Collective Intelligence: Foundations and Radical Ideas
Symposium at Santa Fe Institute, June 20-22 (2023)
Collective intelligence studies the production of adaptive behaviors. Including anticapting problems the system might face. Is the output of the system intelligent or is the system itself intelligence? This is a core question. Org behavior and management is a key application I’m interested in. This is umbrellad by collective phenemona. How do patterns arise in space and time? Statistical mechanics and swarm research are often disparate branches that our conversations here will explore interactively.
Guy Theralaz Building Collective Intelligence Throuch Social Interactions
What are the conditions required for the emergence of collective intelligence? (the book Intelligence Collective by Eric and Guy would be great to review)
- a collective of individuals where it is possible to interact with each other (adapting behvior to face various situations)
- collective behavior is intelligent if it is adaptive and solves generic tasks
- the higher coordination in the group, the better at solving novel problems
In looking at collective intelligence in social insects like bees and ants we study:
- how to decipher interactions between individuals
- how to identify the information that is exchanged
- how to indivudials integrate these multiple interactions
- effect of these interactions
Guy shared Pierre-Paul Grasse’s research on stigmergy. How ongoing activity is guided by the by-product of its work. By this we mean an insect’s actions are determined or influences by the consequence of another inseect’s previous action. Indirect communciation that makes possible coordination and regulation of collective work, and it seems to be following a pre-defined plan. Identifying all the indirect interactions is reuquired to understand the growth and shape of the social insects nests.
He then shared a cool study where they gave wasps colored paper as pulp fodder so that they could look at individual decisions of the wasps over time in their construction of their nest. They were more likely to build a cell where three walls were already built. This shows that the behvior was really influenced by the structures that were already present. This reminds me of research I’ve been wanting to do looking at how developers, particularly on open source projects, choose to work on different repos. What are the coding economies in these projects? Can we think of an open source project as a nest and the contributors the social insects? How might such analysis promote mindful stiumuls into the system to prompt positive construction in these projects?
Some collective intelligence objectives include:
- cooperative transport
- task allocation and division of labor
- swarm thermoregularion
- collective defense
Collective motion in fish schools include, governened by aligment intensity and attractive intensity:
Distance, angular position, and realative heading can be analyzed in the interactions between fish, including analysis of attraction and alignment between two fish.
Guy is supported by Collective Animal Behavior lab (Max Planck Institute), CRCA, CNRS, SFI (sponsors)
Chris Kempes First Principles Thinking in Biology
There are two types of biological first principles:
- universal (invariant)
- contingent (emerge from biology)
“Evolution is life finding the laws of physics”. What a great way of describing the “random” walk that organisms take as they evolve. Chris shared how the fundamental thermodynamic limits include the relation of DNA, ribosome, and protein. Life is only 4 times less efficient than the ultimate physical bound.
Mutation rate is selection coefficeint over length of information. This was explored in DNA copying and allowable mutation rate, and is proposed as a univeral limit, not contingent limit.
It is proposed that there are three levels of life (Kempes and Krauauer):
- material (contingent)
- fixed universal constraints (physical)
- optimization principles (logical)
Geoffrey West From Individual to Collectives in Biosocial Systems: A Physics Perspective
Geoffrey gave us a quick tour of collective intelligence in the history of physics, through moments like the Higgs Boson discovery and the creation of the World Wide Web, ultimatley landing on the foundations of Social physics (Comte, Quetelet, Durkheim). Founders of the science of sociology. Why did this go out of fashion? What is trying to be accomplished? What is being optimized?
Cities are collective intelligent entities creating knoweldge, innovation, ideas, standards and qualities of life. Took a nice path through comparing the relative collective intelligence of different cities. In superlinear scaling, the bigger the city you are, the more interations per capita, number of patents, more total wages, and supercreative employment, but also crime (which also requires a certain kind of intelligence).
Offered the correlation of body mass and metabolic rate as an example of the fitness function that species are optimizing for. Proposes the concept of social metabolic rate. How much energy is needed to keep society going while also factoring the energy needed for collective intelligence.
Nikta Fakhri First Principles of Collective Intelligence in Active Matter
Being alive is a macroscopic state, with things like proteins and DNA serving as the microscopic processes that shape it.
An order parameter is a quantification of the emergent phenomena. So the order parameter space definese the solution space of where that emergence takes place. Active experimental tools for studying nonequilibirium degrees of freedom and frameworks leveraged for prediction. What are the broken symmetries? These include time reveral symetry breaking (fluctuations), spatiotemporal symmetry breaking (protein patterns), chiral symmetry breaking (nonreciprocoal matter and odd elasticity).
Kullback-Leibler Divergence (KLD) measure of the difference between two entropies, which be leveraged to articulate the arrow of time.
