Keynotes

Neuromorphic AI for small, autonomous drones

Spurred by the quick progress in artificial intelligence, autonomous robots promise to revolutionize our economy and society. However, current-day AI requires excessive computational effort, leading to extra robot weight, costs, and energy consumption. This makes robots with AI large, heavy, and dangerous, which negatively affects their societal acceptance. To realize a future of competent and safe autonomous robots, AI should become orders-of-magnitude more efficient.

Neuromorphic sensing and processing promises substantially faster and more energy efficient robotic AI. In order to show its potential, in my lab we tackle the daunting challenge of developing a neuromorphic AI that provides full autonomy to tiny, severely resource-restricted drones. I will present our ongoing work on realizing capabilities such as ego-motion estimation and control purely with SNNs, and share our findings on the actual latency and power usage of existing neuromorphic chips. Finally, I will present Bee-Nav, our honeybee-inspired approach to navigation with tiny, efficient neural networks.

Online gas classification at the edge

Field applications in chemical sensing, from portable safety monitors to mobile robots, require models that are lightweight, fast to respond, and can be recalibrated from limited labelled data. Current gas classification pipelines assume controlled sampling cycles and offline feature extraction, and rely on computationally expensive models. With a few thousand neurons and sub-milliwatt power consumption, insects reliably classify odours in open environments.

Inspired by the insect olfactory system, in this talk we present a neural network architecture that classifies raw sensor array readings using a sparse, high-dimensional binary representation. We showcase evaluations on random sequences of single gases and binary mixtures, introducing new samples dynamically and classifying gas identity from the transient sensor response. In these settings, the architecture outperforms linear Support Vector Machine (SVM) and maintains online class-level frequency estimates, distinguishing familiar from novel stimuli. It relies exclusively on binary weights and Hebbian local learning, remaining computationally lightweight and well suited to ultra-low-power neuromorphic processors.

Finally, we present an on-chip demonstration of the network with a hazardous gas detection experiment running on the full-custom mixed-signal neuromorphic chip TEXEL. The on-chip classifier yielded an accuracy of 74.6% while running at 482 nano Joule per inference and 27 micro Watt of static-power.

From Sound to Meaning: What Human Auditory Cortex Can Teach Neuromorphic Computing

Human hearing transforms continuously changing acoustic signals into meaningful information about the world: who is speaking, what is happening, and which events matter. How does the brain accomplish this transformation, and which computational principles could inform neuromorphic listening systems? 

In this talk, I will present our research combining ultra-high-field functional MRI, perceptual experiments and NeuroAI modeling to investigate how the human auditory cortex represents natural sounds. I will show how this approach reveals sound processing at multiple spectral and temporal scales, and intermediate representations that bridge acoustic features and semantic interpretation. Comparisons with deep neural networks help identify which representations best explain brain responses and human perception. 

Building on these findings, I will outline opportunities for neuromorphic auditory computing, including architectures that integrate information across timescales and learning objectives that capture perceptually meaningful sound structure. I will also discuss how neural and behavioural measurements could provide benchmarks for evaluating such systems. The broader aim is to translate insights into human hearing into testable principles for efficient, adaptive artificial listening.

Neuromorphic Computing for Bidirectional Neuroprostheses: From Neural Dynamics to Sensory Feedback

Bidirectional neuroprostheses must interpret neural activity while also delivering useful sensory feedback, under tight constraints on power, latency, and bandwidth. Neuromorphic systems offer a way to explore both sides of this challenge. 

This talk presents complementary bottom-up and top-down approaches. First, a neuromorphic model of the somatosensory cortex is used to investigate how cortical network dynamics shape adaptation to intracortical stimulation. Second, a mixed-signal neuromorphic system emulates proprioceptive cortical responses from hand kinematics, preserving key temporal and population-level features in hardware. 

The talk also considers how spiking autoencoders can compress large-scale motor-cortex recordings while retaining information relevant to grasp decoding. Together, these studies show how biologically grounded models and efficient neural representations can support future neuroprosthetic systems, including closed-loop interfaces that combine neural decoding with tactile and proprioceptive feedback.