Talk: Evolution of Memory

I presented the following at University of Maryland, College Park, 30 Oct 2017. It summarises three papers with constructive feedback on where to improve their methodology.

The bottom line is simple: we know memory is fallible and that we evolved this sort of memory mechanism rather than just a purely rigidly veridical mechanism — the question is why evolve a seemingly imperfect mechanism?

Abstract: The following three approaches show that updating information in novel situations (rather than a well-defined niche) differentiates the distinctly human form of memory from that which non-human agents possess: we need to update information as time passes and as social arrangements change (not so much the environment in which we must survive and reproduce, but rather, in the uniquely human terrain or social landscape, ie. regarding what is “due” others as well as, or, more importantly, what is “due” us in particular). Rigid memory serves us well (and we seem to possess this just as non-human animals do, in cases such as locating resources ); but it breaks down in social interactions when we must perform so-called moral book-keeping to disentangle our ever-changing social obligations (and, more saliently, what others owe us — as human memory has an ego-driven, self-knowing, meta-representational character).

Mahr & Csibra (2017) “Why do we remember? The communicative function of episodic memory”

Mahr and Csibra argue that common accounts of episodic memory function have shortcomings, and suggest an alternative functional account, claiming that it plays a generative justificatory role in forming beliefs about past events (p 58). Their positive thesis is that “episodic memory is essential to managing our discursive commitments by demarcating the range of beliefs about which we can claim epistemic authority.” To support this thesis they attempt to discredit other accounts of episodic memory, suggesting instead that it has a social discursive role by which we communicate past events. In turn, this communication requires accounting for the truth of an assertion. Finally, they claim their account makes sense of why episodic memory admits of self-reference, and, lastly, how it evolved as a set of cognitive tasks with specific problem to solve.

  1. What is episodic memory?

Tulving (1972): episodic memory is memory for personally experienced past events; this includes information about what happened, when and where. Tulving expanded this definition (1983; 1985; 2002), as it overlaps with semantic memory, by adding that episodic memory has a unique phenomenology: we “know” the contents of semantic memory; episodic memory comes as a re-experiencing of that event as it occurred (p 7). On the face of it, this sounds difficult to support. Weak recreation of past events seems trivial, but a mechanism that guarantees strong recreation seems evolutionarily inefficient (although certainly effective for the task).

Question I had in reading is if either semantic or episodic memory overlap with the process described by Evans’ Generality Constraint (or do one or both act by pumping temporal content for the Generality Constraint to in turn parse). This seemed tangential at first, but the section marked (#) later warrants closer consideration.

Memory-belief congruency (p 20): memory is guided by the present beliefs rather than having us form beliefs grounded in accurate/veridical historical narrative committed to memory.

But is this “I remember what I choose to” (ie. a filter)? Or is more self-referential, and by recalling things that do not contradict present beliefs, one “re-remembers” things in a way that more resembles the current belief (ie. revision)?

“By now, there is an impressive literature showing that it is possible to induce in people vivid, detailed false memories … induced beliefs can guide constructive retrieval.” (p 22)

This is a massive ethical hazard and I believe this warrants discussion. Another issue is that memory retrieval has a tendency to confirm prior beliefs rather than contradicting them — reconstructing memory goes for consistency rather than accuracy (p 22f). Examples: The so-called Mandela effect, the benign controversy over the spelling of the Berenstain/Berenstein Bears childrens books, etc.

  1. What is episodic memory for?

It will not suffice to explain episodic memory by just saying what we can use it for — we must answer what evolutionary pressure (“fitness relevant problem”) led to its emergence and stabilization, over less costly features (p 26). Alternatively, one could just be lazy and go with a Chaos theory approach and claim that episodic memory just happened to emerge as a physiological mechanism, and just happened to stick around due to conferring more advantages than disadvantages. In either case, these costs include the recognition that episodic memory comes with remembering painful episodes rather than just good ones, and that the already calorically expensive human brain must consume more energy in order to host this mechanism.

Klein et al 2000: information about the past is important only in so far as it enables us to make better decisions in the present so as to ensure benefits in the future. Hippocampal lesions lend support to episodic memory having a shared mechanism with future-oriented thinking; the authors identify the ability for future-oriented mental time-travel as a possible fitness relevant problem — in such a view, the utility future planning alone led to developing this mechanism. The problem with this is that it does not explain we retain the past at all (p 27f). Alternatively, and much more efficiently, we could have some kind of inductive learning sub-mechanism that pumps content (not necessarily temporal) based on past lessons but not including the salient narrative phenomenology. Furthermore, as the authors state, just replaying the past does not adequately equip us to deal with contingencies (p 28).

