Day 3 - How animals and robots navigate the world - Andy Philippides, Jason Kerr, Andrea Chiba, Yulia Sandamirskaya

 

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

Morning started out gray with scudding clouds following a good rain last night.  Just before breakfast started at 730am a solitary snorkeler Ed and runner Chiara appeared in the bar area along with Elisa who arrived late last night.

Yulia Sandarsmirskaya opened the session with a short story about a mollusk (I think she is referring to the genus Saccoglossus) that has a brain until it finds a rock to attach to and then digests its brain.

What we get from this is that we often think of neural mechanisms as ways to deal with complicated things like planning and language, but locomotion can be a big driving factor.

Putting it more briefly, navigation can get you further than you think.

Then was the turn of Andy Phillipedes, Professor of Biorobotics at the University of Sussex, who started by giving an outline of his talk: intriguing examples of vision from the animal kingdom, some theories and then some mechanisms people have come up to explain animal behavior.

Andy's field is neuroecology. He does not care so much about vision in itself but in how it evolved as different species are solving a particular task and specializing in their niche.

The first animal example is ants (which seem to be loved in this community as this is just the second lecture of the workshop and the second time we talk about them). Andy has worked with wood ants in Sussex and dessert ants in Australia. Very hot places are interesting for the study of ants because they cannot rely on pheromones: they evaporate too quickly to be of any use. So instead ants employ vision and other mechanisms that we will see next to navigate. Another thing that makes high temperatures interesting is that it enables ants to create their niche there, after other species have gone extinct.

The food source of these ants is temporary so they do not evolve strategies for having stable roots. Instead, they are born with innate mechanisms that enable quick adaptation and exploration of new food sources.

The ant colony contains a couple of reproductive queens and very specialized foragers (I think they are female). These foragers are constantly running around, locating food sources, returning to their nest and going back to the food sources, so they are useful for collecting data at a scale.

After setting the stage,  Andy moved to discuss a specific problem that these ants face: navigation from their nest to food and back.

 

 

An ant starting from its nest and following a path to the food and back

 

In the beginning, the ant does not know where the food is so it needs to explore. It does so by picking a random angle. Once it finds the food the ant returns back and then does the same trip. How does it remember the path?

These ants have a low-resolution vision that they use depending on context. But their impressive feat is their ability for path integration (also known as dead-reckoning). The ant has the ability to sense the distance and direction of its nest and constantly track its position (through the compass neurons we talked about on Day 2. Gaby Maimon, who is an expert in this problem, explained that the ants have 8 sinusoids where the amplitude gives you the depth and the angle the direction ). In general, these ants weigh cues from different modalities based on the uncertainty of the que.

People have done many experiments trying to mess with the ant's job. For example, if, after it returns to the nest, you move the nest while the ant is blind it will move to the wrong target (that would have been the correct one if you hadn't moved it). Another example is that if you stop them before they reach the nest and place them at the food they will start moving in the wrong direction (that would have been correct if you had not displaced them). They can also overcome obstacles that come on their way. Or if you are very mean and blow them with a wind blower a few meters away from the path they will come back to it after, as they clatch on the floor and estimate the wind direction

Different ants follow different paths. They do not follow each other, because foraging in their niche is very uncertain. In contrast, bumblebees communicate the location of food to others when they return to their nest. This is a nice example of different species solving the exploration/exploitation dilemma in different ways depending on ecological conditions.

We then moved to another impressive innate ability of these ants: they can do single-trial learning based on visual cues. For example, if, on their way back to the nest you move them up to 2m away from their path they can find the right way and use it in the next trial.

Perhaps the easiest way to do this is this: with every step, you take a screenshot of your visual field and, since you know your direction, you associate it with the shot. Nearby shots are similar and have similar directions.

We do not know exactly how many screenshots the ants can hold in memory simultaneously, but since they are wide-filed and low-resolution, they are very efficient.

We have also seen that once the ant reaches its nest it follows a convoluted path around it to take a lot of shots from different direction.

We then moved to the next speaker: Jason Kerr from Max Planck Institute for Neurobiology of Behavior. The motivation for this part will be understanding how trade-offs arise in visual systems and how different animals have come up with different solutions but still exhibit commonalities that reveal shared costs and trade-offs.

