Embedded Files
Workshop description
Workshop description
Even putatively simple nervous systems exhibit a high degree of complexity that is not fully understood. The recent discovery of associative learning in box jellyfish, for example, highlights the surprisingly rich behavioural repertoire of such networks and invites a more in-depth study of the underlying neural mechanisms. These insights serve as a promising inspiration to develop technical systems that fully exploit the underlying principles of the unmatched efficiency and resilience of natural systems.
The aim of this one-day workshop is to bring together biologists working on biological model systems with theoreticians developing approaches to understand these systems and engineers working to implement novel, non-conventional neuromorphic computing schemes and applications. Relatively simple - but resilient and efficient - nervous systems might be an excellent inspiration for edge AI, synthetic cells, sensing or miniature robotics. Therefore this workshop aims to distill principles from the evolution of the nervous system to inform theoretical insight and ultimately to design novel neuro-inspired computing hardware.
The workshop is divided into three sessions:
In the first session, “Model Systems for self-organization, evolution and computation”, we will hear from experts studying a diverse range of model organisms, including cnidarians, acoels and Drosophila. These talks will help to delineate biological design principles for functional nervous systems.
The second session, “Theoretical Approaches to understand network computations”, will focus on theoretical studies of network dynamics, the evolution of memory, computational connectomics as well as neural circuits for signal separation and novelty enhancement.
The final session, “Evolving Bio-Inspired Systems”, will highlight how insights into foundational principles translate to emerging applications in neuroevolution, physical circuits and computing architectures.
The workshop will include several interim panel discussions to take stock of the state of the field and find common principles between model organisms, theoretical approaches and bio-inspired and evolving technical systems.
Schedule
Schedule
The schedule is tentative. Times below are local Halifax times.
Abstracts will be provided below.
Tuesday, July 14, 2026 (09:00-12:30)
Session: Model Systems for self-organization, evolution and computation
Jan Bielecki, CAU Kiel, Germany (9:00-9:30)
P. Robin Hiesinger, Free University Berlin, Germany (9:30-10:00)
Vikram Chandra, Harvard University, USA (10:00-10:30)
Coffee Break (10:30-11:00)
Celina Juliano, UC Davis, USA (11:00-11:30)
Session: Theoretical Approaches to understand network computations
Rainer Engelken, University of Illinois Urbana-Champaign, USA (11:30-12:00)
Ekaterina Gribkova, University of Illinois Urbana-Champaign, USA (12:00-12:30)
Wednesday, July 15, 2026 (09:00-12:30)
Session (continued): Theoretical Approaches to understand network computations
André Longtin, University of Ottawa, Canada (09:00-09:30)
Session: Evolving Bio-Inspired Systems
Thomas Novotny, University of Sussex, UK (9:30-10:00)
Sebastian Risi, IT University of Copenhagen, Denmark (10:00-10:30)
Coffee Break (10:30-11:00)
Mark Miskin, University of Pennsylvania, USA (11:00-11:30)
Elisabetta Chicca, University of Groningen, The Netherlands (11:30-12:00)
Panel discussion (12:00-12:30)
Abstracts
Abstracts
Jan Bielecki: Seeing, Avoiding, Learning: Visual Cognition Without a Centralized Brain
The box jellyfish Tripedalia cystophora navigate cluttered mangrove habitats using a unique visual system: 24 eyes distributed across four rhopalia, 8 of which are vertebrate-like camera-type lens eyes that are known to modulate visual behavior. Each rhopalium is a self-contained sensory-motor unit capable of generating swim control independent of a central brain. This talk introduces Tripedalia as a model system for visual information processing in a nervous system that lacks centralized control, and presents three recent advances. First, obstacle avoidance, long known as an innate behavior, can be experimentally shaped, demonstrating that this animal is capable of associative learning despite the absence of a conventional brain. Second, animals use directional light-shaft detection to actively acquire targets, extending the known behavioral repertoire beyond reactive avoidance into goal-directed visual search. Third, a model agent trained with only 16 hidden model neurons reproduces obstacle avoidance behavior, offering a concrete hypothesis for the minimal circuitry sufficient to generate this behavior in vivo. Together these results position Tripedalia as a system in which visual learning and decision-making can be studied at the level of identified circuits.
