The physical implementation of today's technologies (very much including neuromorphic electronics) depends almost entirely on external (to the instantiated object) factories. By contrast, biological systems construct, repair, and evolve, themselves.
This discussion will consider issues such as:
Link to projected handwrittne notes of today's discussion:
https://cloud.capocaccia.cc/s/dMXmnJmebQeGJLG
Link to neuromorphic platforms table:
https://docs.google.com/spreadsheets/d/1taHE4maC2vdloPLpZZJhElE6kZOA-Nx9GHjdPId8uXg/edit?usp=sharing
Logistics (Giacomo, Committe, Organizers)
History (Rodney, Tobi, Andre, Giacomo)
Map of Interests (all):
1. [BBICS] Building Brain-Inspired Computing Systems (22%)
2. [UNCBB] Understanding Natural Computation by Building (21%)
3. [GBARS] Generating behavior in autonomous robotic systems (18%)
4. [R&IPS] Representations and Information in Physical Substrates (14%)
5. [DNP&A] Develop new …
We invite everyone that has taken the time to use Spinnaker2, that have experience with Spinnaker1, or have some intuitions, ideas or suggestions about the direction S2 could take to join for a feedback / discussion session !
We have a fully implemented baseline pipeline for Human Pose Estimation with Event Cameras in Real time that, then, feeds into a full body Fruit Ninja application. The system is ANN based and implemented modularly. A webpage will be made available shortly. If you would like to try seeing …
The goal of this workshop is to investigate how to make the best use of the unique properties of neuromorphic processing solutions to enable natural interactions with social humanoid robots. We would like to collaborate with workshop participants in order to find the most suitable applications for neuromorphic substrate to …
In this workgroup we will investigate the computational capabilities of DYNAP-SE heterogeneous neurons for compressing time series data. Our focus is on exploring how these neurons can efficiently compress both periodic and aperiodic signals using few spikes per neuron, enabling low-latency pattern detection. We will optimize the DYNAP-SE neuromorphic chip …
Fancy running your SNNs 10x faster? Our GPU enhanced Neuronal Networks (GeNN) library is freely available from https://genn-team.github.io/ and provides an environment for GPU accelerated spiking neural network simulations. GeNN is capable of simulating large spiking neural network (SNN) models at competitive speeds on commodity and even embedded GPUs. In …
We'd like to explore self-assembly in neural celluar automata and visualize their behavior in a control task.
Inspirations:
Robot:
SynSense is a Nueromorphic startup founded in 2018 with offices in Zurich and China. We have two family of SNN processors: Xylo for low dimensional signal processing and Speck for vision processing. Both HWs come with open source and user friendly python-based librarys: Rockpool and sinabs that have been co …
Argo is a 65cm hobby sailboat outfitted with RPi computer, solid state wind sensor, GPS, and IMU. The aim is to learn how to get it to sail itself.
This is an umbrella workgroup which aims to organise the mass of related workgroups. We'll hold an initial session where there'll be a bit more time to introduce all the workgroups in this area than will be possible in the main session and then meet every few days to try …
Humans and animals seamlessly control their complex bodies, use new tools, and adapt to injuries.
A highly adaptive body model of some sort (body schema/body image, forward and inverse models, etc.) seems indispensable.
Using the Mujoco simulator we simulate infant-like full-body tactile agents such as iCub and MIMo environment. We …
Lu.i is an electronic neuron circuit mimicking and illustrating the basic dynamics of real, biological neurons. The printed circuit board features a configurable, fully analog implementation of the leaky integrate-and-fire model and visualizes the internal state, the membrane potential, through a VU-meter-style chain of LEDs. The neuron emits a short …
Drone attitude is generally estimated from IMUs containing accelerometers and gyroscopes. In some flying insects, sensors like these are missing, posing the question of how they estimate their attitude. One hypothesis is that they do so using vision and knowledge of their own body and control commands. Recent work has …
Continually Learning Prototypes (CLP) is an algorithm developed for Loihi 2 to address online learning of novel patterns in an incremental fashion without forgetting. In this meeting I will give a quick intro to the algorithm and the three applications we are investigating during this workshop:
This workgroup is about robotics with accelerated neuromorphic hardware. In particular, we will provide multiple BrainScaleS-2 systems that feature emulated neurons with dynamics 1000 times faster than those found in biology.
All systems are equipped with multiple digital and analog real-time interfaces for robotic applications. We can inject and extract …
We bring along multiple BrainScales-2 systems and invite everyone to try some of our demos or implement their own neuromorphic experiments.
The most recent generation of BrainScaleS-2 ASICs features 512 analog multi-compartment neuron circuits with 256 plastic synapses each.
The analog circuits are tightly coupled to on-chip digital event routing …
The ability for learning models to be deployed in new and challenging conditions is a long-standing aspiration of modern AI. Models computing at the edge should be capable of both adaptation and knowledge-transfer, leveraging extensive (offline) pre-training while remaining nimble and avoiding catastrophic forgetting.
In this workshop we …
We will bring a Neuromorphic Drummer to Capo Caccia and launch a neuromorphic drumming control problem, as a low-cost and reproducible toy model for neuromorphic locomotion control.
The problem consists in controlling the tempo and velocity of swinging pendulum + drum-pad system through a fully neuromorphic sensing-control-acting sensori-motor loop.
The …
We will have two Allen Brain researchers with us introducing their last datasets to inspire our modeling and design choices!
