Day 2: Sensing in animals and robots - Florian Engert, Gaby Maimon, and Wolfgang Maass

The 2024 CapoCaccia Neuromorphic Workshop is the latest from the series that started in 2009, 15 years ago.  The workshop is taking place at Hotel dei Pini near Alghero on the island of Sardinia

Interested in updates about future workshops? Subscribe to the mailing list.

today's authors: Tobi Delbruck, Eleni Nisioti, Muhammad Aitsam, 

The second morning started out with snorkeling before breakfast under calm skies and no wind though rain is predicted for later.

Florian Engert from Harvard kicked off things at 9am  after being asked by Saray to put himself in one of the classes of interest. He stated he wanted to talk about control of behavior but Andre said he only gets 20 minutes about "touching himself". He then requested explanations of the categories which the crowd provided. He finally chose categories 1 (Understanding by Building) and 4 (Generating Behavior).  He told the story of how there was a mysterious song produced by Zebrafishes but it was discovered that the bone structure produces the song itselfHe then shared a conceptual epiphany in the context of goal directed behavior.  He stated that during development all that animals do is to minimize error. They compare a desire with measured state and try to reduce this error.   Animals have a "set point", i.e. a goal of particular blood oxygen level and then act to get closer to this goal/set point.

In studying baby zebrafishes, they try to stay stationary. They can get a long way towards this by minimizing retinal slip. Making this hypothesis and measuring the slip by tracking the fish you can then correlate it with the activity of 100k neurons. You cannot measure brain state (behavioral desire) this way but there are other ways to infer this.

Now, moving on to "Touching himself" and optimal control, he asserted that brains mainly simulate the world, or simply predict consequences of their own actions.  Like in MPC (model predictive control), but integrated with estimating state. The model mismatch (just as in MPC) should be reduced over time. 

Specifically, to study this you can study how the fishes *body* is represented in the fish brain and to see how model mismatch appears. 

There is distinction between efference (output) and afference (input). Ex-afference is anything out in the world. Re-afference is stimuli that are consequences of owns actions.   An operational definition that allows distinguishing these ex- and re- signals is a lack of correlation. 

Florian then went though some examples of re-afferent signals during action.

A great example came from audience is blinking. If you blink it is not surprising that it gets dark. If it does not, that means something is wrong. If you close one eye and nudge the other with your finger the world jiggles around, unlike when you move your eye either consciously or unconsciously.

For vision, zebrafish swim in bouts with tail flicks. During the swimming bout they are blind. 

For audition, there is reafferent cancellation of speech and chewing. 

For somatosensory (skin) there is tickling, you cannot tickle yourself. But insects crawling on you or tickling produces a huge surprise signal that is inhibited by efference copy if you create the sensation yourself. 

For olfaction, he posited that flatulation is another example, where humans can smell others much more than own.

That conclude Florian's comments at 1010am.

Next Gaby Maimon from Rockefeller took over and reported how in drosophila commands are sent to the midbrain to learn a mapping to the ellipsoid body compass reckoning system.

He started off by introducing Cataglyphis studied by Rudiger Wehner that has a system that can quickly vector the ant back to nest after finding foot over hundreds of meters. Fruit flies are much more studied and can be obtained with fluorescent labeled neurons.

Mike Dickenson studied this in desert with flies released to banana baits 1km from release site and the ones that make it much fly straight.  So in this condition of being hot without food and in sun it triggers a behavior to fly straight starting in a random direction.  therefore since they are at the limit of endurance. So by timing they must fly straight. 

In the central complex consisting of PB (protocerebral body??) , EB (ellipsoid body), FB (fan shaped body), and NO (nodulus) sketched here consisting of about 10k cells in total, the "tangential" neurons and columnar cells connection in and out of the central complex. They all fire sodium spikes and all existing theories are rate based which.

Key labs are Rachel Wilson, Vivek Jayawan, Gaby Maimon

It has been discovered that EB has a bump of Ca++ activity in the dendrites of the tangential neurons the wrap around the EB. The bump rotates opposite the direction of the fly's turning while walking. The reference is typically the sun or some other reference object in the environment.

