At Rise of AI 2019 in Berlin, Tim Landgraf, professor of artificial and collective intelligence at Freie Universitaet Berlin, argues that building better AI requires first understanding biological intelligence, summed up as 'B comes before A.' He studies honeybees as a model system because their entire life cycle, social interactions and individual navigation can be tracked in ways not possible with humans or larger animals. He describes recording brain activity from bees flying over open fields with electrodes and a miniaturized lab mounted on a copter, and explains how a labeling bottleneck in tracking bee markers led his student Leon to develop RenderGAN, a generative adversarial network that learns to generate realistic training images with controllable labels, later applied to faces. Landgraf closes with an example of how bees share food through a 'social stomach' and proposes a similar energy-sharing model for self-driving electric cars, demonstrated with model cars and induction coils. He ends with calls for more public research funding and closer attention to solutions already found in nature.
Key takeaways
- Landgraf's core thesis is that building better AI requires first understanding biological intelligence: 'B comes before A.'
- His lab records brain activity from bees fitted with electrodes and flown over real fields using a lab miniaturized onto a copter, funded by the Klaus Tschira Stiftung.
- A six-person-month manual labeling bottleneck for bee markers led to RenderGAN, a GAN that generates realistic images with controllable labels, later applied to faces.
- Bees feed hungry foragers through a 'social stomach' that regurgitates nectar, an adaptation Landgraf estimates saves about a minute per forager.
- Landgraf proposes routing self-driving electric cars to exchange energy via induction mid-route, inspired by the bees' social stomach, and demonstrates it with model cars.
Detailed notes
From bees to better AI
Landgraf introduces his lab's approach of studying biological model systems such as fish and honeybees, combining robotics and machine learning to reach research questions otherwise intractable. His central thesis is that in order to build good and better AI, researchers need to understand biological intelligence first, which he shortens to 'B comes before A.' Bees, he says, live only six to seven weeks in summer yet forage over several kilometers and reliably find their way back, despite a brain smaller than one cubic millimeter, which he presents as remarkable in terms of energy and space efficiency compared with human brains and artificial systems.
Tracking a colony's social network
Beyond individual navigation, bees communicate collectively, for example through the bee dance that conveys direction and distance to a food source. Landgraf's lab builds interaction matrices recording who interacts with whom and for how long, generating millions of such matrices over a bee's full life, which the team compresses with machine learning into what he calls the 'circle of life' of bees. Feeding sites where bees have good or bad experiences shape their future behavior, comparable to reward and punishment in humans. In a separate project, funded by the Klaus Tschira Stiftung despite skeptical reviewers, his team records activity from specific neurons by poking electrodes into bees' brains and flying them over real fields using a miniaturized lab mounted on a copter, producing usable spike recordings from individual neurons.
RenderGAN: solving a labeling bottleneck
To track bees, the lab glues numbered markers on their backs that a neural network must locate and decode into three rotation angles and a 12-bit ID. Manually labeling enough images for this task was estimated to take more than six person-months of work and would have to be redone whenever the tag design changed. Instead, Landgraf's student Leon developed RenderGAN, a generative adversarial network that learns realistic image augmentations such as blur and reflections directly from unlabeled data, producing synthetic training images with controllable, known labels. The method was later ported to generating faces with similarly controllable attributes.
What bees teach about EV charging
Landgraf closes with the bees' 'social stomach': a forager bee that finds no food asks a nestmate, who regurgitates nectar to feed her so she can continue working, an adaptation he estimates saves roughly a minute per forager. He proposes an analogous idea for self-driving electric cars: a system that knows a car's planned route and battery level in advance and pairs it, mid-route, with another car that has spare energy to transfer via induction, avoiding the need to stop and charge. He demonstrates the concept with model cars fitted with induction coils bought from a Chinese online retailer, and says simulations suggest it is workable at real-world scale.
Q&A highlights
Asked whether the social stomach evolved specifically to prevent overeating rather than just to save time, Landgraf says he cannot know the original evolutionary reason, only that the behavior is present and can be interpreted after the fact. Asked about chemical trail communication as in ants, he explains that bees navigate mainly through vision rather than pheromone trails, which do not work over several kilometers or in the air. He adds that bees are a valuable model system precisely because their whole colony's social network can be tracked and manipulated in ways not permitted or practical with humans or even larger animal studies, citing disease-spread research on livestock as an example limited to small numbers of animals.
Quotes
In order to build good and better AI, we need to understand the BI, the biological intelligence.
Tim Landgraf
Let's learn a little bit from the bees.
Tim Landgraf
The risk is in the public sector, and that's good for companies spinning out of those labs.
Tim Landgraf
Questions answered in this talk
Why does Tim Landgraf study bees instead of humans for collective intelligence research?
According to Tim Landgraf at Rise of AI 2019, bees can be tracked over their entire six-to-seven-week life, allow interventions like electrode recordings that are impossible or unethical in humans, and their social network shares structural properties with human networks.
What is RenderGAN?
According to Tim Landgraf at Rise of AI 2019, RenderGAN is a generative adversarial network, developed by his student Leon, that learns realistic image augmentations from unlabeled data so it can generate synthetic training images with controllable, known labels, avoiding months of manual labeling.
How could bee behavior inspire electric car charging?
According to Tim Landgraf at Rise of AI 2019, bees share food via a 'social stomach,' and a similar system could let self-driving electric cars exchange energy via induction while en route, rather than stopping to charge; he demonstrated the idea with model cars.