There is a decade-long tradition in which mathematical and computational models of social learning and cultural evolution demonstrate why social learning evolves; when it is useful; why it can be biased toward learning from the majority or from prestigious, successful, or similar individuals; and how learning biases help create the population-level dynamics of cultural change (Boyd & Richerson Reference Boyd and Richerson1985; Henrich & McElreath Reference Henrich and McElreath2003; McElreath & Henrich Reference McElreath, Henrich, Dunbar and Barrett2007). In turn, an understanding of these population-level dynamics has helped us understand and explain a wide range of phenomena including the origins of human cooperation (Richerson et al. Reference Richerson, Baldini, Bell, Demps, Frost, Hillis, Mathew, Newton, Naar, Newson, Ross, Smaldino, Waring and Zefferman2016), civilization (Bowles & Gintis Reference Bowles and Gintis2013), social identity (Smaldino Reference Smaldino2019), hipsters (Smaldino & Epstein Reference Smaldino and Epstein2015), warfare (Zefferman & Mathew Reference Zefferman and Mathew2015), sex-biased tool use in dolphins (Zefferman Reference Zefferman2016), and environmental sustainability (Waring et al. Reference Waring, Goff and Smaldino2017). This theoretical framework already answers a number of questions Veissière et al. care about.
A criticism of this body of literature is that cultural evolutionary theory does not adequately consider cognition (Heyes Reference Heyes2018b). There are indeed questions about cultural evolution and social learning that a better understanding of cognition might help answer. In particular, cultural evolution likely shapes the sociocultural environments in which the cognitive machinery that facilitates social transmission develops. Learning more about the dynamical interaction between cognitive development and cultural evolution is an important and, until very recently, underappreciated research area (Contreras Kallens et al. Reference Kallens PA, Dale and Smaldino2018; Smaldino & Spivey Reference Smaldino and Spivey2019).
We find much to agree with in Veissière et al.'s qualitative description of how cognition and culture interact. However, much of this is well-worn territory. The main advance proposed in the target article is that free energy principle (FEP) and thinking through other minds (TTOM) integrate an impressively large number of theories and hypotheses of cognition, social learning, and cultural evolution. However, given the authors’ extremely underspecified mathematical model, we remain unconvinced by this claim.
The mathematical model presented in the target article is too imprecise to be useful in its current form. The authors do not describe what the terms of the model represent in a cultural or cognitive system or show how the model does any scientific work. Indeed, the FEP is not actually a model of anything, but rather a paradigm for describing the behavior of systems that must themselves be modeled. It is fine to start theorizing with a general model (see, e.g., Frank [Reference Frank1995] on the Price equation), but useful models must eventually map onto relevant aspects of the world. In the target article's entire mathematical appendix, no parameter is described as representing any aspects of culture, brains, agents, ideas, cognitive processes, or any tangible or intangible object related to the purported subject matter. Without this basic modeler's due diligence, we find the discussion of the potential contributions of FEP and TTOM to contribute little.
Take, for example, the “dark room” problem posed in the target article: “if agents aim to avoid unexpected encounters with their environment, we should expect minimally changing sensory environments like dark rooms and correspondingly monotonous sensations to be the most frequently (re)visited states of an organism.” Of course, even a moment's consideration resolves this quandary, as strategic action can often dominate over passive non-action, and natural selection will favor it when it does. What does the FEP add to this non-dilemma? The authors state that:
the FEP deals with the issue of novelty seeking behavior by formalizing action as being in the game of maximizing the epistemic value of action (or epistemic affordance) … [F]ree energy minimizing agents seek to sample the world in the most efficient way possible. Since the information gain (i.e., salience) is the amount of uncertainty resolved, it makes good sense for the agent to selectively sample regions of environment with high uncertainty, which will yield the most informative observations… In effect, agents will act to optimize the epistemic value or affordance of an action before acting on its pragmatic value, which is essentially its expected utility.
