## Deconfounding Hypothesis Generation and Evaluation in Bayesian Models

### BibTeX

@MISC{(liz_deconfoundinghypothesis,

author = {Elizabeth Baraff Bonawitz (liz and Thomas L. Griffiths (tom},

title = {Deconfounding Hypothesis Generation and Evaluation in Bayesian Models},

year = {}

}

### OpenURL

### Abstract

Bayesian models of cognition are typically used to describe human learning and inference at the computational level, identifying which hypotheses people should select to explain observed data given a particular set of inductive biases. However, such an analysis can be consistent with human behavior even if people are not actually carrying out exact Bayesian inference. We analyze a simple algorithm by which people might be approximating Bayesian inference, in which a limited set of hypotheses are generated and then evaluated using Bayes ’ rule. Our mathematical results indicate that a purely computationallevel analysis of learners using this algorithm would confound the distinct processes of hypothesis generation and hypothesis evaluation. We use a causal learning experiment to establish empirically that the processes of generation and evaluation can be distinguished in human learners, demonstrating the importance of recognizing this distinction when interpreting Bayesian models.

### Citations

766 | Vision - Marr - 1982 |

589 | Probabilistic inference using Markov chain Monte Carlo methods (Tech
- Neal
- 1993
(Show Context)
Citation Context ...mpling is a Monte Carlo method that takes this one step further, substituting samples from another distribution (the surrogate distribution) for samples from the target distribution (for details, see =-=Neal, 1993-=-). Thus, if we wanted to perform computations involving p(x), we would generate a set of samples x1,...,xm from the surrogate distribution q(x). We can get away with doing this if we no longer assign ... |

481 |
The adaptive character of thought
- Anderson
- 1990
(Show Context)
Citation Context ... known as the likelihood. Bayes’ rule provides a computational-level theory of inductive inference, being a component of the optimal solutions to a variety of problems of reasoning under uncertainty (=-=Anderson, 1990-=-; Anderson & Schooler, 1991; Freeman, 1994; Geisler, Perry, Super, & Gallogly, 2001; Griffiths & Tenenbaum, 2007; Huber, Shiffrin, Lyle, & Ruys, 2001; Knill & Richards, 1996; Körding & Wolpert, 2004; ... |

284 | Analogical problem solving
- Gick, Holyoak
- 1980
(Show Context)
Citation Context ...t al., 1993). Hypotheses can be generated in many different ways, including detecting cues from context, recognizing similarities to previous experiences, and making analogies to other domains (e.g., =-=Gick & Holyoak, 1980-=-; Gentner, 2002; Nersessian, 1992; Koslowski, 1996). We will not attempt to model these processes here, but for our purposes, it is sufficient to assume that the result of all of these processes can b... |

210 |
Reflections of the environment in memory
- Anderson, Schooler
- 1991
(Show Context)
Citation Context ...kelihood. Bayes’ rule provides a computational-level theory of inductive inference, being a component of the optimal solutions to a variety of problems of reasoning under uncertainty (Anderson, 1990; =-=Anderson & Schooler, 1991-=-; Freeman, 1994; Geisler, Perry, Super, & Gallogly, 2001; Griffiths & Tenenbaum, 2007; Huber, Shiffrin, Lyle, & Ruys, 2001; Knill & Richards, 1996; Körding & Wolpert, 2004; Shiffrin & Steyvers, 1997; ... |

187 |
A Model for Recognition Memory: REMRetrieving Effectively from Memory
- Shiffrin, Stevyers
- 1997
(Show Context)
Citation Context ...; Anderson & Schooler, 1991; Freeman, 1994; Geisler, Perry, Super, & Gallogly, 2001; Griffiths & Tenenbaum, 2007; Huber, Shiffrin, Lyle, & Ruys, 2001; Knill & Richards, 1996; Körding & Wolpert, 2004; =-=Shiffrin & Steyvers, 1997-=-; Weiss, Simonvelli, & Adelson, 2002). As an account of inductive inference, the prior p(h) captures the inductive biases of the learner, indicating which hypothesis a learner will favor when multiple... |

