveal circuit mechanisms through which experience reshapes decision-making in aversive environments.
Finally, I will present ongoing work showing how acute stress induces persistent changes in hippocampal dynamics, linking experience-dependent updating of threat with longer-lasting circuit adaptations that may be relevant for anxiety and PTSD.

The ability to train spiking neural networks is essential for the modeling of biological neural networks as well as for neuromorphic computing. The standard approach when training non-spiking neural networks is backpropagation. In spiking neural networks, however, gradient descent learning is complicated by two fundamental problems: First, the vanishing and addition of spikes leads to disruptive, discontinuous changes in the network dynamics, which are not taken into account by the gradient. Second, exact gradients cannot systematically generate or remove spikes, even if required. These problems have so far been ignored or circumvented by using heuristics.

Here we show that and how they can be solved. Specifically, we demonstrate exact gradient descent learning based on spiking dynamics that change continuously or even smoothly. These are generated by neuron models whose spikes vanish and appear only at the end of a trial, where this does not influence any future dynamics. Among others, neuron models that generate spikes via a self-amplification mechanism and reach infinite voltage in finite time have this property; this includes the standard quadratic leaky integrate-and-fire (QIF) neuron. Besides spike removal, such neuron models also enable gradient-based spike addition by means of what we call pseudospikes. The timings of pseudospikes are continuous and mostly smooth extensions of the times of ordinary spikes disappearing at the trial end. We apply our scheme to individual and networks of QIF neurons using event-based simulations and automatic differentiation to compute spike-based gradients. This allows, in particular, to induce and continuously move spikes to desired times. Further, we match the performance on MNIST of previous studies using time-to-first-spike coding, but starting from an initially silent deep network.

Finally we use a standard benchmark dataset for spiking networks, the Spiking Heidelberg Digits speech classification dataset, to compare training with exact gradient descent and time-to-first-spike coding in QIF and LIF networks. We find that the QIF networks outperform the LIF networks and that the best performing LIF networks have near-simultaneous output, which may be undesirable in practice. Consistent with the performance difference, the loss landscapes of QIF networks appear less rugged than the discontinuous ones of LIF networks and the related loss gradient landscapes contain less noise. By analyzing the loss gradients of single samples, we show that the noise stems from rapid and seemingly random changes of the gradient direction, which occur more often in LIF networks due to the discontinuous changes of spike times.

The research has been conducted together with Christian Klos and Carlo Wenig.