The Accuracy Paradox - When Less is More | Overfitting | Data Science

[@jeffspam2791]

Great quote. “When measure becomes a target, it ceases to become a good measure”.

[@CarbonRollerCaco]

It depends on whether the measure itself is reliable to start. Taste alone stops being reliable because when it comes to nutrition, one can certainly have too much of a good thing, and not all healthy things taste good. Plus the chance of not having access to food for a while is near nil for many people nowadays, making overeating largely *rimshot* useless unlike in the old days where food was far from a guarantee.

[@zyzhang1130]

@@CarbonRollerCaco so the reliability of a metric can be context dependent

[@NesrocksGamingVideos]

That quote perfectly explains why netflix originals are so terrible since they're only created by checklists of what has to be included instead of a director with a vision and goal.

[@hollosou335]

5:07 Yes, Jade, yes; you finally get it. Books really are the only things that matters.

[@upandatom]

hear hear

[@post1305]

Here here, unless you’re talking about audiobooks.

[@tobiasbehnke9895]

Im waiting for the song: "I like big books and cannot lie"

[@radaro.9682]

Racial justice, unifying the working class, and fighting systemic denouement of science are other things that matter.

[@noam_segal]

@@radaro.9682 you're not wrong but you're also not funny

[@matthewb2365]

More factors != More data. In all of these instances, you are strictly better off with more data (in fact, as you add more data, you can afford to make the model more flexible).

[@gylorincz]

Problem is deciding when to stop collecting data, since time and cost are limiting factors.

[@kennethferland5579]

Data acquisition has a cost though in and of it's self, plus the risk of over-fitting once you have said data which is driven by the sunk-cost fallacy of the cost you incurred to acquire the data makes you more likely to give the data undue weight.

[@MrTrevortxeartxe]

Literally. More data leads to more accurate models-less data leads to less accurate models. No other way to say it.

[@The268170]

"What a lucky girl" oh brother xD

[@joshwi4193]

I'm definitely guilty of overthinking myself into choice paralysis, but thankfully when it comes to Up and Atom videos, that never happens :D Great video, I'm glad you chose not to overthink it - as long as you're having fun, I'm sure we all are too.

[@upandatom]

haha it's so ironic. and thanks! hopefully I can stick with it (the not overthinking that is)

[@PaulPaulPaulson]

Joe told me to check out your channel and to think about subscribing. You told me to think less. So i subscribed.

[@tusharchilling6886]

This is such an amazing channel. She literally brings all these scientific subjects to human level where we can study these concepts like everyday life stuff. Just wow to the thinker of the concept of this channel!!!!

[@neurotransmissions]

So here's a question: does "big data" increase or decrease the chance of overfitting? I would guess it increases overfitting because of sheer size, but I'd be curious to hear your thoughts. Also, can't forget that floating "e" at the end of female! Lol

[@upandatom]

haha no I thought it was hidden! And yeah I think big data would increase chances of overfitting but I also think more data = more accurate data so obscuring, leaving out or adding noise to the data would be better than just not having a good amount of data in my opinion. All I can say is machine learning has improved drastically since the internet allowed us to accumulate huge amounts of data so there are definitely good things to come from big data :)

[@feynstein1004]

Hang on. Isn't more data a good thing because it shows the trends, if there are any, more clearly? I mean, if you have like, a million data points and there's still no overall trend, doesn't that mean there isn't one in real life? Either that or your methodology is flawed.

[@R.Instro]

I feel that having ALL THE DATA POINTS(tm) allows one to, e.g., use the obscuring trick to make predictions, then "remember" the missing points to check the results. Also, the 9-pt model catches the reality of the ups & downs better than the 2-pt model (if not their precise location on the time axis for every couple), while the 2-pt model catches the overall trend better while missing the reality of the ups & downs. While the former might give an unrealistic trend line, the latter might result in a couple looking at the graph & expecting smooth sailing w/only minor deviations from the norm: even for most "successful" couples, this seems, um, _flawed._ =D Finally, found Up & Atom via Physics Girl, & I'm so very glad I did. Thanks for your hard work! ^_^

[@juzoli]

Neuro Transmissions Big data definitely makes p-hacking easier... It increases the chance of finding random coincidences, which might seem to be something significant, but they are just ... coincidences.