What is vision good for? Examples are navigation, danger recognition, and orientation. And what trade-offs come to mind? The audience mentioned the problem of removing your self-motion to extract useful information

Our first example species is the rodent. Rodents live in the ground so it's important for them to look up. They have a very large field of view (about 200 degrees)

 

                     

On the left, a rodent's eyes can be seen from the top to the side and even from behind.

The problem that the rodent needs to solve is that when their head rotates their eyes counterrotate as an opto nystagmus reflex to stabilize the vision. This means that their vision depends very much on gravity. They also don't have a fovea. How do you chase a target when your vision is so restricted?

Let's say that the rodent is chasing a cricket. To catch it, it locates it due to its wide field and then runs head straight to it. As it moves toward the cricket its optic flow changes: this is information about the distance and direction to the target. The reason why they need to focus their head on the target is that they have a specialized area in their retina that contains a particular cell type that gets all the optical flow (as it has good accuracy).

So rodents were an example of how you can trade off the need to respond to predators (requires peripheral vision) with the need to hunt (requires precise local vision).

Next animal is the Ferret where there is an interesting saccadic movement quite different from the human ones. They have a fovea, a wide field of view and they really like chasing prey.

We have experiments with ferrets chasing a ball. To do that they focus their fovea (aka area centralis) on the target.

But what about corners? When a ball turns around a corner, the ferret needs to do two tasks: navigation and target following. It seems that these two tasks are solved by two different mechanisms. The still use their fovea for a target following but for navigation, they use a new mechanism: a saccade that happens when the ferret turns its head, and the eye turns in the opposite way. This creates a bump in the optical flow. Then the eye turns back. In contrast to the human saccade, this one does not provide visual information but navigational cues, allowing the ferret to efficiently navigate its environment by sensing changes in the optical flow during head movements.

We then moved to birds, very different from ground species as they are the ones hunting the most.

The story is about the Canadian goose. These animals have two foveas: a high-resolution pointing to the sides and a low-resolution pointing to the center.  They are also interesting in a collective: when a group of geese is feeding all of them are pecking except for one that keeps its head about to keep panoramic vision. 

How do these geese hunt individually? People had them hunt for prey and collected data about whether their head was up (panoramic vision) or down (targeted vision)

On the right, we can see when the goose has its head up and when down. It follows a binomial distribution

A quick coffee break where you need to grab a coffee either very early or very late to avoid the queue and move to the next part of the talks.

Third Speaker: Andrea Chiba will talk to us about understanding behavior in the context of both brain and body.

We first defined what is affordance: it is a design in your environment that enables you to do something.  For example, a cup with a handle enables you to hold it more easily with one hand.

 

 

She then asked people who have worked with robots what kind of limitations they encounter: battery distance, memory versus space, robustness, speed and latency were mentioned among other.

 

Notably, Chiba and Florian admit being very bad at navigation (one cannot but wonder if  this motivated their work?)  

Andrea mentioned that the Nobel Prize was awarded to someone discovering the GPS of the brain (probably referring to John O'Keefe, May-Britt Moser, and Edvard's work, chatGPT is much more useful than Google search for this kind of trivia).

 

Andrea drew a (convoluted) diagram of the brain (with many Latin names and abbreviations that were difficult to note down). You can see the eyes, the visual cortex being far from the eyes and different cortex areas that are wrapped in a c-shape.

 

Bottom: different cortical areas in the brain are wrapped c-shaped layers. Top: different types of cells useful for different navigation tasks

The main point of Andrea was that, depending on the behavioral task you pose on the organism (in this case rodents),  you will discover the different kinds of cells: grid, axis and place cells.

 

Andrea, in the face of a lack of slides and videos, recreated an impression of how the cells fire as the organism moves:

 

 

Andrea demonstrating how neurons fire by walking and making cell sounds

If you record the hippocampus when the rodent is walking, you will see the place cells firing. But these cells are quite dump: if you repeat the same task every day the will need retraining each time.

Experimenters had to study rodents while chasing in order to discover grid cells.

And to discover axis cells (cells that encode direction) they had to place them in mazes.

These different cortical areas collaborate and sometimes don't need the hippocampus. For example, you can take the hippocampus out of a rat and it will still show evidence of remembering locations.