(#) As stated, hippocampal lesions coincide with the ability to imagine personal future (p 27); however, the authors later state that the loss of the capacity for episodic memory alone does not impair significantly people’s ability to draw inferences about the future (p 30). So the authors reject the mental time-travel position.

Michaelian (2012, 2016) proposes that episodic memory simulates the future yet still has reliable veridicality because the mechanism functions for the sake of source monitoring. But the authors clearly demonstrated that retrieval and reevaluation leave memory vulnerable to change and manipulation (whether nefarious or not) — and evolution selecting for an otherwise unreliable mechanism seems odd unless something else is at play (ie. we need that sort of eidetic plasticity). Furthermore, the source monitoring position leads to an infinite regress — the thing guaranteeing reliability will need something else to guarantee its own reliability (30f).

The last position the authors evaluate is one in which episodic memory is a byproduct of our ability to mentally time-travel into the future– but the evidence already mention clearly demonstrates these are independent mechanisms; the authors further state that subjective past and subjective future play different roles in inference and action despite having similar content (p 32f).

  1. Their proposal: a communicative function of episodic memory

Firstly, episodic memory has epistemic generativity with an autonoetic (“self-knowing”; denoting meta-representation) character but is prone to belief-congruency. Secondly, the authors address the content of episodic memory (why we require a capacity for specific past events). Their position is that episodic memory plays a generative role in belief-formation, but committing a complete historical narrative to memory is inefficient — so we store inferences without the steps we took to reach such. The authors invoke the work of the evolutionary psychologists Cosmides and Tooby (2000), who state that such inferences have a dual role: we can use them for human communicative interaction (p 33).

I find that any account of episodic memory ought have to have a dual role — especially if it portends to be an evolutionary account. I find this compatible with the thesis in Opacity of Mind that human mind-reading mechanisms are the same whether directed at the self or upon others. But I also wonder if episodic memory would work in this fashion without theory of mind. Regardless of the answer to this, the authors claim the benefit is aggregate to the group rather than just the individual (ie. constrained self-interest). In the literature, this defector problem goes all the way back to Plato’s Ring of Gyges, and was probably the first novel problem that humans had to face in the savannah.

Consequences of discursive commitment (“avoid reputation damage through consistency”) (p 41).

Epistemic vigilance: source monitoring allows us to defend against or commit to the beliefs and claims that others make (p 48).

Social role specifically instantiated (p 50) — but the literature I dug up from the same time period shows that there is conflict: As far as memory as a proxy for trustworthiness goes, it remains open how this compares to other proxies such as social status, prestige, wealth and so on. One may be more inclined to remember the untrustworthiness of another individual if they perceived said individual of lower standing — this rank-based reason develops in children as young as 3 years of age and before the development of other forms of transitive reasoning — and early than episodic memory. For more on this, see: Cummins’ 1998, 2005 papers on Dominance Theory, as well as Hyungwook Yim, Simon J. Dennis, Vladimir M. Sloutsky 2013 “The Development of Episodic Memory Items, Contexts, and Relations” (in which they claim that episodic memory emerges between 4-7 years of age).

The format of episodic memory (p 52): epistemic generativity allows metarepresentation of the reasons for beliefs about past events so as to give reasons in testimony. So this structure has a Claim-Ground-Warrant structure.

Book-keeping: “social commitments have evolutionary significance because they make it possible for parties in an exchange to gain benefits that would be unattainable in the face of the risk of detection” (p 56).

I agree with this but even amoebas can form coalitions with complexity that would interest game theorists (ie. when dictyostelids form a grex). But I think (for the sake of discussion) the book-keeping aspect is important as human behaviour has dramatically variegated and dynamic landscape, and concrete belief that is closed to revision means that we cannot update information as the landscape changes. This, and the autonoetic portion of episodic memory, make it important — we need to know what we are personally owed even if we throw out the rest of the social state of affairs.

Some evolutionary pressures (are these unique to humans?) include sustainment in a harsh environment: hunting and gathering (locations of resources, how to survive a fight, craft tools), and navigating uniquely (or not?) human terrain: so-called “cheater detection” (but this is also redundant — some physical features are a proxy for untrustworthiness regardless of previous interactions; see also Cummins on transitive rank-based deontic reasoning – citation elsewhere in paper). Are there other pressures, or do almost all of the problems solved by episodic memory as explained by the authors fall into these categories? More importantly, do we still face these exact pressures?

Suddendorf, Addis, Corballis (2009) “Mental time travel and the shaping of the human mind”

Evidence from physiological damage (specifically hippocampal damage) shows semantic memory does not deteriorate as much as episodic memory. On the other hand, so-called semantic dementia (not Alzheimer’s disease; according to Hodges & Graham 2001, semantic dementia results from focal atrophy on the antero-lateral temporal lobes) shows no impact on episodic memory.

