Using Bayesian Approaches & Sausage Plots to Improve Machine Learning

Using Bayesian Approaches & Sausage Plots to Improve Machine Learning - Computerphile

Рет қаралды 37,895

Күн бұрын

Пікірлер: 237

@paulmitchell2916 Ай бұрын

Playing "mastermind" when I was 12, I was astonished to see the decoder make a guess that couldn't possibly be right, then solve the puzzle on the next guess.

@andrewharrison8436 Ай бұрын

That's a nice algorithm in the decoder. The "best" guess is the one where each answer narrows down the possibilities the most. That's a lovely conceptual leap but then implementing in code is an interesting problem.

@rmsgrey Ай бұрын

Yeah, while guessing possible combinations gives you a chance of being right on each guess, so optimises your best case number of guesses, if you want to minimise your expected number of guesses, or your worst-case number of guesses, you'll sometimes want to guess something that can't possibly be right in order to best separate the remaining possibilities. For example, if you guess ABCC and get one black, then guess ADEF and get two blacks and two whites, you then know that the answer has to be one of ADFE, AFED or AEDF. You could pick one to guess, getting either four blacks, or one black and three whites, then guess a second, and finally guess the third if needed, with equal chances of taking 3, 4 or 5 guesses to get the four blacks. Or you could guess ADEE, which would give three blacks, two blacks one white, or one black two whites respectively for the three possible answers, and guarantee taking 4 guesses to get the four blacks. Same expected value, worse best case, and better worst case. I make no claim that those first two moves are in any way optimal, by the way - they're just a way to reach a situation where there are three remaining cases. If you have four remaining cases, where incorrectly guessing one tells you nothing about the others, then making a single impossible guess that distinguishes all four would improve your expected number of guesses too. With fourteen possible responses to a given guess, you can save a lot of guesses with one impossible guess, though coming up with scenarios where you have precisely n different possibilities and can distinguish them with the results of a single guess is challenging and probably impossible for larger n.

@paulmitchell2916 Ай бұрын

@@andrewharrison8436 I didn't understand this presenter at first. I thought he was approximating the whole function.. In that case taking your next observation at the widest point of the max area "sausage" seems best, analogous to the mastermind case. But I guess his task was to find a global min.. which is more likely to be near an already discovered min. How much more likely? That would seem to depend on the variability of the data.. Did he say what his "acquisition function" is? If he did, I didn't understand it.

@Xeridanus Ай бұрын

@@rmsgrey Not impossible. Wordle works on the same principle, but it has a limited space of guesses. Not sure if it's exactly 1:1 comparable because of that and if it isn't, I have no instinct for which would be a more difficult set of choices. The Wordle word list would have embedded information that could be used to help solve it though. I believe Wordle solvers use graph theory in trying to find the optimal path to a solution.

@Xeridanus Ай бұрын

@@paulmitchell2916 Your reply seems more relevant to the video at hand and not the comment you replied to.

@Zahlenteufel1 Ай бұрын

I would appreciate more in-depth videos about all sorts of Bayesian approaches, particularly the neural variety.

@BenitoAndito Ай бұрын

The lower uncertainty between close points sounds intuitive, but it becomes a challenge to define how close observations are, especially in high dimensional space such as in the biology space, like genomics and proteomics. I'd like to pick his brain about applying Bayesian methods in these cases.

@nightmix Ай бұрын

Interesting comment!

@jaiveersingh5538 Ай бұрын

Curse of dimensionality strikes again

@Juan-qv5nc Ай бұрын

When I stumble across these kinds of stuff I try different metrics with an example of which I know the outcome. Then I choose the metric that suits better in such a context, or invent a new measurement system. Sometimes suitable measures don't satisfy triangle inequality.

@petergerdes1094 Ай бұрын

The whole problem is presuming some given background probability measure on the functions being minimized and it flows from that. I mean if it was literally an arbitrary function from R to R you couldn't say anything and even for continuous ones you probably want to assume some bound on variation (if all your measure is concentrated in high frequency functions nearby points don't tell you much)

@D1ndo Ай бұрын

I feel like this is really a question of extracting features from your dataset. Its inherent to all machine learning, and it's more of an orthogonal problem to what Bayesian is trying to solve. For example, when using SVM (Support Vector Machines), there are various kernels available to transform the data features into a hyperdimensional plane, where a nice curve fitting algorithm searches for the best fit. Depending on the kernel, you might get different shapes of the function, but the feature extraction is a separate step done beforehand. If your inputs are whole genomes, you might end up with super large high-dimension search space, which is not really efficient I guess. So you have to reduce the feature set down through a function that is highly dependent on what the data actually represents, and this is where expertise in a given field (i.e. biology) is priceless. In other words proper feature extraction and feature space search optimization are two completely different topics, related to two different stages of training.