[@firdacz]

Overfitting (in AI / NN) happens with too many neurons, not with lot of data. It is like the number of factors, not number of dots. Too few neurons/factors and you get wrong answers, too many factors/neurons and the algorithm starts to remember every sample, instead of finding trend / features.

[@BANKO007]

Haha! Hilarious! I love the way she crossed out children and companionship as unimportant and spending on books as more important!

[@HoD999x]

the trick is to assign the correct importance to each piece of information before making a decision.

[@stevedoe1630]

Dennis Haupt Display two (2) conclusions, as below, then compare/contrast: 1. Conclusion based on all points have same weightage. (This will make it easier for peer review, for academia and maybe even lawyers.) 2. Conclusion based on all points having a discretionary weightage applied. (This will have more practical application, for the rest of the world.)

[@digitig]

So for each possible weighting of each piece of information, I draw up a table listing the pros and cons of giving it that weighting...

[@Ender7j]

Graph theory is useful for this. Weighting edges appropriately given the problem needing an answer and trying to filter the useful points out helps a lot. God I love school

[@Blau_Flamme]

I have watched many science channels,compared to those, your way of explaining things is absolutely amazing

[@uvbe]

I love your videos but your animation is... Perfect. There's nothing better than a paint character being moved with a mouse. It's so silly, simple and funny...

[@upandatom]

haha glad you enjoyed it!

[@ZoggFromBetelgeuse]

In German, there is a saying "not seeing the forest because of all the trees".

[@adhilAKApeekaboo]

Came here because of Joe Scott's video recommending you.... Stayed here because I love you and your videos! 💓

[@Verrisin]

Great! Now I have completely new angles for how to overthink everything! - Now, I have to figure out which regularization algorithm best fits each problem. - I have to correctly label and value in importance and all other criteria (depending on regularization algorithm) all the data points I have. - I have to account for the extra time I spent doing all of this. - I have to account for all of this being simultaneously wrong, since I'm not supposed to overthink things, but how should I not overthink things correctly? ..................................................... ..................................................... I almost just want to write a list of pros and cons... (but now I know how incorrect that likely is, so I cannot)

[@HowardWimshurst]

So glad Thomas frank showed me this channel!

[@ejkitchen]

Great video! I came here from machine learning trying to get a better intuition for accuracy, recall, precision, F scores etc. At first, I was like, "nope, wrong video" but stayed because my brain was low on dopamine and needed a quick hit. But then you said a few things that clicked a lightbulb in my head regarding some basic concepts I had misunderstood on # of factors. So I got dopamine and useful knowledge! Double whammy. Keep up the great work. Your explanations are very well thought out and presented in an intuitive fashion.

[@utubulatortwo]

I have seen several of your videos and enjoy them very much. I have no sophisticated math background but am fascinated by all things physics. Quantum physics is my favorite area but classical mechanics is fascinating as well. You have an utterly charming way of making understandable the mind bending complexities of how things work . I’m very happy to call myself a fan Keep em comin’

[@michaelthomashayden7125]

As a Patron supporter I love your videos. This is one of my favorites. Go with your instincts because the fun you had making this video is obvious.

[@Defeater33ify]

I’ve been reading that book. One of my computer science professors recommended it. It’s actually really good.

[@MozartJunior22]

Every time I am in a restaurant, and I am debating between 2 or 3 dishes I want to get, I just randomly pick one, close the menu and don't look at it again. I now call this "The Restaurant Method" and I admit I have used it in more serious situations in my life. Though it only works if you have to decide between two good things, and not something that might lead to a bad outcome. The idea is understanding you will be happy anyways, and also realizing that if your choice turns out to be not-so-good, it doesn't mean the other one was better.