Question posed in paper: In light of/in spite of this dissociation between semantic and episodic memory at the functional and structural level, what links the two? (end of p 1317). Do the authors adequately explain this?

From this, the paper makes the following claims: episodic memory depends upon semantic memory; recalling past events involves reconstructing episodic memories with semantic details; this dual-process mechanism in turn allows for future-oriented thought; and that communicating such episodes put an evolutionary pressure on humans to come up with languages.

  1. Intro

Semantic memory: memory for enduring facts about the world.

Episodic memory: a personal record of the past (Tulving 1983).

Episodic memory and semantic memory are doubly disassociated, implying distinct mechanisms Evidence for this comes from cases of hippocampal damage coinciding with impaired episodic memory. On the other hand, cases of focal atrophy on the antero-lateral temporal lobes show impaired semantic memory.

Evidence suggests that the episodic memory system plays a role in mental time travel into the future, suggesting a general concept of mental time travel. Elaborating on this, they say episodic memory is better conceived as a conscious act of construction (which leaves it open to revision from outside influences, nostalgic want of a past that never was, nihilism, wishful thinking, and so on), rather than a faithful re-enactment of the past. The authors concede that it does not act in a reproductive manner, duplicating experiences in a faithful way: it admits of error. From this it follows that natural selection acted on what memory offers the present and future (rather than accuracy of recollection).

  1. Neuropsychological and Neuroimaging Evidence

Core network for remembering the past, in addition to a more adaptive function of integrating information from the past to simulate future events. This network relies on regions of the medial prefrontal cortex, the lateral and media parietal cortex, and the lateral and medial temporal lobes, including the hippocampus. The authors identify a methodological problem — these regions light up on past events more than thinking about the future, but subjects are asked about the past more often than the future in that they were asked to think about the future and then were asked to think about what they just thought (ie. a past-directed thought about the future (p 1318).

A further methodological issue I have with this sort of imaging is how finely grained and intricate the temporal content the subject is asked to parse — why just past vs present vs future rather than more intricate clauses (such as when we discuss ongoing events, narration in a book we read, past desires for events contingent on the future, and so on).

  1. Uniquely Human?

“…episodic memory provides a vocabulary from which to construct possible scenarios” (p 1319).

A key methodological issue is that the claim that episodic memory is unique to humans is a default position, we do not have access to direct report from non-human animals. Observing animals runs the general philosophical risk of anthropomorphizing their actions. Humans ascribe intentionality even to computer graphics (Barrett et al 2005 “Accurate judgments of intention from motion cues alone: A cross-cultural study”). Looking at the physiology has its own ethical and methodological issues — other species have hippocampus (but primates’ are dramatically larger — their brains are larger, but what is the wiring density?).

The ethical issue is that this affects the moral status (according the authors) of the animal; but some would say the ability to experience pain is a better signifier.

Furthermore, even if animals had temporal content (satisfying the what-when-where criteria), this does not entail anticipation of future events or past recall. If the Generality Constraint means animals can parse temporal content, this alone does not warrant their possession of a phenomenology of episodic memory. A further blind is that, again, observable animal behaviour can be more easily reduced to instinct than recall due to the methods being used, and we do not have strong evidence of novelty/generativity (p 1319f).

  1. Language

Testing for episodic memory in humans relies upon language.

Language allows past episodes and future plans to be shared. This requires the breakdown of episodes into constitutive representative, narrative elements (p 1320f). Interestingly, the authors note that the ways time itself is represented in language varies widely between language (Mandarin lacks tenses, French has over a dozen). The authors point to evidence that the human hippocampus co-evolved with language use, before noting that language and mental time travel have a nexus in generativity, the temporal constraints of language form a mental spatio-temporal topology (“the array of possible arguments one can make”) with type parameterization (rather than us accurately and completely narrating everything we experience in the world). Question — does this a priori commit us to quasi-anti-realism about memory? This take-away interests me, as it strongly suggests that we have hard narrative typing going on, rather than direct experience of temporal, spatial reality. If memory were to be fully veridical, this would not be the case. This may be an artefact of selection favouring efficiency.

  1. Evolutionary Considerations

Language and mental time travel (probably) co-evolved.

Little exposition needed beyond reiterating the question of so-called “costly fitness” — are the advantages conferred by episodic memory worth the price (we remember painful events, sometimes excessively) and biases (we believe the future to be better than we have warrant to)?

Allen & Fortin (2013) “The evolution of episodic memory”

This paper asks “When did episodic memory emerge?” and concludes that “Unfortunately, the available evidence cannot support a definitive answer at this time” (p 10385). Along the way, Allen & Fortin look at the structural basis of episodic memory, finding it rather widespread, before looking at salient functional aspects.

