In a waking day of 16 hours, you are conscious for 960 minutes.  If a “memory event” is something in the neighborhood of 10s, that’s something like 5000 opportunities for a memory per day.  How many memories do you have from yesterday?  If It’s in the neighborhood of 100, that’s a memory storage fraction in the range of 1-2% of experience.  The experience of remembering episodes often feels somewhat sequential in that we can remember what happened after something else we remembered.  So possibly the storage fraction is higher and we’ll need a better model of the retrieval process.  But whatever the fraction is, it reflects some small slice of our experience.

The simplest possible model of memory storage (encoding) is a simple stochastic function with a probability of occurring of something like 1:50 or 1:100 (1-2%) for each event you experience.  That is, you are experiencing events in a relatively constant sequence and the specific neurobiological processes of the MTL that create a new memory trace/item occur stochastically through your day, laying down a random subset of memories.

What falsifies this model?

The ability to “study” and remember somethings better does not necessarily counter the model.  One basic element of study is to direct attention to the desired material and repeatedly engage with it.  A purely stochastic model will end up with a greater fraction of studied memories simply as a result of this repetition/attention process.

Emotional memory, especially negative emotion, does indicate that memory creation is not purely stochastic as those memories associated with negative experience are more powerful and more long-lasting.  However, the effects of positive experience are a lot more subtle.

Are there any other neurocognitive processes that affect the storage rate/fraction?

One thing wrong with the purely stochastic model is that it doesn’t allow for strong/weak memories, or the process of decay.  We generally use the term “consolidation” to refer to the temporally extended process of memory change that causes memories to generally become weaker (on an exponential/power-law curve) but some memories to become durable and permanent (and non-MTL dependent).

Where is the stochastic process occurring?

A simple neurocognitive model of memory is that information comes into awareness through perception, modified by attention, held briefly in working memory and then becomes a MTL-dependent representation that can become a long-term, durable memory.  Since awareness/working memory seems to be constant, the transition to the MTL is where the probabilistic nature enters.

It’s difficult to tell directly if the failure point is the transfer from WM to LTM, or if everything is at least briefly in the LTM but then fails to consolidate.  The available indirect evidence slightly favors the latter.  The availability of information that is extremely recent, but supra-span in WM suggests that representations are available somewhere (e.g., the MTL) to re-enter WM.  Replay findings in animals that show re-creation of activity patterns in the MTL that mirror experience also argue for experience traces to all be present at some point in the MTL.  Default mode activity activating the MTL fairly reliably is additional weak evidence in this direction.

We can refine the simple stochastic model to one where everything is at least briefly reflected in MTL activity.  Perceptual representations are held transiently in WM circuits, which are then echoed in MTL activity patterns.  We therefore presume that a neurobiological process stochastically occurs that triggers a consolidation process for some small fraction (1-2%) of these representations, some smaller fraction of which then go on to be durable, very long-term memories.

Other than attention/repetition to increase the probability of specific information getting into the consolidation pipeline, what other factors might increase this probability?

The ability to hook onto or connect with existing memories might increase the probability of successful consolidation.

Spacing/reactivation of existing memories might influence this process as well — although since no two experiences are exactly alike, this may effectively be the same thing as connection to existing memories at a neural level.

We should also note that some direct reactivation processes are known to be maladaptive, e.g., remembering where you parked today needs to be separated from prior parking events.  In theory, there’s a pattern separation process trying to help establish that, which would have to be operating in tension with a connection/reactivation process.

The CLS model captures the pattern separation/consolidation tension by positing that memories are initially separated via a sparse-firing pattern mechanism and then a temporally extended, gradual consolidation process incorporates these back into an integrated long-term memory.  We have noted that the sparse representation mechanism is likely to be low-capacity, potentially creating that bottleneck that causes the 1-2% storage rate.  The eventual distributed representations in the LTM store probably have larger capacity than can be filled in a lifetime, but does not contain all of our memories due to the bottleneck, which then causes the effective stochastic/probabilistic nature of what small fraction of memories get created.

A problem with this model is that the neural patterns that are initially sparse and eventually distributed are very different, making it hard to see how prior knowledge could affect novel storage events.  How do we think reactivation, connection and association of memories works when there are completely different neural representations of knowledge?

Being able to answer that question might theoretically suggest approaches for engaging with information to be memorized (learned) that would be particularly applicable to education or training, where the goal is to acquire specified content sets of related information.  It might also provide some insight into successes or failures of everyday memory, which do not actually feel like they are totally stochastic.