@wmoani-2606 Ай бұрын

Thank you, Richard Hendriks

@parhwy Ай бұрын

Thank you for this video. I first encountered Bayesian modelling when I was watching KZbins with Dr Richard Carrier who accompanied it with his review of ancient history and historicity. I never quite got it but your description of those graphs and each joints veracity/accuracy was solid in my brain. Thanks!

@YPOC Ай бұрын

Ever since I've started learning about Machine Learning I've been intrigued by Gaussian Processes, based on the Bayesian approach. However I never got around to studying more complex applications than what is depicted here, a simple 1D-prediction. Surely more dimsensions are possible, but I'd love to see another video about real world implementations and use cases.

@YouTube_User_ab Ай бұрын

Hey, can you tell me where can I learn machine learning from basics?

@yensteel Ай бұрын

@@KZbin_User_ab Statquest is a great start

@YouTube_User_ab Ай бұрын

@@yensteel so, ml is all about statistics and probability ??

@yensteel Ай бұрын

@@KZbin_User_ab Nope, a lot of math and algorithms as well.

@yensteel Ай бұрын

@@KZbin_User_ab So, At the basics, all you really need to learn are: Python OR R, and the machine learning algorithms at an intuitive level. It tends to be the first class in a Data Science Masters program. For example, you learn python, so you can start coding. If/Else, functions, print, and arrays for data. You need to learn to get the data to fit to the parameters of the ML model. Some arrays are nx1, others are n, a 1d array, which are not compatible to create a m x 2 dataset. Others need to be transformed, so n x m is m x n array. Then once you're comfortable, you learn python packages such as pandas, numpy, sci-py, and matplotlib/seaborn. Anaconda is recommended as it installs all the needed components. Pandas reads CSVs and handles time data, categorical data, and can output to new csv. Numpy is the math library, and sci-py has most of the ML stuff. Jupyter notebook is a great way to work with python and get outputs. So, you start working with sci-py's ML tools to work with data. SVM and Random Forest is a great start, and you use Matplotlib to graph the results. Then, you learn neural networks.. It is HIGHLY recommended to code your first Neural Network from scratch, using basic layers in numpy. Keras in Tensorflow gets basic results with a few lines, but will be too limiting later on. Most people moved to Pytorch. You can practice NN for CNN for images, Classification of data, Regression, and LSTM/Transformers forecasting, and get a handle of them. Then, you're close to becoming a data scientist! Halfway there. Then, you can get on with what others love to talk about, LLM and GAN... or be like me and mess around with niche algorithms or your own ML algorithms. I use multi-gene symbolic regression so that I can make low-level ML inference at the Verilog level for HFT tradnig.

@ImmortalDuke Ай бұрын

There is always dots in a sausage stand.

@davidmurphy563 Ай бұрын

No. You should probably see a doctor.

@rmsgrey Ай бұрын

@@davidmurphy563 Or an optician

@TRex-fu7bt Ай бұрын

I thought this was going to be about Gaussian processes based on the sausage visual, so it was cool to learn about Bayesian optimization instead

@bl33kselderij Ай бұрын

Woohoo, awesome topic! :-) more Bayesian stuff please!

@yannkitson116 Ай бұрын

Careful the weather Gods don't like Bayesian stuff!

@GeoffryGifari Ай бұрын

So the sausage plot not only visualizes confidence in what the fitting curve looks like but also directly constrains the shape of the curve (to be inside the sausage)?

Ай бұрын

No. It does not constrain the shape. It just represents the current belief about what the shape could be. The whole point of most Bayesian approaches is to be able to update "beliefs" in response to observations. And here, the shape of that belief can then be used to make decisions about what to do next.

@GeoffryGifari Ай бұрын

Thank you for the clarification

@sillystuff6247 Ай бұрын

One of the best computerphile vids I've seen. ML/AI is the hottest thing on Earth (at the moment). This vid helps a tech person expand their intuitive understanding. Would like real world examples (actual data) showing what is meant by "exploring" & "exploiting" to get "lower function values". For example, maybe apply this Bayesian technique to finding an optimal cookie recipe. Helpful for less abstract thinkers. Appreciate Prof Osborne's calm clear way of explaining.

@jeromethiel4323 Ай бұрын

Oatmeal raisin! Solved it for you. ^-^

@landsgevaer Ай бұрын

Nuclear fusion research is a lot hotter, I would argue...

@tbird-z1r Ай бұрын

Hard to watch with the mask.. Should have recorded it when he was feeling better.

