Using Big Data to Understand Memory and Future Thinking

AbstractImagining the future and remembering the past both involve mental time travel. This commonality could indicate shared mental processes, as held by the Constructive Episodic Simulation Hypothesis (Schacter & Addis, 2008), or else interactive processes that complement one another, a possibility we call the Complementarity Hypothesis. According to the Complementarity Hypothesis, future thoughts are constructed from schemas making them episodically poor, whereas past thoughts are constructed from schemas and direct retrieval of memory traces, making them relatively episodically rich. We tested these hypotheses using machine learning to data mine mental operations in language, much as a geologist can recover physical processes from the geological record. People’s natural, unprompted talk on web blogs was automatically analyzed for past, present, and future references using a temporal orientation classifier. In Study 1, we found that perceptual details were mentioned more often in past than future talk, implying greater use of episodic processing in past than future thinking. In Study 2, a neural network using schemas generated from Latent Dirichlet Allocation better predicted the content of references to the future than the past, implying that constructive processes are more common in future than past thinking. In Study 3, we used the results from the two prior studies to construct an episodic-by-constructive process space. We adapted techniques from fMRI analysis to analyze this space for clusters of activity, as if the frequency of past and future thinking were BOLD responses in cortical space. We found that past and future thinking occupy highly separable regions of processing space, supporting the Complementarity Hypothesis.

Return to previous page