This paper explores the computational logic behind one of our most fundamental cognitive tasks: making sense of a world filled with discrete, persisting objects. The authors propose a Bayesian "open world" model to explain how humans infer the total number of objects in an environment and identify them across multiple encounters. Through three behavioral experiments, they demonstrate that humans don't just count what they see—they use probabilistic reasoning to infer the existence of unobserved objects and learn category properties even when individual identities are uncertain.
The Problem: Identification vs. Object Discovery
Imagine you see a dog in the park, and then another ten minutes later. Is it the same dog or a different one? This is the problem of identification. Now imagine you've seen five different dogs in your neighborhood over a week. How many dogs actually live there in total? This is object discovery (or individuation).
Most psychological models of categorization assume we already know which "tokens" (encounters) belong to which "types" (unique objects). However, in the real world, we face identity uncertainty: we often don't know if two sightings represent the same individual. Furthermore, we must often infer the existence of objects we haven't even seen yet. The authors aim to bridge the gap between low-level perception and high-level categorization by providing a unified statistical framework for these inferences.
A Bayesian "Open World" Model
The authors present a Bayesian generative model referred to as an "open world" model. Unlike standard clustering models that only account for observed data, this model explicitly considers the possibility that the world contains more objects than have been seen so far.
- The Generative Process: The model assumes a world with $n$ total objects. It uses a scale-invariant Jeffreys prior ($P(n) \propto 1/n$) for the population size, which is a standard choice for estimating unknown quantities.
- Sampling: Objects are sampled uniformly at random to create "tokens" (observations).
- Comparison Models: The authors compare their approach to the Dirichlet Process (DP) and Pitman-Yor (PY) mixture models. While popular in machine learning for "infinite" clustering, the authors argue they fail here because they don't naturally represent a finite but unknown population size of a "world."
- Computation: For the small-scale experimental domains, the authors computed exact posterior distributions by enumerating and summing over all possible "worlds."
Experiments and Key Insights
The model was evaluated against three behavioral experiments conducted with human participants, revealing several key insights into human cognition:
1. Humans Infer the Unseen (Experiment 1) Participants observed balls sampled from an urn under UV light, which revealed unique serial numbers. If humans see the same object five times, they infer the total population is small. If they see five different objects, they infer the population is likely much larger than five. This suggests we are constantly "counting" the objects we haven't seen yet.
As shown in the figure above, standard non-parametric models (DP/PY) cannot explain the human "increasing confidence" pattern. These models assume a new object is always equally likely regardless of how many unique objects have already been discovered, failing to match human intuition.
2. Unidentified Data is Still Informative (Experiment 2) This experiment introduced "unidentified" tokens (gray balls without visible serial numbers) to test reasoning under identity uncertainty. Participants successfully used the frequency of unidentified tokens to adjust their estimates of the total population. If gray tokens are common, it suggests a larger "hidden" population.
3. Categorization without Identification (Experiment 3) Using a "Kwiba eggs" scenario, participants had to infer the size variance of a category based on tokens whose individual identities were never revealed. The experiment showed that humans can learn the properties of a category (like how much the size of an egg varies) even if they never know which specific egg is which.
Participants in the "1 egg" condition (where evidence suggested all tokens were the same egg) correctly inferred a narrower category variance than those in the "4 eggs" condition, despite the visual tokens being identical in size across both groups. The "open world" model explains this by integrating over all possible identity assignments.
Limitations and Future Work
While the "open world" model provides a strong qualitative fit to human judgments, the authors note several limitations to address in future research:
- Simplicity of Features: The experiments used very simple features (color, alphanumeric strings, or 1D size). Real-world object discovery involves complex, high-dimensional perceptual data.
- Noise-Free Observations: The current model assumes features are observed without noise. A straightforward extension could incorporate perceptual noise to handle cases where two different objects look nearly identical.
- Scalability: The model was evaluated by enumerating all possible worlds, which is computationally expensive. Future work will need to employ more efficient sampling methods (like MCMC) for complex environments.
- Dynamic Environments: The model assumes a static urn; it does not yet account for objects that appear, disappear, or change over time.
Summary updated