135 |
Bayesian integration in sensorimotor learning
- Koerding, Wolpert
- 2004
(Show Context)
Citation Context ...certainty (Anderson, 1990; Anderson & Schooler, 1991; Freeman, 1994; Geisler, Perry, Super, & Gallogly, 2001; Griffiths & Tenenbaum, 2007; Huber, Shiffrin, Lyle, & Ruys, 2001; Knill & Richards, 1996; =-=Körding & Wolpert, 2004-=-; Shiffrin & Steyvers, 1997; Weiss, Simonvelli, & Adelson, 2002). As an account of inductive inference, the prior p(h) captures the inductive biases of the learner, indicating which hypothesis a learn... |

133 | Edge co-occurrence in natural images predicts contour grouping performance
- Geisler, Perry, et al.
- 2001
(Show Context)
Citation Context ...onal-level theory of inductive inference, being a component of the optimal solutions to a variety of problems of reasoning under uncertainty (Anderson, 1990; Anderson & Schooler, 1991; Freeman, 1994; =-=Geisler, Perry, Super, & Gallogly, 2001-=-; Griffiths & Tenenbaum, 2007; Huber, Shiffrin, Lyle, & Ruys, 2001; Knill & Richards, 1996; Körding & Wolpert, 2004; Shiffrin & Steyvers, 1997; Weiss, Simonvelli, & Adelson, 2002). As an account of in... |

100 | Structure and strength in causal induction
- Griffiths, Tenenbaum
- 2005
(Show Context)
Citation Context ...underdetermined hypotheses. In the last decade an increasing number of papers have argued that people solving inductive problems act in ways that are consistent with optimal Bayesian inference (e.g., =-=Griffiths & Tenenbaum, 2005-=-; Goodman, Tenenbaum, Feldman, & Griffiths, 2008; Xu & Tenenbaum, 2007). However, most of these analyses operate at what Marr (1982) termed the computational level, using Bayesian inference to identif... |

84 | Children and adults as intuitive scientists
- Kuhn
- 1989
(Show Context)
Citation Context ...rs possess all relevant hypotheses before seeing data is at odds with numerous findings suggesting that generating appropriate hypotheses can be one of the hardest parts of inductive inference (e.g., =-=Kuhn, 1989-=-; Klahr, Fay, & Dunbar, 1993). We thus consider the consequences of separating the processes of generating hypotheses and evaluating those hypotheses, assuming that learners perform Bayesian inference... |

80 |
How do Scientists Think? Capturing the dynamics of conceptual change in science
- Nersessian
- 1992
(Show Context)
Citation Context ...ated in many different ways, including detecting cues from context, recognizing similarities to previous experiences, and making analogies to other domains (e.g., Gick & Holyoak, 1980; Gentner, 2002; =-=Nersessian, 1992-=-; Koslowski, 1996). We will not attempt to model these processes here, but for our purposes, it is sufficient to assume that the result of all of these processes can be summarized in a single probabil... |

80 |
Motion illusions as optimal percept
- Weiss, Simoncelli, et al.
- 2002
(Show Context)
Citation Context ...; Freeman, 1994; Geisler, Perry, Super, & Gallogly, 2001; Griffiths & Tenenbaum, 2007; Huber, Shiffrin, Lyle, & Ruys, 2001; Knill & Richards, 1996; Körding & Wolpert, 2004; Shiffrin & Steyvers, 1997; =-=Weiss, Simonvelli, & Adelson, 2002-=-). As an account of inductive inference, the prior p(h) captures the inductive biases of the learner, indicating which hypothesis a learner will favor when multiple hypotheses are equally consistent w... |

69 |
The generic viewpoint assumption in a framework for visual perception
- Freeman
- 1994
(Show Context)
Citation Context ...des a computational-level theory of inductive inference, being a component of the optimal solutions to a variety of problems of reasoning under uncertainty (Anderson, 1990; Anderson & Schooler, 1991; =-=Freeman, 1994-=-; Geisler, Perry, Super, & Gallogly, 2001; Griffiths & Tenenbaum, 2007; Huber, Shiffrin, Lyle, & Ruys, 2001; Knill & Richards, 1996; Körding & Wolpert, 2004; Shiffrin & Steyvers, 1997; Weiss, Simonvel... |

