Marr's Three Levels of Analysis: A New Lens on Social Behavior

In cognitive science, David Marr's three levels of analysis—computational, algorithmic, and implementation—serve as a powerful tool to break down how the brain processes information. However, in their paper Levels of Analysis in Computational Social Science, Peter M. Krafft and Thomas L. Griffiths propose an intriguing expansion of this framework beyond individual cognition and into the realm of social systems (Krafft & Griffiths, 2018 [1]). 

한국어 버전:
2024.10.03 - [Study/논문요약(한글)] - 계산 사회과학: 집단 행동을 이해하는 새로운 렌즈 (Marr의 3단계 분석 방법)up behaviors through computational lenses. 

Illustration of Marr's Three Levels of Analysis (Krakauer et al., 2017[2])

 

In cognitive science, David Marr's three levels of analysis—computational, algorithmic, and implementation—serve as a powerful tool to break down how the brain processes information. However, in their paper Levels of Analysis in Computational Social Science, Peter M. Krafft and Thomas L. Griffiths propose an intriguing expansion of this framework beyond individual cognition and into the realm of social systems (Krafft & Griffiths, 2018 [1]). The result is a fresh perspective that allows us to conceptualize social structures and behaviors as analogous to information processing systems, unlocking new insights into group dynamics, social hierarchies, and collective decision-making.

This blog post will explore the significance of Marr’s framework, its application to social systems, and how it helps clarify complex group behaviors through computational lenses. We will delve into three core examples provided by Krafft and Griffiths —waiting in line, status hierarchies in animals, and rumors— illustrating how Marr's levels illuminate the mechanisms underlying these phenomena.

Marr’s Three Levels of Analysis

David Marr originally designed his framework to understand cognitive processes by breaking them down into three distinct levels:

1. Computational Level: This level defines the "what" of the system. What problem is the system solving?
2. Algorithmic Level: This level describes the "how." What procedure or set of rules does the system use to solve the problem?
3. Implementation Level: This level addresses the "with what." How is the algorithm physically realized within the system?

The primary utility of this framework lies in its ability to disentangle these questions, allowing for clearer understanding and investigation of how systems—whether biological or social—function. Krafft and Griffiths argue that this framework can be successfully adapted to explain social systems, offering a new way to analyze how groups of people, animals, or organizations process information and solve problems. But unlike traditional cognitive systems, social systems often face unique challenges, such as distributed control and the lack of a centralized processor.

Applying Marr’s Framework to Social Systems

In their paper, Krafft and Griffiths explore how Marr’s levels of analysis can be applied to understand collective behavior and decision-making in social groups. For example, in social systems, the **computational level** might focus on a problem such as resource allocation or social coordination. The **algorithmic level** would explore the strategies or rules individuals follow to address this problem. Lastly, the **implementation level** focuses on how these strategies are realized through human interaction, psychology, social norms, or physical environment.

The approach has profound implications for computational social science, allowing us to understand seemingly chaotic social dynamics as structured, problem-solving mechanisms.

Example 1: Waiting in Line

Take the example of waiting in line, a simple yet universal social phenomenon.

- Computational Level: The problem at hand is managing the order in which people are served. The group must process who arrives first and ensure that individuals are served in that order, solving the "first-in, first-out" (FIFO) problem.
- Algorithmic Level: The group solves the FIFO problem by following an informal rule where each individual keeps track of who is ahead of them. This forms a queue or line, with each person holding the position they arrived in.
- Implementation Level: In physical terms, the solution is implemented by standing behind the person in front. If the system works well, everyone gets served in the order they arrived. However, failures can occur when individuals cut in line or the queue collapses into a disorganized crowd.

This example shows how social behaviors can be viewed as computational processes. The line is, in essence, an information-processing system that handles the complex problem of determining and maintaining order.

Example 2: Status Hierarchies in Animals

Krafft and Griffiths also apply Marr’s framework to animal hierarchies, such as those found in many species that establish social ranks.

- Computational Level: The social system faces the challenge of resource allocation with minimal conflict. A hierarchy can reduce competition and help distribute resources such as food, shelter, or mates.
- Algorithmic Level: In this case, the hierarchy is established through pairwise comparisons or contests, such as fights or displays of dominance. These comparisons effectively sort individuals into ranked positions, solving the computational problem of ordering.
- Implementation Level: Animals maintain these rankings through individual memory or social cues. The hierarchy is reinforced by behaviors like submissive posturing or avoidance of higher-ranking individuals.