[@Seegalgalguntijak]

This is a great video about overthinking, which I (as probably many others, too) also tend to do.

[@kristianbabic9211]

Joe Scott recommended your channel glad he did

[@MrCardeso]

I was gonna get married and take a hot air balloon ride, but then I saw that Jade published a video, so I decided to watch it instead.

[@upandatom]

haha wise decision ;)

[@post1305]

No WiFi at the church ?

[@physicslover4951]

Smort

[@jaimeduncan6167]

This is the best video I have found about this issue. This is the 3th video I saw from you (not count the one with then PG). All of them very good. I sometimes struggle to explain that USA corporations tend to have metrics associated with everything and everyone. Few important metrics improve performance and help avoid subcontinent bias, like sexism against men or women, or how nice the person is or how it likes what the manager likes. But too much metrics lower performance: people start to work for the metrics and security is changed by blind compliance.

[@Salsuero]

I don't overthink when it comes to your videos. I like listening to your voice. I enjoy your humor. So, I watch. Doesn't really matter what the subject is.

[@r1s1112]

It's so obvious that your channel is going to be a huge success, you are doing a great job, you have good content, good quality and good presentation of it, keep it up!

[@upandatom]

thank you so much! n_n

[@bunklypeppz]

This topic reminds me of a study that was mentioned in a book I read a while back. The study was focused on the game of chess and what it is that makes great chess players so good, compared to mediocre players. The study concluded that it was not that great chess players could consider a greater number of options before making a move, but rather they were less likely to consider options which led to unfavorable outcomes. So according to that study, the best chess players tend to consider much fewer possibilities than average players. I think most people assume in a game like chess, that the best players consider the largest numbers possible choices and outcomes.

[@BabelRedeemed]

This reminds me of 'thin slice' judgments that we make about other people who we meet. Certain social science experiments suggest that we judge aspects of character better given less time rather than more.

[@sharktroubles]

The "What is calculus?"animation sequence and affected voice of your brother was priceless! Thank you for your awesome work, lovely Jade.

[@HagenvonEitzen]

2:57 Yes, removing those pesky data points not supporting my desired result has always been helpful ;)

[@lachlanpfeiffer1589]

This reminds me of a machine learning course. In which you use a train test split model selection, to test wether or not your overfitting and adjust how much data you split between both training and testing to not allow for overfitting or under fitting.

[@goodlookingcorpse]

I noticed that thing about measures becoming targets when working in call centers. Because people are measured by how quick their calls are, they make the calls shorter. The easiest way to make calls shorter is to not give the caller the information they need.

[@KeeDono]

Joe Scott sent me here. Thank you for that ;)

[@PunnamarajVinayakTejas]

I'm sorry that letter 'e' floating around at the start is everything I needed to see on a bad day thanks

[@BenjaminCronce]

I now have a new term to overuse from everything to maximizing grades to maximizing objective performance reviews. Empirical metrics are generally correlated with useful results and when people attempt to maximize a metric they destroy the correlation.

[@gorgolyt]

Nice video, but I think some of the things you said will confuse newcomers to the subject. You keep talking about "more information" sometimes being bad, but information really means data, and more data is always better. What you're referring to is *model complexity* being bad, but model complexity isn't really the same thing as "information". Also, I don't see the connection between the tastebuds thing and overfitting. It's a confusing example. If anything that's an example of *underfitting* -- the reason the model didn't generalise well is that it was too simple. Not too complex.

