As experimentation in the behavioral and social sciences moves from brick-and-mortar laboratories to the web, techniques from human computation and machine learning can be combined to improve the design of experiments. Experimenters can take advantage of the new medium by writing complex computationally mediated adaptive procedures for gathering data: algorithms.
In a paper to be presented at CrowdML’16, we consider this algorithmic approach to experiment design. We review several experiment designs drawn from the fields of medicine, cognitive psychology, cultural evolution, psychophysics, computational design, game theory, and economics, describing their interpretation as algorithms. We then discuss software platforms for efficient execution of these algorithms with people. Finally, we consider how machine learning can optimize crowdsourced experiments and form the foundation of next-generation experiment design.
Consider the transmission chain, an experimental technique that, much like the children’s game Telephone, passes information from one person to the next in succession. As the information changes hands, it is transformed by the perceptual, inductive, and reconstructive biases of the individuals. Eventually, the transformation leads to erasure of the information contained in the input, leaving behind a signature of the transformation process itself. For this reason, transmission chains have been particularly useful in the study of language evolution and the effects of culture on memory.
When applied to functional forms, for example, transmission chains typically revert to a positive linear function, revealing a bias in human learning. In each row of the following figure, reprinted from Kalish et al. (2007), the functional relationship in the leftmost column is passed down a transmission chain of 9 participants, in all four instances reverting to a positive linear function.
Transmission chains can be formally modeled as a Markov chain by assuming that perception, learning, and memory follow the principles of Bayesian inference. Under this analysis, Bayes’ rule is used to infer the process that generated the observed data. A hypothesis is then sampled from the posterior distribution and used to generate the data passed to the next person in the chain. When a transmission chain is set up in this way, iterated learning is equivalent to a form of Gibbs sampling, a widely-used Markov chain Monte Carlo algorithm. The convergence results for the Gibbs sampler thus apply, with the prior as the stationary distribution of the Markov chain on hypotheses. This equivalence raises the question of whether other MCMC-like algorithm can form the basis of new experiment designs.
For more, attend CrowdML at NIPS in Barcelona or see our full paper: Rethinking experiment design as algorithm design.