58 |
Theory and evidence: the development of scientific reasoning
- Koslowski
- 1996
(Show Context)
Citation Context ...rent ways, including detecting cues from context, recognizing similarities to previous experiences, and making analogies to other domains (e.g., Gick & Holyoak, 1980; Gentner, 2002; Nersessian, 1992; =-=Koslowski, 1996-=-). We will not attempt to model these processes here, but for our purposes, it is sufficient to assume that the result of all of these processes can be summarized in a single probability distribution ... |

48 | A rational analysis of rule-based concept learning
- Goodman, Tenenbaum, et al.
- 2008
(Show Context)
Citation Context ...n the last decade an increasing number of papers have argued that people solving inductive problems act in ways that are consistent with optimal Bayesian inference (e.g., Griffiths & Tenenbaum, 2005; =-=Goodman, Tenenbaum, Feldman, & Griffiths, 2008-=-; Xu & Tenenbaum, 2007). However, most of these analyses operate at what Marr (1982) termed the computational level, using Bayesian inference to identify the hypotheses that an ideal learner with part... |

43 | Perception and preference in short-term word priming
- Huber, Shiffrin, et al.
- 2001
(Show Context)
Citation Context ...imal solutions to a variety of problems of reasoning under uncertainty (Anderson, 1990; Anderson & Schooler, 1991; Freeman, 1994; Geisler, Perry, Super, & Gallogly, 2001; Griffiths & Tenenbaum, 2007; =-=Huber, Shiffrin, Lyle, & Ruys, 2001-=-; Knill & Richards, 1996; Körding & Wolpert, 2004; Shiffrin & Steyvers, 1997; Weiss, Simonvelli, & Adelson, 2002). As an account of inductive inference, the prior p(h) captures the inductive biases of... |

37 |
Heuristics for scientific experimentation: A developmental study
- Klabr, Fay, et al.
- 1993
(Show Context)
Citation Context ...ll relevant hypotheses before seeing data is at odds with numerous findings suggesting that generating appropriate hypotheses can be one of the hardest parts of inductive inference (e.g., Kuhn, 1989; =-=Klahr, Fay, & Dunbar, 1993-=-). We thus consider the consequences of separating the processes of generating hypotheses and evaluating those hypotheses, assuming that learners perform Bayesian inference with only the set of hypoth... |

23 | Priming, analogy, and awareness in complex reasoning - Schunn, Dunbar - 1996 |

21 | Learning Causal Laws - Tenenbaum, Niyogi - 2003 |

14 | Two proposals for causal grammars
- Griffiths, Tenenbaum
- 2007
(Show Context)
Citation Context ... being a component of the optimal solutions to a variety of problems of reasoning under uncertainty (Anderson, 1990; Anderson & Schooler, 1991; Freeman, 1994; Geisler, Perry, Super, & Gallogly, 2001; =-=Griffiths & Tenenbaum, 2007-=-; Huber, Shiffrin, Lyle, & Ruys, 2001; Knill & Richards, 1996; Körding & Wolpert, 2004; Shiffrin & Steyvers, 1997; Weiss, Simonvelli, & Adelson, 2002). As an account of inductive inference, the prior ... |

11 |
Analogy in scientific discovery: The case of Johannes Kepler
- Gentner
- 2002
(Show Context)
Citation Context ...es can be generated in many different ways, including detecting cues from context, recognizing similarities to previous experiences, and making analogies to other domains (e.g., Gick & Holyoak, 1980; =-=Gentner, 2002-=-; Nersessian, 1992; Koslowski, 1996). We will not attempt to model these processes here, but for our purposes, it is sufficient to assume that the result of all of these processes can be summarized in... |

2 |
Perception as Bayesian inference. Cambridge: Cambrdige
- Knill
- 1996
(Show Context)
Citation Context ...ms of reasoning under uncertainty (Anderson, 1990; Anderson & Schooler, 1991; Freeman, 1994; Geisler, Perry, Super, & Gallogly, 2001; Griffiths & Tenenbaum, 2007; Huber, Shiffrin, Lyle, & Ruys, 2001; =-=Knill & Richards, 1996-=-; Körding & Wolpert, 2004; Shiffrin & Steyvers, 1997; Weiss, Simonvelli, & Adelson, 2002). As an account of inductive inference, the prior p(h) captures the inductive biases of the learner, indicating... |