[@Lucky10279]

Good point. What she's really saying is that taking into account too many details can obscure things. More data points might be better, but more detail might not be. I think about the data like a forest. A bigger forest shows the trends better, but only if we zoom out enough. Basically, we don't want to miss the forest for the trees. I use that phrase a lot when explaining linear regression to statistics students -- I make the point that we _could_ always just find a polynomial that perfectly fits the data set and it's not very hard if we use a computer, but it's often better to use regression instead both because it's easier to work with linear functions and because they're often better at showing a general trend. (Of course, we often need different kinds of non-linear regression as well, but the course I tutor only covers linear regression).

[@christianosminroden7878]

Mikayla Eckel Cifrese Polynomials are pretty much the worst approach you can choose if your goal is to extrapolate predictions beyond the range of a given data set; once the curve leaves the range of known data, it‘ll go off into outer space with no connection to anything at all. Obviously, it’s always best if you know what general kind of curve the phenomenon you’re describing should be showing and go with that, but if you don‘t have any idea about the underlying workings and have to rely solely on empirical data points, you’ll rarely find cases in which the best estimate won‘t be either linear, exponential (or logarithmic for that matter) or logistic. Try these and cautiously (!) go with the one resulting in the smallest range of relative errors. Polynomials are rarely more than a funny toy to play around with for a bit in this context.

[@nathanborak2172]

Yes, she gets it completely backwards, almost saying you should throw out data. Really, you should throw out parameters in your fit, not data. Don't fit to at + bt^2 + ct^3 + .... . Fit to at, or maybe at+ bt^2. Get rid of c, d, e ... Videos like this make me pretty mad because they get things very wrong but talk as if they are an authority. A basic understanding of math would show you instantly what the problem is. If I have 9 parameters and 9 data points, I can always fit the function to every point exactly because I'll have 9 equations and 9 unknowns. It's not some deep mystery and she gets it completely wrong.

[@MrEmlish]

This the video I have been looking for in ages. I am used to overthinking stuff and delaying decisions by trying to look for more information. I usually don't comment on KZbin but this video worth the effort. Thank you =====> Subscribed, Liked and Shared.

[@sebastianmorataboada9795]

"I bet you the whell was invented by someone overthinking 'pushing' " -Randall Monroe

[@dm9910]

Intuitively it makes sense to me given the Pareto principle: 80% of the effects tend to come from 20% of the causes. So if you have a model consisting of 5 equally important variables, you've probably done something very wrong. Another thing to think about is the Fermi estimate which demonstrates that you can often get reasonably accurate predictions with very little data, purely based on the idea that over- and under-estimates tend to cancel out. Similarly here, the factors you don't consider are likely to cancel out to some extent, so adding in those factors to your model might provide a smaller increase in accuracy than intuitively expected.

[@Roboterize]

Perfect to SciShow Statistics.

[@pearhams2]

Proper weights for the data points are needed. More data points brings an imbalance of lesser weights over the larger ones skewing the proper understanding of the trend. Also, as you think of more data points the weights of previous data points can change because they may have interactions with each other as well.

[@lorenzomaglio176]

You'll probably never read this comment. But I'm a controller, and the biggest part of my job is manipulating data. You can't even imagine how many times I've been asked to produce a forecast, then asked to redo it by adding x more parameters. Given that a forecast is by definition random + math, iI had already realised that the more info in, the crazier the output. Explaining it to my bosses has always been a particular type of masochism and I just dropped trying and I delivered the craziest output ever. And trust me when I say that the ultraviolet catastrophe was a joke if compared. But I used the data they wanted and they were happy.

[@fran6b]

Very interesting. I have the impression that overthinking is not good when come the time to make a decision, but thinking a lot, i.e. thinking a subject in a lot of details, can be a good think to reach a better understanding of how things works. Athletes could be a good exemple of that. To get at the top, they have to «overthinking» their sport. At elite level, details make a huge difference. But when the time comes to perform, overthinking become their enemy, so they have to keep things as simple as they can.

[@fran6b]

Davin Green Thanks for your insight :)

[@VladTepesh409]

Great video. I'm going into data analysys field, so this is really good to take into account. Thanks!

