Can the Insect Path Integration Memory be a Bump Attractor?

Ioannis Pisokas, Matthias H. Hennig

Preprint posted on 7 April 2022

How can insects remember how far they’ve travelled if there are no landmarks? Investigating the “bump attractor” hypothesis.

Selected by T. W. Schwanitz

Background: A “bump attractor” is effectively the neuronal equivalent of that little line in videogames that tracks how far you go, a stamina bar—connected neurons are lined up together topologically with both inhibitory and excitatory synapses. Excitatory synapses link nearby neurons along the line, while inhibitory synapses link neurons that are further away. Any given neuron along the line excites its immediate neighbors and inhibits those further down the line in both directions, thus creating a “bump” of activity. This type of network therefore has a placeholder function: neuronal excitation remains centered on one spot, which could then be moved by external neuronal stimulation that overrides the balancing act along the line (see Fig. 1B in the preprint).

Bump attractor models are a common explanation for the food-seeking behavior of Cataglyphis fortis, a Saharan ant that forages far from its nest in a desert that is (presumably) devoid of landmarks it can use for orienting itself. Ants of this species travel hundreds of meters away from the nest in a non-linear search, yet they manage to head back home in a straight line. To do so, it is argued that they use some kind of path integration, or “dead-reckoning,” where the ants always keep in mind an invisible line that leads to their nest. Creating such a mental tether requires the ants to maintain a log of both the distances and directions they travel. That is where the bump attractor comes in—it is a theoretical neural network model that could explain how the ants actually track distance travelled. Intriguingly, there are even candidate structures for such a network in the central complex of the ant brain.

In this preprint, Pisokas and Hennig investigate the plausibility of the bump attractor model as an explanation for insect navigation. Despite using their own models coupled with data from two separate studies, they find little support for this neat hypothesis, demonstrating the need for a new hypothesis to explain the neuronal basis for path integration in insects.

Key findings:

  1. If ants were using a bump attractor neural network, stochastic drift would lead to the neuronal activity bump shifting at a constant rate. Ants were experimentally prevented from returning to their nest for different time periods in two separate empirical studies, and the authors’ analyses show that indeed their homing distance accuracy does deteriorate at a constant rate. This finding is consistent with the bump-attractor hypothesis.
  2. However, the required number of neurons needed for such a network far exceed what is reasonable for an ant. The authors devise models that try to fit the empirically observed rate of drift to an approximate number of neurons, and the resulting quantities of neurons are simply too high to be biologically reasonable, e.g., 47,000 neurons (for context Drosophila melanogaster has around 200,000 neurons in its entire brain).
  3. Smaller and more plausible networks (on the order of hundreds of neurons) might be possible if the neurons can sustain their action potentials for longer periods of time; however, the possibility of this has only been demonstrated in mammalian neurons in vitro. It is not clear if this is a biologically relevant possibility for ants in vivo.
  4. The observed decay rate of homing distance accuracy in ants is biased toward shorter distances, and this bias becomes more pronounced over time. The bump attractor neural network predicts that they should be equally likely to incorrectly make longer trips as they would incorrectly make shorter trips. Trying to correct for this by systematically biasing the neural network toward shorter trips creates a constant distance decay rate, but this does not match with the accelerating decay rate in ants.
  5. In short, modeled ant behavior based on the bump attractor hypothesis is different from observed ant behavior.

Questions for the authors:

  1. Your estimate of the required number of neuronal circuits depends upon “reasonable” assumptions about the underlying biology and wiring of an ant’s brain. How big a range of possibilities deviates from the reasonable assumptions you make? Are there any conceivable (albeit perhaps unlikely) sets of assumptions that would lead to a much smaller set of necessary neurons other than the longer action potentials that you discuss?
  2. Could you speculate on a conceivable alternative model to the bump attractor?
  3. Is it possible that the desert ants are using landmarks that we simply do not notice because they are small or otherwise difficult for us to detect? Could some combination of landmarks plus a simple bump attractor-based path integration be feasible?
  4. How can a constant decrease in homing distance accuracy create an accelerating decrease in homing distance?

Why I think the work is important:

Models can help guide empirical studies, yet implausible models could potentially lead to empirical studies that produce less useful data than if it had been gathered with a more likely model in mind. By examining the flaws in an attractive but ultimately unlikely model, the authors save empiricists time and give theoreticians additional impetus to develop a better hypothesis. Falsification of existing hypotheses clears the ground for better ideas.

The problem of how ants (and insects more broadly) can actually handle path integration on a neuronal level is an interesting one, and, since their central nervous systems are smaller, it may be a more tractable problem than similar ones in the larger nervous systems of vertebrates. Furthermore, lessons learned in invertebrates could provide transferrable insights into how neural networks manage to make such calculations.

Tags: ant, insect, memory, neural circuits, neurobiology, neuroscience, path integration

Posted on: 11 May 2022 , updated on: 20 June 2022


Read preprint (2 votes)

Author's response

Ioannis Pisokas shared

Thank you Tim for the thorough and very well written review of our preprint. Here are some answers to your questions.

As you say, the values we assumed for the biophysical and network parameters would certainly affect the exact numbers of neurons and the time constants needed. A sensitivity analysis when different parameters are changed might be a worthwhile addition to the preprint. However, we will still need too many neurons to get the required stability with parameter values in ranges we have seen in animals. If the ‘graded persistent activity’ neurons turn out to exist in insects, and they can sustain prolonged activity in vivo (without unrealistically high concentrations of muscarine), the bump attractor’s stability will be dramatically improved. Then the number of required neurons may fall to plausible sizes. But even then, the other problem remains, the state loss dynamics of attractor networks also differ from the animals (random walk in attractor networks vs monotonic decay in the animals). These are two fundamentally different physical phenomena we are observing here. Too bad for attractor networks!

Regarding your other excellent point, the experimentalists are very careful to rule out the possibility that ants use alternative sources of information. The ants are captured in the neighbourhood of their nest and then transported and released far away in places they have never been. The surrounding view is checked to ensure no remote landmarks are visible from both locations, e.g. mountains. In particular, the Cataglyphis ants mentioned in the preprint live in vast flat areas without prominent landmarks. Experimentalists have studied these ants extensively to test when/if they use visual cues, chemical trails, wind, magnetic fields etc. They are fantastic ants; they rely on super-accurate path integration (direction and distance estimation) when deprived of familiar surroundings.

On your question about the constant decrease in the homing distance accuracy and accelerating decrease in the homing distance: Imagine that the ant walks a shorter homeward distance, but the distance it walks varies more and more between trials as time passes. Like a spear thrower, the more tired he/she gets, the shorter the throw distance and the less accurate the targeting. Of course, this is just a metaphor; I am not claiming that the results are due to the ants getting tired!

Have your say

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Sign up to customise the site to your preferences and to receive alerts

Register here