In an article titled “Georgetown Law professor terminated after ‘reprehensible’ comments about Black students,” we hear about a professor who is being fired from Georgetown Law after some inappropriate remarks made on a recorded classroom after class.  Predictably, this is triggering some pushback from people overly concerned about “cancel culture” because of the nature of the quote published in that newspaper report:

“I hate to say this. I end up having this angst every semester that a lot of my lower ones are Blacks,” Sellers said in the video. “Happens almost every semester. And it’s like, ‘Oh, come on.’ You get some really good ones, but there are also usually some that are just plain at the bottom. It drives me crazy.”

It’s a pretty good example of implicit bias and a professor who would probably gain something from some education about education and challenges of implicit bias in the classroom.  It’s also pretty much impossible that this is the sole reason for the professor being fired (it seems she was quitting/resigning when placed on leave so there had to be some more things going on).

For comparison, I had a conversation in our department with Prof. Jen Richeson (before she moved to Yale) who studies race, bias and cognitive consequences that included the question “is there an achievement gap at Northwestern?”  If you know the research, you’d know the term ‘achievement gap’ reflects puzzling results where URM students underperform unexpectedly.  That’s kind of what the Georgetown Law professor is observing and the issues is not the fact that it exists, but what you think should be done about it.

When I was talking to Jen, our context was along the lines of ‘what are we doing wrong as educators that allows this to persist?’ as part of trying to identify how to reduce implicit bias effects in the classroom and/or if there were broader social or cultural effects on these students at the University that could be improved.  Jen’s work includes seminal studies showing cognitive costs to things that I would describe as implicit bias (e.g., reductions in cognitive control associated with worrying about cross-race interactions).  I think Jen’s idea is likely right and the practical question is how to ameliorate this.  For me, it prompts self-reflection about classroom and other university interactions where I might be accidentally reflecting implicit bias and what can I do to continue to make progress on reducing that.

The problem with the out-of-context quote from Georgetown is that in many other cases, these kinds of observations are followed by more seriously inappropriate statements, e.g., repeating incorrect claims about innate ability of URM students or an attack on diversity measures in recruiting.  Noting that your URM students are struggling isn’t a problem.  Arguing that they should be excluded is a really big problem, a fireable problem.

When Jen was here at NU, I used to say to her a lot that people really need to think more about the fact that implicit bias is implicit.  If you’ve worked a long time in a field that needs to improve its racial and gender diversity, it will be the case that the younger generation will not look like the older generation.  Because that does not fit with your implicitly acquired expectations, it may feel weird to you and you won’t know why.  However, it’s still your obligation to work through that, get used to it and get past any mistakes it causes in your interactions.  Her response, fwiw, was that I should be going around saying that more, not her (she does enough public and education work in this area already).  I don’t get asked to go on Oprah to talk about it, though, I just write things like this blog post.

And on the general topic of “cancel culture,” I’ll repeat what I’ve said here and elsewhere that I continue to be in favor of it.  There are a lot more people not getting canceled who should be than people who are canceled inappropriately.  When that balance shifts (a whole lot more), I’ll officially worry about it.  Being completely, perfectly fair is an extremely hard thing to do.  If you think a professor losing her job is a bigger deal than the students whose careers are damaged, you might need to do a little self-reflection on implicit bias yourself.

Chess’s recent surge in popularity, plus the Netflix show “Queens Gambit” featuring a female chess prodigy have unsurprisingly inspired a flurry of new conversations into the old question of differences in chess ability for men versus women.

Prof Wei Ji Ma (NYU, Psychology) had a few recent pieces arguing that the over-representation of men at the top level merely statistically reflect the larger number of men than women who play chess.  A recent piece published in Slate (https://slate.com/technology/2020/12/why-are-the-best-chess-players-men.html) which summarized an analysis posted at Chessbase.com (https://en.chessbase.com/post/what-gender-gap-in-chess).  These focus on chess players in India and largely replicate a prior study of chess players in Germany (Bilalic, Smallbone, McLeod & Gobet, 2008).

Both studies are fairly straightforward and follow from the basic idea that if you sample more often from a normal distribution, you’ll get more values in the extreme.  More men play chess than women, so you get higher top ‘scores’ for men.  Note that this is related to but different from the “fat tails” hypothesis since this simple fact occurs even if both groups are drawn from distributions with identical means and variance (the “fat tails” idea is that the distribution of men has higher variance).  His conclusion is that the difference among top players is completely predicted from the differences in the population size.