[@wmherndon]

Long time ago I had a program on my Atari Computer called decision Maker. To make a decision you would input 10 categories you then you would weigh them with a score of 1-10 , then how important from 1-10. Example ( should I date this girl) category 1.. Personalty 2. Eye color 3. Weight appropriate..etc.. I remember being very surprised at the result. It made me think more deeply about any decision I would make. It helped me to know myself.

[@TheErgunPascu]

Probably one of the most insightful and pragmatic of videos. Gracias! 👍

[@housellama]

"When a measure becomes a target, it ceases to be a good measure". There are so many problems in the educational and business sectors caused by not understanding this.

[@roquepardal]

LOL I love how angry jade got to be in the outro!!! I rewatched the sibling rage bit laughing way harder than i should

[@thatonecommentor7758]

I never knew that bit about why we find some foods more appealing than others. I know that wasn't the key point of the video, but still, thank you for sharing!

[@GroovingPict]

I think a real life AI example of this is Akinator. When I tried it a few years ago, it was scarily amazing at guessing. Now it takes forever to guess even the most obvious of persons (like, Einstein if youre trying how it does on physicists for example). It seems to me that as it has been trained more and more through people using it, the noise to signal ration has worsened, and the more information/training has resulted in a worse AI

[@lexmtaylor]

The benefit of pro can lists is not how many pros or cons but the exercise is to think critically about the effects of the decision. If you have fifty pros and one con it’s still maybe correct to avoid a choice if the one con is particularly onerous. But by listing many pros and cons you are less likely to be surprised by a pro or con.

[@goodlookingcorpse]

At the risk of being 'that guy', 'quod erat demonstrandum' is Latin for 'that which was to have been demonstrated'.

[@shoryaagarwal561]

These videos seriously need more upvotes 🤓!!

[@violet_broregarde]

Nice real-world examples of overfitting. Especially when you went into biology. Taste buds were a really clever analogy. Big fan :D

[@Jopie65]

Yessss this! I always fall into the trap of thinking too much. I read somewhere that people are generally happier with an impulsive purchase than with a well considered one. So this is now even mathematically proven 👍😄

[@Posesso]

Fantastic. I think it is great advice to make hard things easier, like going for a run. Also, it suggests that one should not constantly monitor oneself and make conclusions. Thanks!

[@ctso74]

Nice Golden Spiral/Triangle! And, I particular liked the "congestive heart failure" ending.

[@upandatom]

thank you! :)))

[@theyxaj]

I can't believe there's fewer than 6k subscribers!

[@eng.minanagynasr]

most people became idiots wanna eat, sleep, and f**k and nothing more ...

[@feynstein1004]

+Mina Nagy Lol why so salty? Every channel starts out with a few subscribers. You gotta be patient.

[@harrycartwright6666]

A month later and subscriber count has increased by 250%!

[@BrutalistJr]

18k now

[@bludshock]

More than 32k Now!! In just 5 months!! Congrats :)

[@jenniferarmstrong8879]

Overfitting isn’t a problem of having too much data. In machine learning, it is a problem of having too many variables in comparison to the available data. As such a the model can just memorize the data to guess perfectly every time. A better example for math class would be if the teacher only every tested you on a few questions. So for the test you could copy down the answer from memory without having a good understanding of the underlying principles.

[@bfortner100]

You have good taste in books. Steven Johnson's book called "Farsighted" includes the pros and cons list, along with other great ideas. "Algorithms To Live By" was a lucky find - a favorite of mine. Keep up the great videos.

[@mrsteele1781]

My job is in data analytics, and you are interchanging "data point" with "data feature" -- these are two different things. It is true that you can have too many data features or features that show no real correlation. But with *data points*, the real danger is that it becomes difficult to process (i.e. during ETL) -- NOT that it causes overfitting. Overfitting is usually a result of the chosen feature set and the machine learning model that is used. If you choose the correct features, the correct model, and the proper fine-tuning parameters, you can largely avoid overfitting regardless of the number of data points.

