Mixing memory and desire: How memory reactivation supports deliberative decision-making.
read more
Citations
A Context-Based Theory of Recency and Contiguity in Free Recall. Commentary. Authors' reply
Fixation patterns in simple choice reflect optimal information sampling.
Replay in minds and machines.
Developmental change in prefrontal cortex recruitment supports the emergence of value-guided memory
References
Reinforcement Learning: An Introduction
A spreading-activation theory of semantic processing
Cognitive maps in rats and men
Planning and Acting in Partially Observable Stochastic Domains
A Theory of Memory Retrieval.
Related Papers (5)
Memory Interference as a Determinant of Language Comprehension.
Frequently Asked Questions (18)
Q2. What have the authors stated for future works in "Mixing memory and desire: how memory reactivation supports deliberative decision-making" ?
A primary direction of future research is understanding how various factors influence the temporal dynamics of memory retrieval. Further work is needed to understand how within-trial dynamics affect the integration of information about potential future states. Evidence suggests the influence of at least the following terms: 1. semantic distance ( e. g. as estimated using word embeddings ; Chadwick et al., 2016 ), 2. episodic distance ( Polyn et al., 2009 ), and 3. the spread of probability mass across associations at each kind of distance ( Socher et al., 2009 ).
Q3. What is the role of the CA3 region in the computation of state inference?
The interaction between internally-generated sequences and the properties of external input is a critical feature of computational work on state inference , a necessary function for online planning in environments with uncertain latent contingency structure (Kaelbling et al., 1998; Rao, 2010) .
Q4. What is the role of the hippocampus in learning of uncertain states?
With its ability to extract sparse codes from sensory inputs, hippocampus is implicated in the learning of uncertain states by representing the latent contexts that give rise to observations (Gershman et al., 2010; Sanders et al., 2020) .
Q5. What is the key insight of the theory?
The key insight of the theory is that rather than encoding place in an absolute sense, the place cells encode a predictive representation of future states that reflects the relational structure between them (Stachenfeld et al., 2017) .
Q6. What is the main question raised by this framework?
One question raised by this framework is whether the temporal dynamics of memory reactivation are fundamental, adapt to the time available, or are modulated by the content of representation or computations being performed.
Q7. What is the main argument for the idea that memory retrievals influence decision-making?
Supporting the idea that memory retrievals’ influence on decision unfolds over time is theobservation that longer delays before choice lead to greater memory influence on decisions (Foerde & Shohamy, 2011) -- and, in particular, greater influence of extended retrievals from memory (Bakkour et al., 2019; Eldar et al., 2020; Gordon et al., 2014) .
Q8. What is the definition of an emerging framework?
An emerging framework describes this phenomenon as a simulation-driven estimation process, in which decision-makers examine what might result from each available action by consulting memories of similar previous settings.
Q9. What is the significance of the reactivation of previous experiences?
As reactivations of those previous experiences echo both previous sensory inputs and also latent, non-sensory information, such as the inferred contingency structure of the environment and the value of rewards available at the time, all of these lead to the subsequent reactivation of the same sorts of action-tendency or value associations as does sensory input.
Q10. How do the authors propose that such jumps arise?
the authors propose that one mechanism by which such jumps arise is via parallel sampling from multiple internal evidence sources which produce evidence at different latencies and frequencies.
Q11. What is the way to model the arrival of evidence samples from different distributions?
One especially promising approach for modeling the arrival of evidence samples from different distributions, called Lévy Flight models (Fig. 2), considers a variety of intermittent “jumps” that augment and alter the Brownian motion of Equation 7.
Q12. Why do Shohamy and Shadlen argue that memory-guided decisions take time?
In an insightful evaluation of this question, Shohamy and Shadlen ( 2016) propose that one reason memory-guided decisions take time, rather than acting instantly on internally-available information, is because a limited-bandwidth thalamocortical pipeline enforces serial processing.
Q13. What is the main reason why the researchers proposed that multiple forms of decisions depend on retrieval dynamics?
The proposal that multiple forms of decisions depend on retrieval dynamics that vary as a function of associative distance may explain why choices and response times appear to covary between tasks that examine how subjects weigh options across many kinds of such distances, for instance in intertemporal choice, patch foraging, and model-based planning (Kane et al., 2019; Shenhav et al., 2014) , each of which have been independently shown to depend on long-term memory representations (Palombo et al., 2015; Peters & Büchel, 2010; Schmidt et al., 2019; Vikbladh et al., 2019) .
Q14. What is the significance of the study of memory-guided decisions?
findings about the neural architecture of evidence integration in these other modalities are likely to apply to the study of memory-guided decisions, especially when studies employ stimuli whose predictiveness is estimated via associations that emerge across experience (Yang & Shadlen, 2007) .
Q15. What is the role of pattern completion in decision-making?
Pattern completion may be especially useful to decision-making because it allows past choices and outcomes to come to mind in situations that are similar to, but not exactly the same as, past encounters.
Q16. What are the features of memory representations that mediate the impacts of past experience?
the authors describe multiple kinds of memory representations, how they differently represent aspects of past experience, and how they lend themselves to different retrieval and transformation dynamics that later affect decision-making.
Q17. What model was used to reanalyze previously collected data?
Bornstein and colleagues ( 2017) also used the same model to reanalyze previously collected data from a four-choice decision task (Daw et al., 2006) , which further revealed that in addition to participants’ choices, neural decision variables measured in fMRI were better explained by a memory sampling model than by TDRL.
Q18. How does the alternative compute the reward for each state?
rather than computing this estimate by updating a cached value function with each experience (Eqn. 1), the alternative computes it dynamically, possibly even on-demand (Eldar et al., 2020) , by sampling past encounters with the states of interest (and, potentially, generalizing from similar states) and averaging the resulting values.