TLDR: This research paper explores how humans flexibly construct internal models of the world, arguing that analogy plays a central role. By formalizing analogies as partial homomorphisms between Markov Decision Processes (MDPs), the authors propose that humans reuse solution-relevant structures from past experiences, effectively amortizing the computational costs of model construction and planning. This process involves building a library of abstract, composable modules that can be adapted and combined to understand and navigate novel situations, offering a powerful mechanism for flexible adaptation that current AI systems often lack.
Humans possess a remarkable ability to navigate and adapt to an incredibly diverse range of situations, from simple daily tasks like vacuuming to complex social interactions or solving intricate math problems. Unlike most artificial intelligence (AI) systems, which often rely on pre-programmed models for specific tasks, humans are constantly constructing and refining their internal understanding of the world on the fly. This flexibility is crucial because the real world is far too complex and varied for a fixed set of internal models.
A recent research paper, titled “Analogy making as amortised model construction,” delves into this fascinating human capability. Authored by David G. Nagy, Tingke Shen, Hanqi Zhou, Charley M. Wu, and Peter Dayan, the paper proposes that analogy plays a central role in how we build these internal models, effectively reducing the computational effort required for both understanding new situations and planning actions within them. You can read the full paper here: Analogy making as amortised model construction.
The core idea is that when faced with a novel situation, instead of building a completely new mental model from scratch, humans draw upon past experiences and existing mental structures. This process is akin to reusing blueprints or modular components. The authors formalize this by viewing analogies as “partial homomorphisms” between Markov Decision Processes (MDPs), which are abstract mathematical frameworks used to model decision-making in environments.
Consider a robot vacuum cleaner. Its internal model of the environment is designed by an engineer, focusing only on relevant aspects like obstacles, dirt, and its charging dock. This model is fixed. Humans, however, must act as both the ‘user’ and the ‘designer’ of their own internal models. This presents a significant challenge: how to create a model that is both accurate enough to be useful and simple enough to be computationally manageable. This is a version of the classic “frame problem” in AI, which asks what information is relevant and what can be safely ignored.
The paper argues that analogy provides a powerful solution to this problem. By recognizing similarities between a new situation and a past one, we can transfer solution-relevant structures and even partial solutions. For instance, a child learning about an email account might be told that the password is like a ‘key’ for their account. This simple analogy allows the child to immediately transfer knowledge from their understanding of physical keys and doors – such as the need to guard it, that only a specific one works, and how it grants access – to the new digital context, even though the physical actions are entirely different.
This transfer of knowledge is not just about understanding; it also amortizes the cost of planning. If a login screen is seen as a ‘locked door,’ the child immediately knows that attempting to click ‘login’ without a password will be futile, narrowing down the range of possible effective actions. The paper suggests that the brain extracts structural regularities from diverse experiences and compiles them into a library of reusable, abstract modules. These modules, like ‘door’ or ‘key,’ become increasingly abstract and broadly applicable over time, forming fundamental conceptual primitives.
Furthermore, humans can compose multiple modules to understand more complex situations. For example, understanding an office door might involve an analogy to an apartment door, while understanding a projector in a conference room might draw on experience with a school projector. The challenge here is ensuring these combined solutions still work effectively in the new configuration.
The authors also discuss how these modules are extracted and refined. As a module is reused across varied contexts, aspects that are consistently irrelevant are discarded, making the module more abstract and versatile. This process is similar to how a concrete metaphor can evolve into a highly abstract concept over time, like the ‘key insight’ in a theory.
Also Read:
- A New Framework for Flexible Self-Correction in Robotic Task Planning with Large Language Models
- Unlocking Advanced Reasoning in AI: A New Approach to Learning from Thought Processes
Ultimately, the paper proposes a conceptual framework for how humans achieve such flexible and efficient model construction. While current AI methods for learning world models still struggle with adapting to novel situations, the insights from human analogy-making, particularly the idea of reusable, abstract modules, could pave the way for more robust and adaptable artificial agents in the future.