We will discuss about the differences and trade-offs between different design styles and hardware implementation solutions for implementing spiking neural network processing systems.
Participate in our ongoing study to understand human behavior in different cognitive load conditions. |
In this discussion session, I would like to identify critical components for a holistic approach to learning in neural networks together with other participants. The focus will be on how we can combine the best of the two worlds: offline training for creating the initial knowledge base and online training …
Workgroup created during the week with the idea of embedding an SNN based on the Central Complex of insects for a Path Integration task. The goal is to utilize a car-like robot equipped with an event-based camera mounted on top to estimate linear and angular velocity. Subsequently, these estimates will …
Event-based learning of delays in SNNs
In a collaboration between Melika Payvand's EIS Lab and Mihai Petrovici's NeuroTMA group, we've been developing a fully event-driven training algorithm for learning delays and weights in SNNs. In the established Fast&Deep algorithm, spike times are treated as the quantity central for information propagation. …
This is a 'fession session' open discussion - whose aim is share experience with the workshop.
No presentations, no company pitches. Just the lessons learned, for good and/or for bad.
Exact company affiiliations may (for reasons of NDAs) need to remain anonymous - this is a discussant's choice.
There have been investigations into developing standard protocols for connecting spike-based sensors & computational units but these have so far remained disparate academic efforts. As we try to take SNNs & neuromorphics into the commercial arena, having an easily usable & agreed standard will make it easier for groups …
Dendritic computation has become of great interest to the community as evidence from neuroscience has shown the importance of intra-neural mechanisms in creating realistic behaviour and computational experiments have demonstrated that these mechanisms (e.g. local learning within a synaptic cluster and homeostasis) may be of practical use in spike-based …
In the brain, synaptic learning is constrained to use spatio-temporal local information. However, the majority of the learning algorithms that we use to train neural networks violate spatial or temporal locality (such as RTRL and BPTT).
How can the brain efficiently assign credit using only local spatio-temporal information?
In this …
Gradient-based learning may not be what the brain is doing and there are lots of issues regarding data hunger etc but, lot of people are using gradient-based learning rules to train SNNs and other event-based models like eGRU to do really cool things. However, most of these event-based networks …
Participate in our ongoing study to understand human behavior in different cognitive load conditions.
Follow-up on Tuesday's session, focused on non-monotone (range-localized) dendritic nonlinearities and a proposal for CMOS realization. https://www.science.org/doi/10.1126/science.aax6239
Long time-constants are expensive and difficult to implement in hardware. Can long-term memory be formed without them? In other words, are long time-constants necessary?
While perhaps not necessary, bio-inspired mechanisms are arising in SNNs adding long time constant to improve the computational power (Adaptation [1], Delays [2]). So maybe long …
Understanding the relationship between genotype and neuronal circuit phenotype necessitates an integrated view of genetics, development, and plasticity/learning. Challenging the prevailing notion that emphasizes learning and plasticity as primary drivers, we delineate a tripartite framework to clarify our view of the respective roles that learning and plasticity might play in …
We will continue today's discussion with Wolfgang, about what type of biological constrains are interesting to include into SNNs, and how we can incorporate them so that they emerge in a self-organized manner.
Join us to be a participant in our study about human pose response to cognitive load.
Experiment duration: 10 min
Join us to be part of amazing study.
Compare/contrast various local rules that approximate gradient descent on a global learning cost function and how they are implemented in hardware. Eg electrical circuits that implement "coupled learning" to do supervised learning tasks on their own, without a processor. Seeking speakers!
Let's have an efficient and productive discussion on key differentiating elements of Computation vs. Control and sharpen our arguments for the need for a) closed loop systems; b) recurrent structures; c) feedback and control; d) handware-software co-design; etc. Let's also discuss continuous vs. discrete control in brains and machines and …
It seems as if the mainstream ANN & LLM models and techniques are moving inexorably towards sparser and lower-resolution encodings and representations. In this sense, they are converging on sparse 1-bit representations and learning mechanisms that have historically been inspired by computational neuroscience and which have at least a …
The discussion is related to all the new memory devices, not only memristors.
The idea is to share problems and solutions we found in our labs.
The problems can came from device integration, characterisation, circuit design, modeling, network simulation...
We need 10 more participants to finish our study about human pose response to the cognitive load. If you haven't participated yet and want to see your name on the leaderboard on Saturday, then be part of this scientific study. It will not take more than 10min per participant.
We will look into methods for computational modeling of neural development, in particular, agent based modelling. The software Biodynamo will be introduced.
We will discuss different representations for genomic bottleneck algorithms, desired properties, and the role of growth.
Walking meditation is an old meditative practice in which one focuses all their attention on the act of walking at extremely slow paces.
This simple practice can make us discover and become aware of the mechanisms that make us move in a direct, experential way.
Besides its recreational value, walking …
A series of training sessions in the beautiful art of grappling (Brazilian Jiu Jitsu, a.k.a. the art of folding clothes with people still in them). Only very few things are as satisfying as being in full control of your opponent while whispering softly into their ear ;)
For all skill …
We are a bit late, but let's try to get some bikes for tomorrow in the coffee break today to go to the beautiful grotto di Neptuno on Sunday (after breakfast, come back AFTER lunch).
Explore ways to present your research (workgroup) results using artistic expression
Meet your fellow partiipants