Maxim Raginsky Information, Compression, Generativity in Neural Nets: First Principles
What is the relationship between inner simplcity and external complexity? “Human beings, viewed as behaving systems, are quite simple. The aparent complexity of our behavior over time is largely a reflection of the complexity of the environments we find ourselves”, Herbert Simon.
The mechanics of compression can be defined by two functions, an encoder function (f) that maps a symbolically represented object to a compressed represenation, and then a decoder function. There is a difference between lossless and lossy compression.
Natural languages must follow these constraints (Zellig Harris, Chomsky’s advisor):
- partial order
We can hort-circuit evolution with symbols using our LMMs. Replicating evolution more efficiently, but when we do this we’re training these systems on the wrong set of signs. GPT does not use langauge for prediction or control of reality, yet this is how humans can leverage the technology of language.
Jessica Flack First Principles Thinking for Information Processing Systems
Foundations of computation are to be discovered not invented. Can we define a formal language for micro-macro maps? This requires holding the macro constraints, so we can hold in check the subjectivity inherent in observation. Physics is dominated by concepts like pressure, temperature and entorpy, whereas biology makes use of comparable collective concepts, metabolism and conflict management and robustness. Adpative systems produce order by addition of information processing.
Different scaling depends on whether energy or information dominate. Energy dominated variables scale sublinearly, whereas information dominated macrscopic variables scale superlinearly. Components collectively compute their macroscopic world. Course-graining connects the large and small scale.
How is nature couse-granining, particularly given there is subjectivity and error?
Elements of collective computation include:
- circuit (stocahstic causal description of microsocpic regularities)
- notion of correctiveness (how a program knows to stop)
Social power distribution is a course grained, slowly changing encoding of the collective view an individual can use succesfully. With inductive game theory we can build a circuit that shows how the macro can be derived from the micro. Downward causation (tuning behavior from macrocopic behavior).
Information bottlenecks reduce uncertainty about the future. Policing reduces the cost of fights.
Ian Couzin The Geometry of Decision-Making
Ian led us through a whirlwind tour of relationship between individual and collective cognition.
Placazoa is a non-neuronal animal. Humans can’t study turbulence at the reynolds numbers we care about, so we look at the energy costs of schooling in fish.
All the software they are working on is open source and easy to use. Excitd to check it out!
Influence across networks can be non-reciprocal. Computation can be an emergent propery of the structure of networks. Regulating social contagions. Modulating behavior in response to increased risk. Sensig of long-range environmental gradients.
Leadership in collective decision making. Can these groups come to consensus? How do grouping animals make informed unanimous decisions? Associating goal-orineted vecotrs with social vectors in the brain. Making decisions on the move. Below a critical angle of opinion, these creatures can be very accurate finding the average vector (that seems evolutionarily logical). If the groups are not equal, above the critical point (where it bifurcates), they will always choose the majority. Paper ( uniformed individuals promote domacratic consensus in animal groups). Being opionated doesn’t always work, because if you add unfiormed individuals, you can return the control to the majority.
Collective minds. Discussion topic: how do neural collectives (individual brians) make decisions as they are moving through space?
Really cool research showing real fish interacting with projected fish in the “matrix”. They can do full parametric scans of agents in virtual reality to see their decision making.
Reinforcement of nerual ensembles that have similiar directional representations (goal vectors). Neural ring attractor network. Instead they used a flock spin-system mapping onto a ring-attractor spin system to measure specfic firing rate of individuals. Principles of decision-making - the actual geometry of the problem makes a huge difference. Othe principles include perceived difference between options and time. If the brain can go through the geometrical bifurcation it is easier to make a decision (this is a study they recently did). The brain breaks down the world into a series of bifurcations (a series of binary decision) and uses that structure to make decision. The interaction of the geometric representation and then the decision that leads to movement through spacetime makes this especially complex. Representations in eliptical geometry can prevent the brain from getting stuck in traps.
In flocks, requires a for mof explicity ihnibition (forgetting) can maintain performance when more than two decisions are presented to the group. Konstanz - Center for the advanced study of collective behavior (center fo thte advanced study fo collective behavior) - cool place to work! Would love to visit. Would be good to imagine what studies we’d like to do in such a lab.
Twitter communication network is not how we evolved to communicate. Truth can’t catch up in natural communication networks can’t catch up with the spread of collective stupidity in these newere communiction networks.
Melanie Mitchell The Meaning of Intelligence in AI
How do we think about inteligence in systems that are not biological? Do AIs understand the data they process? Does it even matter? LMMs are superhuman in their ability to extract information from the world’s database of text (Terrance Sejnowksi)
So how do we evaluate LMM?s
We can give them tasks that actually require understanding. For instance, we can assess their physical understanding, or a task requiring strategic social manipulation
Do these systems create latent concepts? Can they produce abilities that are fundamentally different and enable newunderstanding that humans are incable of accessing? This seems like a good line of inquiry.
A viable concept would involve causal structure and enable predictions, reasoning, and common sense. And thse concepts can be accessible to associates through new metaphor to other concepts.