This is qualitatively appealing, but adds little value to the current state of scientific understanding. There is already a long history of modeling the trade-off between exploring an uncertain environment and “acting on its pragmatic value” in the social and biological sciences (e.g., Hills et al. Reference Hills, Todd, Lazer, Redish and Couzin2015; Rendell et al. Reference Rendell, Boyd, Cownden, Enquist, Eriksson, Feldman, Fogarty, Ghirlanda, Lillicrap and Laland2010; Reference Rendell, Fogarty, Hoppitt, Morgan, Webster and Laland2011). To demonstrate how a model can add value, contrast the musings above with a mathematical model studied by Perreault et al. (Reference Perreault, Moya and Boyd2012). In their model, a population of agents uses Bayesian learning to integrate environmental and social cues and eventually make decisions in an uncertain environment. They show that individuals optimally weigh social cues (relative to environmental cues) more heavily when the environment is more stable and when environmental cues are more uninformative. They also show that conformist-biased social learning (weighting common behaviors above chance) readily outperforms unbiased social learning across a broad range of conditions, especially when environments are novel or cultural transmission is error-prone. Most importantly, they provide an explicit model where agents make decisions and perform actions with measurable outcomes that potentially provide insight into the world. They have gone from the “first principles” of a Bayesian-learning process and evolutionary selection to the consequent changes in individual agents’ cognition in response to socio-environmental forces.
What added value does the FEP give above and beyond the type of modeling done by Perreault et al.? Would an agent using the FEP or TTOM outperform their Bayesian-learning agents? Or is the Bayesian-learning agent a special case of an agent using FEP or TTOM? Should the overall model dynamics be analyzed using principles of the FEP, and if so, why?
Veissière et al. assert that agents will maximize the epistemic value of an action before its pragmatic value. However, a long history of models in many disciplines suggest fundamental trade-offs between these goals. As researchers interested in marrying cognition to cultural transmission, the determination of whether the FEP or TTOM are useful cannot be assessed unless their proponents can show how they do useful work in explaining the world.
There is a decade-long tradition in which mathematical and computational models of social learning and cultural evolution demonstrate why social learning evolves; when it is useful; why it can be biased toward learning from the majority or from prestigious, successful, or similar individuals; and how learning biases help create the population-level dynamics of cultural change (Boyd & Richerson Reference Boyd and Richerson1985; Henrich & McElreath Reference Henrich and McElreath2003; McElreath & Henrich Reference McElreath, Henrich, Dunbar and Barrett2007). In turn, an understanding of these population-level dynamics has helped us understand and explain a wide range of phenomena including the origins of human cooperation (Richerson et al. Reference Richerson, Baldini, Bell, Demps, Frost, Hillis, Mathew, Newton, Naar, Newson, Ross, Smaldino, Waring and Zefferman2016), civilization (Bowles & Gintis Reference Bowles and Gintis2013), social identity (Smaldino Reference Smaldino2019), hipsters (Smaldino & Epstein Reference Smaldino and Epstein2015), warfare (Zefferman & Mathew Reference Zefferman and Mathew2015), sex-biased tool use in dolphins (Zefferman Reference Zefferman2016), and environmental sustainability (Waring et al. Reference Waring, Goff and Smaldino2017). This theoretical framework already answers a number of questions Veissière et al. care about.
A criticism of this body of literature is that cultural evolutionary theory does not adequately consider cognition (Heyes Reference Heyes2018b). There are indeed questions about cultural evolution and social learning that a better understanding of cognition might help answer. In particular, cultural evolution likely shapes the sociocultural environments in which the cognitive machinery that facilitates social transmission develops. Learning more about the dynamical interaction between cognitive development and cultural evolution is an important and, until very recently, underappreciated research area (Contreras Kallens et al. Reference Kallens PA, Dale and Smaldino2018; Smaldino & Spivey Reference Smaldino and Spivey2019).