However, two other analyses suggest this approach does not account for differences in other countries.  Jose Camacho Collados (https://josecamachocollados.medium.com/the-gender-gap-in-top-level-chess-15591d8990ba) argues that the observed gap is bigger in many countries than should be by simply population sampling.  Nikos Bosse (https://followtheargument.org/gender-differences-among-top-performers-in-chess) makes a similar point.

The approach and ideas are familiar and serve as a good example that every one of these analyses suffers from two major flaws.  First, the statistical sampling approaches do not take into account that they are sampling performance data from a restricted and high performing range.  Second, even if something about chess ability is innate, chess is a learned skill and you cannot draw any real conclusions about performance data without considering influences on the learning process.

The first, sampling point is fairly simple math, actually curiously impactful, and I have never seen it fully considered before in distribution studies.  All the analyses of chess players estimate the distribution of performance (as a proxy for ability) from rated players — which is one of the reason people use chess as an expertise model, since the rating system provides a nicely quantifiable metric of current ability.  However, only the relatively better chess players have a rating, so rated chess players is not an effective representation of the distribution of chess ability in the population.  The implication of this is that this restriction leads to underestimating the population variance and that really throws of your estimate of the tails.  It’s an effect of easily 2x or 3x of your estimate of the expected number of people in the tails.

So you can’t really look at the rated chess population and infer anything about there should be X men/women at the top unless you have an unbiased estimate of the variance in the broader population, which you don’t get from a sample that overrepresents experts.  This seems like a pretty basic sample/population issue and it’s almost weird that some of the people making this mistake are trained statisticians who should know better.

The second point is much more commonly made. It’s also pretty clear that chess rating (ELO) is quite strongly affected by learning and experience and this will be dramatically impacted by societal factors that discourage women from pursuing chess (or limit opportunities or access to high level training).  One of my minor quibbles with the Queens Gambit is that Beth Harmon doesn’t seem to need to actually learn any chess.  Even Bobby Fisher lost to top players as he moved up to playing professionals.  Maybe the point of the scriptwriters was to show that the only way a woman could crack the top echelons of chess would be to be extraordinarily talented from the beginning.  But if so, maybe they should have made that point clearly enough for people to see it.

Over the holidays, I discovered a new chess competition variant being played and streamed by Chess.com: Puzzle Battle.  Chess puzzles are a well-known training device that also have something of a micro sub-community within the chess world for artistry in creating positions with a difficult to spot but winning move.  They are highly recommended to people trying to improve at chess.

A new approach to this idea has popped up on all the various online chess sites where the idea is to speed-solve these puzzles, usually ones that aren’t as hard as the original “chess studies.”  These get called things like ‘puzzle rush,’ ‘tactics trainer’ or the latest: ‘Puzzle Battle.’  In a battle, two top-level players solve as many puzzles as possible in 3m or until they make 3 mistakes where mistake = incorrect move in the winning line.

The top performers are routinely hitting 50 puzzles solved in 3m.  This has to be seen to be believed, e.g., in the linked clip from the quarter finals of a current tournament being run now.  In fact, if you aren’t very, very experienced in chess, there’s a very good chance you’ll have no idea what is going on.  If so, slow the video way down…

What is happening is that they are getting a new chess position with up to 30 pieces on the board.  Somewhere is a winning move for the side to move.  The position must be parsed, analyzed, and the correct winning move selected and implemented.  Sometimes the puzzle requires a sequence of 3-5 accurate moves to prove the win.  The best solvers are doing 50 of these in a row at a rate of 3.6s each.

At least for me, this pushed my understanding on the upper bound of human cognitive performance out another step or two.

Potentially of note, in this linked battle (spoiler alert), an untitled puzzle specialist is beating one of the top-10 GMs in the world.  Apparently, the database of puzzles on Chess.com only has about 20,000 examples and some of the competitors have memorized a significant fraction of the db. The super-GM is certainly better at chess than his opponent, but is solving more of the puzzles on the fly so he’s tending to score more in the 40s.  Then there’s the question of how one memorizes the winning line in 20k chess puzzles so as to be found, retrieved and executed in 3.6s.

Hello,
We are working on a group speech for a school project and need to reach out to an expert in the field of psychology. We have a few questions about the topic of brain capacity and we would greatly appreciate it if you would take a look at them and get back to us.
Here are our questions:
Does exercise affect how well the brain functions?
Can certain exercises be done to improve brain capacity? If so, what are some examples?
How much information can the brain hold?
Where did the myth of only 10 percent brain capacity come from?
What general methods do you think work best to fix brain injuries?