P. Robin Hiesinger: Neuromorphic Growth in a Biological Brain
Artificial neural networks underlying current approaches to artificial intelligence are designed to contain little to no information before they are switched on to learn. By contrast, biological brains grow through a genomically encoded processes, are not switched on at any distinct time point, and contain a lot of information prior to learning. What information, if any, is missing in today's AI given that there is no genome and no growth process? How much - and what type of information - exists in a biological neuron and network compared to synaptic weights of a pre-trained artificial neural net? I will approach these questions from the perspective of developmental neurobiology.
The genome does not describe the brain, it contains information to grow the brain. The genetically encoded growth process has surprising features: it is non-deterministic, flexible and robust to perturbation. Yet, adult neural circuitry is precise enough to ensure function. Moreover, there is no fundamental limit to how much information the growth process can encode in a neural network – prior to any learning. Self-organization is key to these features of genetically encoded brain development. Individual neurons need to fend for themselves and make local choices. Our findings using Drosophila as a biological model highlight that pattern formation during growth and the kinetics of live neuronal interactions restrict synapse formation and partner choice for neurons that are not otherwise prevented from making incorrect synapses in this system [1, 2]. The seminar will explore self-organization of neural network topology with a focus on information theory and how growth leads to precise, flexible and robust outcomes in brain wiring.
[1] Agi, E*, Reifenstein, ET*, Wit, C, Schneider, T, Kauer, M, Kehribar, M, Kulkarni, A, von Kleist, M*, and Hiesinger, PR* (2024). Axonal Self-Sorting Without Target Guidance in Drosophila Visual Map Formation. Science, 2024 Mar 8;383(6687):1084-1092.
[2] Kehribar, M., Wit CB, Krasikova K, Agi E, Reifenstein ET, Wolterhoff N, Wriedt LQ, von Kleist M, and Hiesinger PR (2026). Selective adhesion preserves eye patterning as axonal retinotopy in the Drosophila brain (2026). Current Biology, 2026, doi: 10.1016/j.cub.2026.01.007
Vikram Chandra: Distributed neural computation and the evolution of the first brains
The origin of brains was a landmark in animal evolution, enabling new behavior and natural histories. We do not know how the first brains were organized or how they functioned. Acoel worms, the likely sister lineage to all other animals with brains, offer a unique window into the possible functional organization of early brains. Here, I will present recent work in which we develop the acoel Hofstenia miamia, a marine predator that hunts planktonic invertebrates, as a new model for neuroscience. We found that H. miamia has an unusual ‘diffuse brain’: a subepidermal network of dense neuropil exhibiting little regionalization or stereotypy in its gross anatomy or distribution of neural cell types. Remarkably, we found that behavior in H. miamia is robust to large, arbitrary amputations of brain regions, suggesting that most regions can perform most computations. More brain tissue improves performance, especially on challenging tasks, but no specific brain region is required. These results lead us to propose that H. miamia’s brain is composed of computationally pluripotent “tiles” that interact to generate coherent behavior. This architecture suggests a trajectory for nervous system evolution in which early brains may have arisen through the condensation of ancient diffuse nerve nets into unregionalized brains, with regionalization evolving secondarily. I will end with some thoughts on how we might use acoels and other marine invertebrates to study the evolution of neural architecture, computation, and behavioral affordances.
Celina Juliano: Hydra as a model for understanding nervous system design, function, and regeneration
The freshwater cnidarian Hydra possesses a simple diffuse nerve net that undergoes continual neuronal replacement throughout adult life and can fully regenerate following injury. This provides an opportunity to investigate how nervous systems maintain function despite ongoing cell turnover and tissue loss. Recent advances in single-cell genomics, molecular genetics, and whole-animal imaging have enabled the construction of a molecular and spatial atlas of the Hydra nervous system, the identification of neural circuits underlying discrete behaviors, and the characterization of developmental trajectories that generate neuronal diversity. In this talk, I will discuss how these resources are being used to uncover the design principles of a simple yet dynamic nervous system, from neuronal identity and circuit organization to continual cell turnover and regeneration. Together, these studies establish Hydra as a powerful model for linking molecular mechanisms, neural circuits, and behavior at the scale of the whole organism.