We find much to agree with in Veissière et al.'s qualitative description of how cognition and culture interact. However, much of this is well-worn territory. The main advance proposed in the target article is that free energy principle (FEP) and thinking through other minds (TTOM) integrate an impressively large number of theories and hypotheses of cognition, social learning, and cultural evolution. However, given the authors’ extremely underspecified mathematical model, we remain unconvinced by this claim.
The mathematical model presented in the target article is too imprecise to be useful in its current form. The authors do not describe what the terms of the model represent in a cultural or cognitive system or show how the model does any scientific work. Indeed, the FEP is not actually a model of anything, but rather a paradigm for describing the behavior of systems that must themselves be modeled. It is fine to start theorizing with a general model (see, e.g., Frank [Reference Frank1995] on the Price equation), but useful models must eventually map onto relevant aspects of the world. In the target article's entire mathematical appendix, no parameter is described as representing any aspects of culture, brains, agents, ideas, cognitive processes, or any tangible or intangible object related to the purported subject matter. Without this basic modeler's due diligence, we find the discussion of the potential contributions of FEP and TTOM to contribute little.
Take, for example, the “dark room” problem posed in the target article: “if agents aim to avoid unexpected encounters with their environment, we should expect minimally changing sensory environments like dark rooms and correspondingly monotonous sensations to be the most frequently (re)visited states of an organism.” Of course, even a moment's consideration resolves this quandary, as strategic action can often dominate over passive non-action, and natural selection will favor it when it does. What does the FEP add to this non-dilemma? The authors state that:
the FEP deals with the issue of novelty seeking behavior by formalizing action as being in the game of maximizing the epistemic value of action (or epistemic affordance) … [F]ree energy minimizing agents seek to sample the world in the most efficient way possible. Since the information gain (i.e., salience) is the amount of uncertainty resolved, it makes good sense for the agent to selectively sample regions of environment with high uncertainty, which will yield the most informative observations… In effect, agents will act to optimize the epistemic value or affordance of an action before acting on its pragmatic value, which is essentially its expected utility.
This is qualitatively appealing, but adds little value to the current state of scientific understanding. There is already a long history of modeling the trade-off between exploring an uncertain environment and “acting on its pragmatic value” in the social and biological sciences (e.g., Hills et al. Reference Hills, Todd, Lazer, Redish and Couzin2015; Rendell et al. Reference Rendell, Boyd, Cownden, Enquist, Eriksson, Feldman, Fogarty, Ghirlanda, Lillicrap and Laland2010; Reference Rendell, Fogarty, Hoppitt, Morgan, Webster and Laland2011). To demonstrate how a model can add value, contrast the musings above with a mathematical model studied by Perreault et al. (Reference Perreault, Moya and Boyd2012). In their model, a population of agents uses Bayesian learning to integrate environmental and social cues and eventually make decisions in an uncertain environment. They show that individuals optimally weigh social cues (relative to environmental cues) more heavily when the environment is more stable and when environmental cues are more uninformative. They also show that conformist-biased social learning (weighting common behaviors above chance) readily outperforms unbiased social learning across a broad range of conditions, especially when environments are novel or cultural transmission is error-prone. Most importantly, they provide an explicit model where agents make decisions and perform actions with measurable outcomes that potentially provide insight into the world. They have gone from the “first principles” of a Bayesian-learning process and evolutionary selection to the consequent changes in individual agents’ cognition in response to socio-environmental forces.
What added value does the FEP give above and beyond the type of modeling done by Perreault et al.? Would an agent using the FEP or TTOM outperform their Bayesian-learning agents? Or is the Bayesian-learning agent a special case of an agent using FEP or TTOM? Should the overall model dynamics be analyzed using principles of the FEP, and if so, why?
Veissière et al. assert that agents will maximize the epistemic value of an action before its pragmatic value. However, a long history of models in many disciplines suggest fundamental trade-offs between these goals. As researchers interested in marrying cognition to cultural transmission, the determination of whether the FEP or TTOM are useful cannot be assessed unless their proponents can show how they do useful work in explaining the world.