She then talked about rhythms in the brain and how cells fire in correlation with them. For example, here's a theta rhythm (extending across two papers) and the place cells firing at the beginning at the peaks and progressively eariler with each peak. This gives the organism a temporal code aligned with the spatial code, which offers a way to implement both retrospective and prospective coding.

 

Crucially, the different areas offer coding at different scales and the cortex needs to multiplex all this information.

We then moved to understanding how movement happens. Even if we know that the thalamus offers head direction, there are still many questions. How do we place the information of place cells in a sequence in order to remember trajectories? 

There is an area that is very close to the virtual and motor cortex and provides inputs to the hippocampus and collects information from many other ares. It is called the retrosplenial cortex. When this area is damaged in people we know that they remember things, like where they have been but not the sequence of how things happened. So it must be a crucial area for sequencing.

Another example of how navigation can be weird: all of us have had occasions when driving/going back home and remembering how we got back there. You don't actually need the memory of your trajectory so your hippocampus is not involved, instead, your cortex is occupied by planning.

A concluding remark from Andrea particularly relevant for robotics: if your robot get's stressed or confused due to noisy information, you would like it to enter an automatic mode and return to its "nest" nevertheless. Understanding how animals leverage the immense redundancy in brain signals to function even when different areas functionalities shut down can be a good step toward this autonomy.

 

Final speaker: Yulia Sandamirskaya from Zurich University of Applied Sciences talked very fast so was hard to write down. She works on (her exact words were "worships") dynamic field theory and probably a good place to start to understand her work is https://dynamicfieldtheory.org/. She is interested in solving behavior through simple neural networks, without the complex tools that mathematical analysis often imposes, and implements them in neuromorphic hardware as it is best for efficiency. Another argument in favor of biological inspiration is that nature is the only way we know to work.

 

She first mentioned two problems for navigation with robotics: SLAM (Simultaneous Localization And Mapping), where you predict the next map and location based on a series of observations and place recognition (similar to our ant example). When building these biological mechanisms on robots we need to take into account two constraints: 1) autonomy, meaning that you don't have someone tell you when to do something and how and 2) using a neural substrate.

We then looked at a specific robotic task and went through the whole process of designing neuromorphic hardware to respond to different challenges: a robot can turn around its axis and is located in a room with various objects. It needs to be able to reach out to objects and recognize when an object has changed. The example provided by Yulia was about a robot surrounded by LEDs that are blinking at different frequencies, and the task of the robot is to reach or direct towards the specific LED. The robot should also be able to detect if there are changes in the whole surroundings, e.g., if one LED stops blinking.

On the right, a schematic of the neuromorphic circuit solving the tas

 

Yulia went through increasing steps of difficulty. First, the robot needs to recognize its direction. It can do so using the compass neurons, that are by now known to us. The challenge arises when we want it to recognize rotations of different speeds and directions, so we cannot hardwire them. But we can create asymmetrical connections in a two-layer network, which will give you a gradient. 

And what if I want to detect how quickly the robot is spinning? You can do it through the activations of neurons but that would require updating all connections. Instead you can add a motor neuron that controls for it.

Next, I want to make the neuron plastic, so I want to add some learning rules, like Hebbian learning. You then need to come up with a network that will allow you to compare two representations encoded in two populations of neurons. One population shows you your desired rotation and one your current rotation. You can then build something like a lookup table that subtracts the two. This is much more useful than having a single neuron to represent the error, as that would not contain enough information for the learning algorithm

The session concluded with a discussion on how one can build such systems, especially in the future. Unfortunately, we do not have a way to translate a task (e.g. what your CEO asks for) to software running on some hardware. Instead, we need thousands of engineers and programmers working on writing compilers. Could we imagine a future where Large Language Models will solve this issue for us? Some skeptics and some optimists in the crowd, but the discussion was dissolved as both prioritized moving to the lunch break.

!!! LUNCH BREAK !!! Filled with fruitful discussion and delicious food.

Later, everyone joined their workgroups and started working on their selected projects.

The afternoon was filled with workgroup meeting, sports and "real work". Tomorrow will be another exciting day and perhaps more sunny weather. Till tomorrow goodbye!