[@douglascampbell66]

My new favorite KZbin Channel...physics with charm.

[@fredrickemp7242]

I was not thinking about any of this and now I can’t stop. Do Another video on how to get this one out of my head.

[@martypoll]

Over fitting can be a problem but the main problem I see in your examples is extrapolating too far beyond the available data. How far is too far? That depends the decision you plan to make using the extrapolated region of the model. Note also: fitting data only leads to a model.

[@deebur8777]

That was a concerningly relatable list

[@erictaylor5462]

0:20 why is the "e" in female moving around? Did you misspell "female" and "fix" it in post?

[@cr10001]

73 was a pretty good age in Darwin's time. I've seen some of those curve-fitting exercises and they're not infallible. We had consultants prepare a guide to estimating costs of future projects - they took all our past construction projects for the last few years and the idea was to produce 'unit rates' that engineers could use for future budgeting. Trouble was, past projects varied enough in their individual details and circumstances that comparing them was misleading. The consultants did their best, but it became painfully apparent that they'd just plotted a lot of disparate points and used Excel to best-fit a curve through them, and some of the curves were hyperbolae, some were exponential, some were straight lines, some parabolic - there was no logic to it at all.

[@stevedoe1630]

I often attribute overthinking to taking formal education too seriously. One of my favorite quotes on the matter: “The more formal education you receive, the more common sense you lose.”

[@patbak235]

i would say "the less common sense you have, the easier it is to lose." less to do with formal education more to do with your inability to think for yourself.

[@Simulera]

A specific case of diminishing returns, in this case diminishing predictive value of added parameters for a particular (class of) fitting technique. Over-fitting can have certain analytical use value as well. And….. have you done a video on Simpson’s paradox? I think you may have, but in any case, 5 years on, this video is very good and is current-event timely since massive adaptive data fitting of various kinds are driving the current and latest “AI revolution”. Plus the meta problem of when to stop editing this video is perfect.

[@dracopalidine]

It's like the coastline paradox where the smaller the unit of measurement you use the longer the coastline appears to be.

[@mossbanksy]

Excellent video, Jade.

[@crusiethmaximuss]

4:34 now this just made absolute sense.

[@sirtanon1]

Is that a Star Trek pendant? Awesome. Yet another reason to love you.

[@gearhead1302]

I just love her animations! So funny. I really like this channel.

[@matthieujoly424]

When you've got children / a family AND overthinking, is, that you have to take decisions... quite quickly sometime, but still.. thinking after haven taking them. Pure nightmare, allover again.

[@ultradude5410]

Great, now when I’m overthinking something I’ll start to overthink whether or not I’m overthinking

[@xandercorp6175]

That floating "e" though... I want to give her a hug and smile.

[@salsa123freak567]

First: keep up the good work! Second: Too much information and unnecessary model complexity are two different things. I think you could have picked a better example. There are cases where more data doesn't help (much). Instead, less complex models reach better quality on unseen data. Just wanted to make this rather technical point. I think you understand. Third: Keep up the good work! Enjoying your vids :)

[@frenstcht]

Thanks for the book recommendation!

[@carryon2197]

Complimenting shouldn't be overthinking. My compliments. You are pretty.

[@nicolaiveliki1409]

Reminds me of path integrals of QFT where the weirder paths cancel each other out and contribute less to the overall outcome

[@nibblrrr7124]

2:53 This sounds like conflating model complexity and sample size. More data is always* better; in fact, a bigger dataset allows you to use more complex models without overfitting. If you don't use all the data you have for training/fitting the model, it's usually because you hold out some for testing how good it actually performs on new data. * Assuming they are of the same kind, sampled in the same way, from a large population... Obviously, crappy data in -> crappy model out :P

[@upandatom]

hey thanks for the correction :)

[@Skeithization]

Very enlightening, it's certainly a new perspective to look at things. :)

[@DarcyCowan]

Sudden deep depression after ten years and infinitely miserable sounds like marriage to me. Lasted 16 years.