The coffee break was filled by furious argument about the meaning of signals in the EB.

Gaby continued with discussing how the EB is connected to other blocks. 

The experiments they and Viveks lab have done show that this bump sustains itself and cannot be made 

much wider or move around much. 

When the flies turn 90deg in darkness the bump in these "compass neurons" also moves so it has an efferent copy of its turning.

In this system structure relates to function very tightly. 

In the PB/EB anatomy, the wiring is such that there are 45 deg offset connections that cause this rotation.  When the fly turns the red pathway activates to cause the activity to move. 

Next "landmark association".  The bump is somewhere in darkness, now turn on the sun. The "ring neurons" synapse onto the compass neurons. The ring neurons have a  small RF in a particular direction (one of 200). The particular wing neuron fires and strengthens the connection to the active ring compass neuron. 

Now the fly turns right 45 deg. Now the bump moves CCW 45. Another ring neuron gets activated. Over time these ring neuron connections are mapped to the correct ring neurons for that particular environment. In this way the fly can learn into synaptic connections a useful mapping for each new environment over a period of a few minutes or less.

Gaby concluded with a story about drosophila copulation that lasts 21+-0.5m. The male ejaculates after 6+-0.5m. How is this precise but very long timing work? In a set of 4 neurons there is a coupling from electrical to biochemical phosphorylation then back to electrical. It is like the formation of the EB mapping to a particular environment, which mixes spiking with synaptics formation.

Gaby had a couple of summary points:

1. There's no central clock
2. Cell types are important
3. Structure results in function
4. Learning only occurs in very specific synapses in this EB mapping setup
5. Ants/flies/locusts? all have 50 of these EB neurons, so it is conserved structure
6. Keeping ant in box for 2 days it can still remember the way back to nest, but not 3 days.

Wolfgang Maass took over pointing out that grandmother neurons are special in terms of learning rules (compared with V1), data structures and something else (didn't catch that). And these areas also have neurons that encode relations.

His talk was about a recent paper fom Stocekl, Yang and Mass, Nature  2023 "Local prediction-learning in high-dimensional spaces enables neural networks to plan".

Observations (sigma) and actions (a) are embedded in high D space S as Q and V which is "cognitive map" in this illustration and many operations can take place in S.

The goal of learning for Q and V are to predict the next Qt+1 from current Q and action a with its embedding V such that you can move the system in a desired direction in observation space.

Therefore in contrast to the need for a value function for RL, you can just learn this mapping from learning how to predict the next Q from the current one and the embedding of the action you take. This Cognitive Map Learner (CML) could be suitable for rapid online learning in some applications with moderately high dimensionality.  The key idea is that to get to a desired observation (the goal) you choose the action with the biggest dot product with the Q vector pointing in the desired direction.

Once the mappings are learned, structure emerges in the mapping when reduced to lower dimensionality so that this rule makes sense.

The paper shows this approach can drive a simple simulated insect robot. 

It would be interesting to see if this approach scales to problems like the ones so convincingly solved by RL in high D robotics, e.g. https://www.science.org/doi/10.1126/scirobotics.abc5986  as shown in the video 

After this summary of the paper there were many raised hands but it was time for lunch so discussion continued there energetically.

After lunch workgroups started meeting to discuss their projects, starting with all the learning-related projects. 

Just before dinner there was a lively no-money poker game by the bar. 

The evening session had an excellent introduction to using Xylo from Mina Khoei (Synsense), which is a simple tiny digital SNN chip with 16 input neurons, 1000 hidden neurons, and 8 output neurons and a handy toolchain that makes it a lot easier to train for standard simple and shallow MLP and maybe vanilla RNN architectures.

There was lots of intense discussion of projects in the disco. 

Finally the first-ever night snorkeling thanks to two aliexpress underwater headlamps, where the two French Gabriels and Tobi saw several octopus and cuttlefish in spooky darkness. The big octopus was not afraid and we could get within a foot of it. But none of us dared to touch it!