Rainer Engelken: Sparse Chaos and Event-Based Computation in Recurrent Spiking Networks
Brains and neuromorphic systems compute with networks that are sparse, recurrent, nonlinear, and event-based. This raises a basic design problem: how can such systems remain reliable without losing the rich dynamics needed for flexible computation? I will discuss this question using Lyapunov spectra as a quantitative language for stability, sensitivity, and state control in recurrent spiking networks. In balanced networks, the discreteness of synaptic input and the rapidness of spike initiation can qualitatively change the collective dynamics. Increasing spike onset rapidness transforms conventional dense chaos into a regime of sparse, localized instability, in which long periods of near-stable dynamics are punctuated by brief bursts of sensitivity. External input spike trains can then suppress chaos especially effectively, leading to reliable state control by sensory drive. I will also briefly discuss how event-based structure can be exploited computationally through SparseProp, suggesting a bridge between dynamical systems theory, biological design principles, and efficient neuromorphic computation.
Ekaterina Gribkova: Evolution of Memory: From Basic Foraging Decisions to Cognitive Map Construction
Cognitive mapping builds internal representations of the world and is essential to episodic memory and mental imagery. Using a bottom-up modeling approach we show how circuitry of basic foraging decision can be straightforwardly expanded for affective valuation and cognitive map construction in an agent-based foraging simulation, ASIMOV, reproducing likely potential evolution. In foraging, behavioral choice is governed by reward learning and motivation which interact to assign subjective value to sensory stimuli. These qualities characterize foraging generalists in variable environments and are precursors to more complex memory systems. ASIMOV's core behavior model is based on neuronal circuitry of cost-benefit decision in the predatory sea-slug Pleurobranchaea californica. The ASIMOV agent affectively integrates sensation (olfaction, nociception), motivation (hunger), and learning to make cost-benefit decisions for approach or avoidance of prey. We expanded the ASIMOV model with the Feature Association Matrix (FAM), which shows how simplest mechanisms of classical conditioning give rise to more complex sequence learning, and even a simple episodic memory which enables spatial learning for obstacles and distant landmarks in the ASIMOV agent. These model expansions show how the neuronal circuitry of foraging decision may have served as the framework for cognitive mapping in evolution.
André Longtin: Online extraction of novelty: noise-cancellation the hard, biological way
Early on in evolution, organisms have had to learn to differentiate between sensory stimuli they generate, e.g. by moving, vs those generated by outside sources. Further, and often relatedly, they must attenuate redundant stimuli to highlight - and even predict - novel stimuli. Higher evolved animals make use of a corollary discharge that warns sensory organs of impending stimuli caused by self-movement. Distinguishing self from non-self inputs can nevertheless still be tricky. Also, learning and disambiguating such percepts must continually co-occur with action and perception in biological systems, in contrast to artificial networks with distinct learning and testing phases. After reviewing these related general biological goals, I will present the intricate mechanism by which a south American wave-type weakly electric fish can continuously filter out redundant low frequency signals online, i.e. without a corollary discharge – in contrast to its African pulse-type cousin with such a discharge. They can quiet down single sinusoids, superpositions of sinusoids, and even low frequency random noise, akin to noise cancellation headphones. The mechanism
relies on a “teacher” pathway that continually synthesizes a negative image of the predictable signal to be subtracted via the cerebellum from the “student” pathway. A cleaned-up signal is then sent by the student towards higher brain where useful novel
signals (like food) emerge. This involves a complex set of dynamical ingredients: bursting from back-propagation, STDP , multi-delay feedforward circuitry, intrinsic noise and input signal modulation. This “hard way” of achieving the stated biological goals likely provides
the flexibility for the known multi-tasking of these cells and circuits. This algorithm can also be viewed as a special form of reservoir computing that learns to separate and even predict signal mixtures.