David Krakauer The Meaning of Intelligence in Brains + Collectives
I really enjoyed how David and Melanie both are being clear and brave in their scholarship in a moment when people who are rather new to the machine intelligene story are opining and grandstanding.
“Intelligence is not like mass, intelligene is like architecture. You’d be an idiot to compare an adobe house with a skyculture.” Boom! This reminds me of what my undergraduate advisor Ellen Winner taught me about her work at Harvard Project Zero. When you center multiple intelligence, it becomes difficult to take the intelligence “arms race” as a quest to accumulate a bunch of it and perhaps enough to hit this magical AGI tipping point. I find this sad at a time in history where we have barely scratched the surface and potential of situated cognition, emergence of dynamic knowledge generation in high-trust collaborative teams, and the many ways of knowing that never made their way into our academic text books but shaped the course of human history all the same (for better and worse).
A definition of architecure offered by Frank Lloyd Wright is that architecture helps people understand how to make life more beautiful, the world a better one for living in, and to give reason, rhyme and meaning to life. Another boom! What a gorgeous defintion of architecture, and what a great way to turn up the music on other aspects of the human experience we want in the room with us whenever we are discussing AI. To not consider beauty, making the world a better one for living in, and giving reason, rhyme and meaning to life when discussing technology designed for human use would be unethical.
So, let’s make sure everybody’s in the building as we discuss AI, and remember that collective intelligence includes communication, coordination, and consensus. We haven’t arrived at the metahpor of AI, so we are in a state of radical uncertainty That’s what we’re reaching for together through conersation. The analogy of LMMs as intelligence is wrong, and if anything they are an intelligence cheat, as they have access to the entire library.
David ended with a well state reminder that Inventions emerge from cognitive limitations. Constraints lead to our creativity, so the only way we get to be clever is by being limited, and I believe that limitation is the open gaze of all of our collective ideation about what the AI metaphor should be. This means we can not grandstand and say we’ve already determined that AGI is even possible or the goal. We get to embrace the inherent constraints of collaboration across diverse perspectives, and I think that’s exciting because it means we are priming a canvas for true discovery.
David Ha Collective Intelligence and Open Source AI
Cesar Hidalgo The Nature of Collective Intelligence in Economic Settings
Orit Peleg The Dynamics of Collective Intelligence
Communication infrastructure in animal systems are beutifully complex.
Orit shared communication research from both fireflies and honeybees, orienting us first with three motivating questions:
- How should organisms choose an optimmal signal modality?
- How should they integrate the sginals spotiotemporally?
- How should they respond to signals?
The approach her lab is taking to answer these questions involves both computational and mathematical modeling as well as behavioral experiments.
Every firefly species has its own flash pattern that is used in mating. We can almost think of this binary on-off switching over time as an encoding like morse code. She then shared findings on chimera states for coupled oscillators, which I found very compelling but for the life of me this morning I can’t remember why (lol). Something to follow up on.
Now when it comes to bees, drones will amplify the queen phermone out through the network. Information propagating through a network. The bee’s life is a like a magic well (Karl von Frisch) bettering the stars (Richard Newman).
Mirta Galesic Beyond Collective Intelligence
This talk explores How Collectives Adapt.
Collective Adaptive Framework building bloks include:
- social cognitions (belief networks)
- social networks
- problem landscapes
What trajectories can collectives take while navigating multipe and ever-changing problems? (how and why did we get here? where are we going?)
Some key areas of research we can explore include:
- Social Learning (rules for learning from one another)
- Collective Problem Solving (how groups engage in explore/expolit in solving specific problems)
- Wisdom of Crowds
- Group decision making
- Belief dynamics
- Cultural evolution (how people and collectives learn over generations)
- Game theory
- Group minds
When we think about how this comes together we can think of the interacting topologies of these knowledge domains like a multidimensional Ruth Asawa-ish lava lamp, as opposed to lego-like building blocks. It matters what mental toys we toy with.
With social cognition we are integrating social information (social learning strategies, belief updating, group decision-makign rules)
The integration itself can be grouped in three classes:
- frequency dependent strategies (i.e. majority)
- averaging strategies (blending ineheritance, contagion rules)
- model-based strategies (like averaging rules with extreme weights - i.e. follow the leader, liked, best, etc.)
All the decisions for this advisor are happening in a public Slack channel. Comments can be made their or in the tech spec.
Catherine Hicks (Developer Success Lab) Where we’re going with developer productivity?
Beliefs we seem to be holding include: * Productivity is hard * Productivity is never stopping * Productivity is lonely
What does thriving on engineering teams look like? Access to the socio-cognitive resources you need to sustain innovation.
- learning culture
- motivation and self efficacy
- support and belonging
Developed a model that showd that developing thriving leads to developer producitivity. Empirical evidence backing why thriing matters. Check out their whitepaper if you’d like to leverage this evidence at work!