Insights from Electric Fish: Lessons for NeuroAI Research

Electric fish are among the most captivating examples within nature’s treasury of wonders. They “perceive” their surroundings and one another by detecting — and creating their own — electric fields. This remarkable skill presents a significant opportunity for investigation in the burgeoning field of NeuroAI, which delves into the perceptual and cognitive abilities of both biological and artificial systems.

For Kanaka Rajan, an investigator at the Kempner Institute and associate professor in the Department of Neurobiology at Harvard Medical School, electric fish symbolize a promising resource for broader understanding of “collective intelligence,” which arises from interactions among various aim-driven entities. 

Rajan and her team propose that investigating weakly electric fish might serve as a gateway to comprehending multi-agent intelligence, potentially illuminating the intricate dynamics characteristic of animal and human communities.

Rajan and her collaborators focused on the Gnathonemus petersii, commonly known as Peter’s elephantnose fish, a weakly electric species renowned for the trunk-like extension on its head. Indigenous to the rivers of western and central Africa, elephantnose fish thrive in murky waters and shaded, slow-flowing streams. By utilizing electro-sensory abilities, these fish can navigate through muddy terrains by generating pulses referred to as electric organ discharges, or EODs. 

“Their entire lives revolve around electric pulses,” explains Rajan. “Their communication, mating rituals, pursuits, aggressive behaviors, cooperation, competition — everything is dictated by these pulses they produce.”

As Rajan and her team investigate how agents — which may include humans, animals, robots, or language models — interact with each other across varying scales, elephantnose fish present a valuable opportunity to study coordinated communication, or agent interaction, within a relatively straightforward framework. 

This is where weakly electric fish play a crucial role. Their “language” comprises sequences of matching electrical pulses. The “sentences” differ based on the frequency of these “syllables” emitted by each fish. Even though their modes of communication are less complex than those of humans, they exhibit emergent, coordinated actions that could offer insights into much deeper social dynamics. Modeling these behaviors could also aid in the creation of advanced AI systems.

The central concept when examining collective intelligence is that social dynamics do not merely represent the aggregate of individual behaviors, or even pair-to-pair interactions. For instance, the unique behaviors that can develop in groups of humans at a social gathering are significantly more intricate than a mere collection of two-person dialogues. This complexity may arise because each conversation retains memories from prior discussions, even when they involved different individuals, or due to the subtleties inherent in triadic dialogues that cannot be entirely captured by the total of every possible two-person chat. Moreover, the overall environment plays a vital role concerning human behaviors — whether it is a work gathering or a casual after-hours event, for instance, or whether the scenario unfolds among colleagues or within a hierarchical structure.

Elephantnose fish exhibit this kind of collective intelligence in remarkable ways. In a study conducted by Federico Pedraja and Nathaniel Sawtell, a key collaborator of Rajan at Columbia University, elephantnose fish within a group displayed the ability to leverage each other’s food-finding skills. If one fish discovers a viable food source, it sends out pulses that can be sensed by nearby fish. Consequently, other fish can conserve energy by following the lead fish instead of foraging independently. Through this form and other variations of collective communication, the fish are capable of demonstrating social behaviors that enhance their chances of survival and success in specific surroundings.

Rajan and her team developed computer simulations using artificial agents that replicate the behaviors of elephantnose fish. By altering elements that would be impractical to control in actual fish experiments, Rajan and her team successfully modeled and examined how collective intelligence arises across different situations. 

Their research at the Kempner Institute has begun to uncover some captivating findings. Evolutionary simulations, where artificial fish must endure to transmit their traits to succeeding generations, indicate that food availability influences whether the fish primarily engage in cooperation or competition. Diminishing the reliability of food sources raises the probability that the population will display competitive behaviors. Cooperation and competition are not innate traits of these artificial agents. Instead, these behaviors develop gradually over the course of successive generations through a simulated model of natural selection.

Rajan intends to broaden this initiative to gradually include increasing levels of complexity, including a deeper comprehension of the learning processes in individual agents and the emergence of diverse survival strategies.

A significant question that the team is beginning to tackle is whether there exist universal principles governing social interactions. Is there, for instance, a “critical mass” of agents needed for optimal cooperation? Or a threshold beyond which competition becomes excessive? 

This type of research has fascinating ramifications for human collaboration but is also pertinent to applied AI inquiries. Principles of collective intelligence could play a crucial role in how distinct AI systems interact and communicate amongst themselves. Cooperative teams of AI agents — often referred to as “swarms” — hold promise for technological applications, potentially facilitating quicker, more complex, and adaptable problem-solving capabilities. Only time will reveal whether such AI “hive minds” prove beneficial, but in the interim, foundational work like that of Rajan and her collaborators is illuminating how both biological and artificial agents can collaborate to generate collective intelligence.