Thomas Novotny: Training SNNs with exact gradients: Progress and Challenges
The “Eventprop” method published by Wunderlich and Pehle In 2021 uses the adjoint method to calculate exact gradients in spiking neural networks (SNNs) of LIF neurons. It has successfully been used on proof of concept, small machine learning problems, and we were able to extend it to a larger class of loss functions to solve mid-sized problems, such as keyword recognition on the SHD and SSC benchmarks. We also extended it to networks with delays and to learning delays in addition to weights. When run on neuromorphic hardware, trained networks suffer minimal accuracy loss and realise considerable energy savings.
Pehle’s PhD thesis also contains equations for general hybrid SNN models beyond the scope of the Eventprop paper. Based on this, we have used Python’s sympy symbolic math package to implement an “auto-adjoint” method in our mlGeNN software, not unlike the powerful “autodiff” methods in PyTorch, TensorFlow or JAX. This allows users to define neuron and synapse models of their choice, which mlGeNN automatically turns into equations and code for gradient descent on the chosen model.
However, there are issues that mean that adjoint learning in SNNs is not yet as plug-and-play as autodiff-based error backpropagation in time in artificial neural networks.
In this talk, I will briefly introduce the overall method and will then discuss the main advances mentioned above and the challenges for a wider adoption of this technology, drawing on examples from our recent works.
Sebastian Risi: From Self-Assembling Networks to Self-Adapting Machines
This talk discusses how principles from the evolution and development of biological nervous systems can inspire a new class of adaptive artificial systems. Rather than treating neural networks as fixed architectures with hand-designed parameters, we explore how artificial networks can be grown through compact "neural developmental programs", local interactions, neuronal diversity, and lifetime plasticity. This perspective connects neuroevolution, developmental biology, and computational neuroscience by asking how complex adaptive behavior can emerge from simple developmental and learning rules. The talk will also highlight applications to adaptive control and robotics, where grown and plastic neural networks can improve robustness under changing conditions, including novel environments and morphological perturbations. More broadly, it argues that understanding how nervous systems evolve, self-assemble, and adapt offers a promising route toward machines that are more resilient, embodied, and capable of continual adaptation.
Marc Miskin: Learning Diffusion Models on Physical Electrical Networks.
Though diffusion models are at the heart of AI image generation, their roots run through non-equilibrium physics. These models learn to time-reverse random processes, forcing a swarm of random walkers back to their initial, unknown distribution. Yet currently, both training and reconstruction live in digital simulations, imposing tremendous energy costs to execute. Here, we suggest that, given the strong physical foundations of these models, they might be better realized using physical substrates that embody diffusion. Specifically, we leverage the formal correspondence between Markov chains and electrical networks, using the voltage in the network as a physical representation of state probability. We present a local-rule training algorithm that can encode the time-reversal operator in electrical conductance, having only seen examples of the chain running forward in time. Our approach is supported by both theory and numerical demonstrations of image generation and filtering. Finally, we present a preliminary hardware realization, based on sub-threshold transistors, paving the way to scale up with microfabrication.
Elisabetta Chicca: Event-Based Vision for Obstacle Avoidance and Egomotion Estimation
To navigate complex terrains, biological organisms rely on distributed, asynchronous neural units that effortlessly transform spatio-temporal visual cues into motor actions. In this talk, we demonstrate how modeling these insect pathways allows us to identify robust computational principles and better understand biological navigation.
First, we present an obstacle avoidance network inspired by insect neurobiology that robustly replicates animal-like navigation. By combining a Time Difference Encoder (TDE) with a shallow spiking network architecture, our system generates emergent closed-loop behaviors across diverse environments using a single set of parameters. Second, we examine how global self-motion is resolved by extending this pipeline to match the properties of fly T4, T5, and VS neurons. By integrating visual event streams through complementary matched filters, we mimic how the fly brain tracks complex optic flow fields like clockwise, anti-clockwise, forward, and backward motion. Finally, we demonstrate that when these parallel biological principles are directly mirrored in neuromorphic silicon hardware, they achieve high tracking precision with an ultra-low mean power of 914 pW. Ultimately, this end-to-end spiking framework serves as a concrete working hypothesis, yielding testable predictions that can be directly validated in future neuroscience experiments.
Sponsor
Sponsor
Page updated
Google Sites
Report abuse