|
Episodic memory and causal reasoning through counterfactuals| title | Episodic memory and causal reasoning through counterfactuals |
|---|
| start_date | 2023/05/09 |
|---|
| schedule | 16h15-17h45 |
|---|
| online | yes |
|---|
| visio | https://univ-grenoble-alpes-fr.zoom.us/j/92229286091?pwd=aGpnKzBSRm5GdTRNdHQ1cEFhREx3Zz09 |
|---|
| location_info | On zoom |
|---|
| summary | Over the past several years, there has been a shift in researchers’ thinking about the functional role of episodic memory (Klein, 2013). Rather than focusing on how memory represents the past, recent literature often presents memory as ultimately dealing with the future – helping the organism to anticipate events and increase its adaptive success (Suddendorf & Corballis, 2007). We claim that episodic memory does yield adaptive success because of its crucial role in causal reasoning. To causally reason whether event A caused event B, the subject needs to cognitively evaluate a diachronic counterfactual of the following type: 1. If A had not happened at t1, B would not have happened at t. Such counterfactuals require episodic memory for their evaluation because, according to Lewis (1973), counterfactuals of type (1) are true just in case: 2. The not-A world closest to the actual world (where A and B occurred) is also a not-B world rather than a B-world. Current associative networks in AI struggle with diachronic temporal reasoning because these networks do not explicitly represent change but rather update representations as new information comes along (Hoerl & McCormack, 2019). Second, they learn by altering the weights/strength of connections through accumulating data over time. This approach significantly differs from human learning, which, in some instances, only requires single examples and allows rapid, radical re-evaluation of the causal relationships between events. Our paper presents a causal inference model based on the predictive processing framework of brain functioning (Clark, 2013, 2015; Hohwy, 2013; Friston 2005, 2010; Rao & Ballard, 1999) and minimal trace account of episodic memory (Werning, 2020). According to our model, because of its truth condition (2), evaluating counterfactuals of type (1) involves a) episodic memory to construct a scenario of the past (Cheng, Werning, Suddendorf, 2016), b) negation of the target event A in that scenario, and c) evaluation of the closeness relation between the newly attained non-A scenario and the remembered scenario. For example, to understand that the fact that your train was 30 minutes late caused you to miss the flight (rather than a 10-minute queue at the airport), you need to simulate the world as it existed when you expected the train to arrive according to the plan (a). Then, you need to simulate a world where the train did not get delayed by negating that event in the respective scenario (b). Finally, using semantic knowledge, you let the scenario unfold in time and ask yourself whether you would still have missed the flight or not (c). |
|---|
| responsibles | Righetti, Werning, Kourken, Andonovski |
|---|
Workflow history| from state (1) | to state | comment | date |
|---|
| submitted | published | | 2023/04/20 14:30 UTC |
| |
|