Thank you!

Today the above email arrived and reflects questions that come up fairly often to me as a memory researcher.  I decided to answer them here on the blog for availability for future reference (I might have actually done this before, at some point I should look and collect similar posts).

General answers:

• Physical exercise probably does not have much immediate impact on brain function other than being fatigued after exercising might very temporarily slow function.  Longer-term, cardio-vascular fitness appears to be important in healthy aging.  Maintaining physical fitness through middle age and late life looks to be very helpful in keeping your brain working well.
• Cognitive exercises also seem to help healthy aging.  It is less clear how well cognitive exercise helps younger people.  Most things that keep you cognitively active result in learning and for younger people and the information learned is probably the most important thing.  Being cognitively active mainly means doing interesting things that make you think.  Staying active when you are older probably helps with something like ‘capacity’ while doing this when you are young makes you smarter in a more direct way.
• The brain appears to have the capacity to hold everything you are capable of learning over the course of your life.  It doesn’t appear to work that way — we forget things a lot — but this is due to problems storing everything permanently more than running out of room to store them.  Storing memories is slower than experiencing things, so a lot of your experiences don’t end up in your long-term memory.
• The brain is made of neurons.  Neurons are electrically active when they are functioning.  You definitely don’t want them all firing at the same time — that would cause a seizure, as in epilepsy.  If the 10% idea has any basis at all, it’s related to about how many neurons could be firing at about the same time.  There is no hidden, reserve unused capacity in your brain.
• The major problem with brain injuries is that neurons don’t grow back.  Almost all treatment and rehabilitation is training new processes to work with the remaining uninjured parts of the brain, and learning to work around the damaged parts.  This can work surprisingly well to recover function, but you can never really recover or replace lost brain tissue, unfortunately.

On the theme of demonstrations of exceptional skills via youtube, I recently ran across the channel of the LockPickingLawyer (https://www.youtube.com/channel/UCm9K6rby98W8JigLoZOh6FQ/about). He posts videos of picking various kinds of locks together with evaluations on how effective the locks are as security devices. I found this to be highly interesting for a variety of reasons.

First, this seems to be a wonderful example of a highly implicit skill. The mechanical interaction between the tool and the internal elements of the lock cannot even be seen. You could try to explain to me how to do this, but there’s absolutely no question you’d need lots of practice to carry out the procedure successfully. And yet, people obviously not only learn the skill, but get very, very good at it.

Second, this skill is even more pointless than learning to yo-yo (or speed solve a Rubik’s cube).  Locks are peculiar security devices in that they are a minor deterrent at most to actual criminals.  In most circumstances there’s a brute force way around a lock (bolt cutter, break a window) if somebody is determined to break in.  Probably the mostly likely case of somebody picking a lock is a locksmith helping you with a door when you’ve lost or misplaced the key.  And locksmith’s have access to tools that make the perceptual-motor skill relatively less critical.

But if you read the comments on the Lock Picking Lawyer’s videos, you’ll quickly discover this is a hobby that seems to have a reasonably sized interest base.  It appears to be called Lock Sport (http://locklab.com/) where people compete on speed or challenge themselves with increasing difficulty in a way reminiscent of puzzle-solvers (there’s a robust puzzle solving community on youtube as well, but puzzle solving seems like a very explicit process).

I’ve never met anybody who is into this — that I know off.  But if I was picking locks for fun, I don’t think I’d talk about it with people outside the community all that much.  People would likely think you were some kind of aspiring criminal.

Which makes it a great example of a skill some people get really good at, that takes many, many hours of practice and has no particular external value in achieving.

So why do people get good at it?  I can hazard a couple of guesses. In the video comments,some people report the process of practicing to be calming in a way that is reminiscent of ‘flow’ states, which we have thought might be related to dopamine.  Relatedly, the process of picking a lock probably produces a real substantial RPE (reward prediction error) feeling where you struggle with the task for a long time, then suddenly get an unexpected payoff of success.

Honestly, it looks like it might be a fun thing to learn.  But I think I’m not going to go buy tools and try it because I don’t want people to judge me.

Many people have emailed me the following comic:

https://www.smbc-comics.com/comic/conscious

Yes, I appreciate it.  No, I didn’t have anything to do with the author/artist.

Happy New Year!

P.S. I wish I could embed it more directly here, but I don’t want to deprive the author of links/traffic. It’s about implicit memory.