A theory of learning to infer.

Ishita Dasgupta, Eric Schulz, Joshua B. Tenenbaum, Samuel J. Gershman

TL;DR

This paper proposes a new theory of "learning to infer" where the brain doesn't use a fixed, general-purpose Bayesian inference engine. Instead, it learns to infer by optimizing a recognition model based on the distribution of queries it encounters. This "learned inference" explains why humans deviate from optimal Bayesian reasoning in systematic ways, like under-reacting to prior information or evidence, and how these deviations can be controlled by the learning history.

What Problem Does This Paper Address?

Traditional Bayesian models of cognition assume humans are rational agents who perfectly integrate probabilities. However, a wealth of empirical evidence shows systematic deviations from this ideal, often in seemingly contradictory directions. For instance, people sometimes "under-react" to prior probabilities (base rate neglect) and other times under-react to new evidence (conservatism). These errors aren't random; they're systematic and context-dependent. The challenge is to explain why these deviations occur and how they can be so varied, rather than simply assuming people are irrational or use ad-hoc heuristics. The authors aim to bridge this gap by proposing a theory that accounts for these systematic, context-dependent inferential errors within a rational framework.

Methods

The core of the paper introduces the "Learned Inference Model" (LIM). This model is framed as a Learned Inference Model (LIM), which uses a neural network function approximator to learn an approximate posterior distribution. Crucially, this is not a fixed algorithm but one that optimizes its parameters based on the distribution of queries it encounters. This optimization process, akin to blackbox variational inference, allows the model to learn to infer by adapting to the query distribution. The model's "computational bottleneck" (e.g., the number of hidden units in the neural network) imposes resource limitations, forcing it to prioritize accuracy for high-probability queries. This leads to systematic errors when dealing with less frequent or less informative queries. The authors also explore amortization, where computations are shared across queries, and regularization, where prior experience influences current judgments.

Datasets

The paper draws on and re-analyzes data from several existing experimental paradigms, including:

  • Balls-in-urns (or poker chip) setups: Used to study base rate neglect and conservatism.
  • Taxi cab problem: A classic example for base rate neglect.
  • Coin spin experiments: Investigating sensitivity to sample size and diagnosticity.
  • Payoff prediction tasks: Examining inference in continuous domains.
  • Scene statistics experiments: Studying memory effects and subadditivity.
  • Believable vs. unbelievable scenarios: Investigating belief bias.

The authors also describe a new experiment involving an urn task to test their predictions about how query distribution manipulation affects sensitivity to priors and likelihoods.

Key Insights and Ideas

  • "Learning to Infer" as a core mechanism: The brain doesn't have a fixed Bayesian inference algorithm; it learns to infer by optimizing an approximate posterior based on past experience.
  • Query distribution shapes inference: The frequency and informativeness of past queries dictate where the inference engine allocates its limited resources, leading to systematic biases.
  • Limited capacity explains errors: Computational constraints (like a fixed bottleneck in a neural network) force the model to make trade-offs, leading to under-reaction to less informative sources.
  • Amortization as memory: Reusing computations across queries (amortization) creates a form of memory that influences current judgments, explaining sequential effects and belief bias.
  • Context-sensitivity of errors: The same inferential error (e.g., under-reaction) can manifest differently depending on the learned query distribution, resolving apparent contradictions in previous research.

Key Results

  • The Learned Inference Model (LIM) successfully reproduces several key findings from the literature on probabilistic reasoning errors, including:
  • Under-reaction to likelihood and prior: The model shows systematic under-weighting of both evidence and base rates, consistent with empirical findings.
  • Sample size effects: The model exhibits differential sensitivity to evidence strength and weight, mirroring human behavior.
  • Belief bias: The model shows closer adherence to Bayesian norms when probabilities are consistent with prior experience.
  • Memory effects: The model demonstrates how past judgments can bias subsequent ones, especially when queries are similar.
  • Effect of experimental design: The model explains why within-subjects designs lead to greater sensitivity to base rates than between-subjects designs.
  • The model's performance is sensitive to its capacity (e.g., number of hidden units), with lower capacity leading to more pronounced inferential errors.
  • The authors' new experiment confirms that manipulating the query distribution (i.e., the relative informativeness of priors vs. likelihoods during training) systematically influences how subjects (and the model) weigh these sources of information.

Limitations and Future Work

  • Model representation: The current model uses a neural network function approximator without explicit representational structure. Future work could explore whether other functional forms, particularly those incorporating compositionality, better capture human inference.
  • Complexity of real-world data: The current model focuses on relatively simple, well-defined data-generating processes. Real-world learning involves noisy, high-dimensional inputs and the need to infer causal structure, which is not explicitly modeled.
  • Unification of inference methods: The paper discusses integrating variational inference with sampling-based approaches (like Monte Carlo methods) and suggests avenues for future research.
  • Flexible computational bottleneck: The assumption of a fixed computational bottleneck could be relaxed to explore how capacity might dynamically adjust based on the cost-benefit of greater accuracy.
  • Causal structure and prior knowledge: Future work could incorporate how pre-experimental knowledge and intuitions about causal structure influence inference.

Summary updated

Abstract

Bayesian theories of cognition assume that people can integrate probabilities rationally. However, several empirical findings contradict this proposition: human probabilistic inferences are prone to systematic deviations from optimality. Puzzlingly, these deviations sometimes go in opposite directions. Whereas some studies suggest that people underreact to prior probabilities (<i>base rate neglect</i>), other studies find that people underreact to the likelihood of the data (<i>conservatism</i>). We argue that these deviations arise because the human brain does not rely solely on a general-purpose mechanism for approximating Bayesian inference that is invariant across queries. Instead, the brain is equipped with a recognition model that maps queries to probability distributions. The parameters of this recognition model are optimized to get the output as close as possible, on average, to the true posterior. Because of our limited computational resources, the recognition model will allocate its resources so as to be more accurate for high probability queries than for low probability queries. By adapting to the query distribution, the recognition model learns to infer. We show that this theory can explain why and when people underreact to the data or the prior, and a new experiment demonstrates that these two forms of underreaction can be systematically controlled by manipulating the query distribution. The theory also explains a range of related phenomena: memory effects, belief bias, and the structure of response variability in probabilistic reasoning. We also discuss how the theory can be integrated with prior sampling-based accounts of approximate inference. (PsycInfo Database Record (c) 2020 APA, all rights reserved).
BibTeX
@inproceedings{dasgupta2020theory,
  author = {Ishita Dasgupta and Eric Schulz and Joshua B. Tenenbaum and Samuel J. Gershman},
  title = {{A theory of learning to infer.}},
  booktitle = {Psychological},
  year = {2020},
  doi = {10.1037/rev0000178},
}

Loading related content...

Figure