The hierarchy allows the group to function efficiently by reducing costly fights over resources. Each animal’s status acts as a form of "cached computation, " enabling the group to allocate resources with minimal ongoing conflict.

Example 3: Rumors and Collective Sensemaking

One of the most socially relevant examples in the paper involves rumors and collective sensemaking in times of uncertainty.

- Computational Level: The group is faced with the task of making sense of incomplete or ambiguous information—essentially, a distributed inference problem. 
- Algorithmic Level:  Rumors act as a form of information transmission, where people share hypotheses or interpretations of uncertain events. This algorithm helps the group collectively arrive at a consensus or understanding of the situation.
- Implementation Level: The algorithm is implemented through social communication, such as conversation, social media, or other channels. Individuals pass on information, sometimes distorting or amplifying certain aspects, but the overall function remains to make sense of uncertain environments.

In this case, the spread of rumors can be seen as a functional solution to a shared problem, helping the group reduce uncertainty. However, the process can also lead to failures, such as the spread of misinformation or panic.

Benefits of Applying Marr’s Levels to Social Systems

Krafft and Griffiths argue that applying Marr's levels of analysis to social systems provides several key advantages. First, it offers a deductive approach to discovering mechanisms. Rather than relying on post hoc explanations of social behavior, Marr's levels encourage researchers to think in reverse: start by identifying the problem the group is solving, then trace the algorithms and implementation strategies they use.

This approach is particularly useful in fields like functionalism and analytical sociology, where the goal is to understand how social structures emerge to solve collective problems. Marr’s framework provides the conceptual tools to rigorously test functionalist hypotheses by linking abstract social functions to concrete behavioral mechanisms.

Moreover, the computational lens can aid in the design of social systems. Once a social problem is defined in computational terms, it becomes possible to search for alternative solutions—either by improving the algorithms that people use or by redesigning the physical or social environment in which they interact. For example, instead of relying on physical queues, some institutions use ticketing systems to achieve the same computational function with less human coordination.

Challenges and Critiques

However, the authors also acknowledge several challenges to applying Marr’s levels to social systems. One critique stems from the differences between social systems and artificial systems. In computer systems, for example, communication between agents is precise and reliable. In social systems, communication is fraught with ambiguity, misinterpretation, and cultural differences. This makes it harder to design or even accurately model algorithms for human behavior.

Another criticism comes from the methodological individualism perspective, which argues that group behavior is reducible to the actions of individuals, and thus, focusing on group-level functions may be unnecessary. Krafft and Griffiths respond by pointing out that while individual behaviors are crucial, group-level phenomena—such as social hierarchies or norms—often exhibit emergent properties that cannot be easily reduced to individual actions alone.

Finally, there are limits to the framework’s applicability. Not all group behavior can be meaningfully described as information processing. For example, in cases where collective behavior is chaotic or dysfunctional, Marr’s levels may not offer much explanatory power.

Conclusion

Marr’s three levels of analysis offer a powerful framework for understanding not just individual cognition, but also collective social behavior. By applying this framework to social systems, Krafft and Griffiths demonstrate how seemingly simple social behaviors—like waiting in line or spreading rumors—can be seen as solutions to computational problems. While challenges remain in adapting this framework to the social sciences, its potential to clarify and improve our understanding of group dynamics is undeniable.

As the field of computational social science grows, frameworks like Marr’s levels will likely play a crucial role in guiding research and informing the design of better, more efficient social systems. Whether it’s organizing queues, designing fairer algorithms, or mitigating the spread of misinformation, the computational approach promises to bring new clarity to the complex world of social interactions.

 

 

---

References

[1] Krafft, P. M., & Griffiths, T. (2018). Levels of analysis in computational social science. Cognitive Science. http://dblp.uni-trier.de/db/conf/cogsci/cogsci2018.html#KrafftG18

[2] Krakauer, J. W., Ghazanfar, A. A., Gomez-Marin, A., MacIver, M. A., & Poeppel, D. (2017). Neuroscience needs behavior: Correcting a reductionist bias. Neuron, 93(3), 480–490. https://doi.org/10.1016/j.neuron.2016.12.041