  1. Episodic Memory Capacity Across Species

Receiver Operating Characteristics approach: informational responses differentiate episodic recollection against familiarity in a recognition task by comparing signals. Drawback is that it does not seamlessly work across species.

Events-in-context-approach differs from this in parsing behaviour based on the criteria of content (object + spatial location + temporal location), structural integration forming an “event,” and flexibility (ie. novelty/generativity).

Three models for events in context (p 10380):

Memory for what-where-when (higher threshold for behavioural demonstration, but the authors note this has been observed in many animals with the caveat that “there is no clear evidence that all episodic memories contain all three types of information”).

Memory for what-where (association between specific items in specific places).

Memory for what-when (a defining feature of episodic memory; but also indicates awareness of sequentially occurring events — Pavlovian conditioning might not work without this).

Methodological question: in vitro (brain anatomy; fMRI and other imagery – can we have the same neural hardware perform different tasks (within and across species)? ) vs in vivo (behavioural expression – do have access to this; content of thoughts – no access to this). If episodic memory has temporal content, at some point we have to say just what that content is (a computer can do this for us if we use natural language processing) — the what/where/when criteria serves as one candidate. Although we can extract this from more general narrative content, it becomes more difficult to concretize it from observing behaviour (especially in non-human agents). It becomes doubly insidious as it is easy to put a human-oriented intentional gloss on the actions of non-human agents.

  1. Brain Structures Important for Episodic Memory

“the hippocampus, parahippocampal region and prefrontal cortex form a neural system that is thought to underlie episodic memory in humans, but this basic neurobiology is not unique to humans…” continuing that other mammals have it, a comparable circuit exists in birds, and reptiles and bony fish have homologous regions. This supports a last common ancestor developed this rather than it being the product of parallel evolution (p 10382).

But structure alone is not enough to claim full similarity in function, especially at this level of complexity and timescale since last common ancestor: they may all have a method of parsing (temporal) content, but the modalities conveyed by that content may radically differ, may depend on factors including ecological niche and perceptual apparatus. Furthermore, this literature does not mention hippocampal-size-to-body-size ratio – primates have a much larger ratio than other mammals.

One might just push the question back, and say that this circuit generates, stores, and provides (rather than processes) temporal content – for another (species-differential) circuit to in turn process (I feel somewhat comfortable making this claim on the basis of (# “that the loss of the capacity [eg. hippocampal damage] for episodic memory alone does not impair significantly people’s ability to draw inferences about the future (Mahr & Csibra p 30)). In this way, we can reconcile structural similarities with differences in consciousness.

  1. Neural Mechanisms Underlying Episodic Memory

Object and spatial awareness form parallel, segregated streams, but temporal awareness is less understood; they are integrated at the hippocampus.

“Episodic recall is thought to occur when the integrated event-in-context representation is reactivated, involving a pattern completion process that can be initiated by cueing the hippocampal network with elements of the event or context” (p 10383). Discussion point: might this pattern-filling mechanism be a point of bias or revision introduction? I can go into the hypothetical mechanics of this pattern-filling if time permits and interest exists.

  1. Functions of Episodic Memory Across Species

Prediction (is this an extension of the pattern-filling mechanic?).

Planning (requires capacity for novelty/generativity).

Social role (memory needs to be updatable rather than static if this can get off the ground).

Differences across species (humans have hard-link with language, autonoesis, empathy and theory of mind; other considerations are ecological niche, mating habits, food sources, energy constraints)

Concluding remarks for discussion not pertaining to any paper in particular

My core worries are ethical as well as epistemic. I have mentioned the ethical worry. The epistemic one is that we no longer live in the same environment where this feature was selected for. Is it useful still? Has it become redundant or obsolete via technology and access to information? Additionally, if non-human animals have episodic memory, does this warrant a claim that we have the same sort of episodic memory? Or will there always be a regressive, lack-of-access situation going on?

One other issue I have is that the brain is calorically expensive and any non-necessary component comes at a direct energy expense, and even if that were not the case, episodic memory still seems like a costly adaptation. The most parsimonious answer might be that it just wasn’t worth it for other animals to develop, retain, or pass on this feature.

But to get even remotely close to a more accurate answer means conceding that our sort of memory is partially fallible. We have the kind of semantic memory possessed by non-human animals. This is not controversial at all. However, we possess a soft-soldered mechanism over and above common animal memory that comes with the moral hazard of being manipulable through access. The cost seems justified by allowing us to account for dynamical social interactions in novel settings and facing new challenges that did not feature in our evolutionary background.