[@slobodanreka1088]

Wife: "You left the fridge door open again." Me: "Stop putting so much emphasis on each individual data point."

[@thinkpiece4334]

thanks for this! been meaning to read that book and i think this was helpful

[@briancherry8088]

Withholding data didn't work too well for the Challenger shuttle. When discussing the likelyhood of temperature affecting shuttle launches, a graph was created showing the failures for the O-rings, and at what the external temperature was. This showed that failures happened on cold days, as well as warm days, so the temperature must not be a factor. However, when data was added to include ALL flights (both successes and failures), it showed that not only were most warm weather launches a success - it showed none of the cold weather ones were.

[@Toastmaster_5000]

I would argue that less data for a clearer plot line is more paradoxical. Assuming the data isn't faulty, the more you get, the closer you are to the truth. If you can't get a truth you "like", what's to say there's supposed to be one? Take for example flipping a perfectly balanced coin: it's a 50% chance you'll get heads. Well, let's say you only flip the coin 6 times, and 5 of those times you got tails (which is totally possible). If that's where you decide to stop, that means you would determine heads can only be achieved 17% of the time, which just simply isn't true. Keep flipping the coin and the more data you collect, the closer to 50% you'll become. 50% is the truth, and the only way to actually get it is to collect an a _lot_ of data. For more complex things (like the marriage happiness example), if too much data isn't giving you a "satisfactory" trend line, you might be collecting the wrong data, or parsing it wrong.

[@JimyLindner]

Answers with Joe brought me here! Great channel! Keep up the good work!

[@upandatom]

thank you and welcome!

[@tesseraph]

When you presented ways overfitting relates to life, I felt like some of the connections were a bit contrived and "overfitting" started to just sound like a synonym for "being wrong". Other than that, great topic and great video about it!

[@upandatom]

thanks :)

[@nibblrrr7124]

I mean on some level overfitting is analogous to "missing the forest (real model) for the trees (data points)". ^^ I agree though. (Also, if you stretch that analogy & want to make sure you don't overfit, you can use the null model: everything is just random fluctuations, nothing to be learned here! :D)

[@KuraIthys]

You know, one of the weirder things I came across (which was incidentally speculation by a physicist, though admittedly not based on anything substantial) was that everything is entirely random and the apparent order of the universe is a side effect of how we think and remember things. If you imagine a mind that remembers things that fit a pattern and throws out stuff that doesn't seem to fit a pattern... What do you think that mind's model of an endless stream of totally random events would look like? Overfitting taken to it's ultimate extreme - thinking a pattern exists where there is nothing at all except randomness. Quite the idea... XD

[@SirKaracho77]

Someone should present that to the movie industry. I get that production companies want to follow trends, but pasting actual plot points into a new script usually makes it worse.

[@ahmedalmerza1707]

I'm binge-watching your channel rn

[@Playzon]

I imagine that weighting the categories would fix part of the problem.

[@Vanished_Mostly]

0:13 Wait, did you forget the letter E and just add it in post, but forget to compensate for camera movement?

[@macsnafu]

They married in 1839, and Darwin didn't die until 1882, so they were "happily married" for more than 40 years.

[@maythesciencebewithyou]

And he was very close to his wife, and loved his children. So crossing out those things wouldn't be his priority.

[@nastysimon]

Some of this reminds me of my own information bias psychological problems

[@divyamkumar3877]

I love your channel !! Keep it up !

[@RichardOles]

I just discovered your channel. I'm an Instant fan! Very enlightening, and you're adorable. Lol. Much luck to you and I look forward to your future videos.

[@monishnasa9880]

i really watch ur videos with a passion ....and i mean it