@Kirillissimus Ай бұрын

You forgot the most important thing about the topic - the model used and the initial assumptions made about the system. It is clear that you assumed that observations are perfect and some points have 0 uncertatinity, but what makes you think that between the sample some assumed paths are more probable than the others? Why you think that some short paths interpoladed by polynomials or splines are more likely to be "correct" than just linear segments or the path oscillating at random. What does a value of uncertainity even mean for a path that can never be estimated perfectly? There are many ways to understand it, from the mean distance to the expected path curve to adding another dimension and having an individual probability distribution for each point. It must all be strictly defined before we even attempt any statistical modeling. Otherwise it is all just meaningless handwaving.

@arturaugustyniak212 Ай бұрын

Kinda sorta reminds me gradient descent mixed with binary search

@charstringetje Ай бұрын

And of Thompson sampling

@arturaugustyniak212 Ай бұрын

@@charstringetje I wanted to grab a beer and chill out and now I have to read about it ;) Thanks! :D

@DF-ss5ep Ай бұрын

This probably applies to recommender systems as well, right? For example, in online commerce and other web environments, it is possible to do experiments dynamically

@isaac10231 Ай бұрын

You should do a video on long term memory and data corruption cause it seems people forgot masks were a thing.

@Species1571 Ай бұрын

Masks were a thing, their effect was not a thing.

@snex000 Ай бұрын

Even worse, some people forgot that decades of studies showed that masks don't stop airborne viruses.

@ThingsAreGettingTooSpicy Ай бұрын

@@snex000 Please cite those studies.

@snex000 Ай бұрын

@ Pubmed has always been there, dumdum.

@ThingsAreGettingTooSpicy Ай бұрын

@@snex000 Cool then you should have no problem actually citing it. Go on, I'll wait.

@superman39756 Ай бұрын

We need a numberphile or computerphile video on conformal prediction for uncertainty quantification! It is time as CP is rising in popularity in 2024

@puffinjuice Ай бұрын

I'd like to know how effective this is compared to just adding a datapoint midway. When I do experiemts I usually get a few datapoints first and fill in the gaps once I have identified the regions of interest. I wonder if a Bayesian approach like this would help me to identify the underlying behaviour faster?

@andyhall7032 Ай бұрын

you had me at sausage plots.

@stefanolassandro886 Ай бұрын

So cool!

@LoganKearsley Ай бұрын

My Bayesian prior is that everyone complaining about the mask is a right-wing American.

@andrewharrison8436 Ай бұрын

You would need a confidence level with that (100% isn't allowed but might be close).

@LoganKearsley Ай бұрын

@@KerrySoileau No application was necessary, as I was born to citizen parents on a US Air Force base.

@snex000 Ай бұрын

It's almost like they tried masks in 1918 and found through decades of scientific studies that they don't stop airborne viruses.

@mrbell2827 Ай бұрын

Boom

@northernmetalworker Ай бұрын

A little reductive don't you think?

@sahilsaraswat2588 Ай бұрын

Nice

@rqrqrqrqrq Ай бұрын

It's an ML video people. That's the contract when you're clicking to the video. He could've wear the Gandalf's hat and wouldn't matter a bit. edit: i am glad this comment became irrelevant

@codycast Ай бұрын

? Okay and?

@rqrqrqrqrq Ай бұрын

and no reason to polarize over nothing. when we can already have a common ground of shared interests

@error.418 Ай бұрын

wat edit: oh, idiots making negative comments about the mask, I'm with ya

@georgesos Ай бұрын

I wonder why he calls "observations" what are actually guesses. Yes it can be right, but it can also be far from being right.

@newerstillimproved Ай бұрын

Excellent visualization and explanation!

@Danny-hj2qg Ай бұрын

Could've been a Numberphile video.

@Juan-qv5nc Ай бұрын

probably

@HalfDoughnut Ай бұрын

lovely to see someone masking and showing it in video!!

@northernmetalworker Ай бұрын

When was this filmed?

@AgentM124 Ай бұрын

Machine Learning is so contagious these days :)

@nielsSavantKing Ай бұрын

But why this human is wearing a muzzle?

@Drakonak Ай бұрын

So happy to see an N95 casually in a video

@snex000 Ай бұрын

Why? You hate science or something?

@northernmetalworker Ай бұрын

Why would that make you happy? That's a sign of illness not health?

@error.418 Ай бұрын

@@snex000 wat

@snex000 Ай бұрын

@@error.418 Sorry you can't read scientific papers or look at the data on mask usage wrt viruses.

@error.418 Ай бұрын

@@snex000 Sorry you're clueless and cherry-picking

@jeromethiel4323 Ай бұрын

Sawsages! I watch this channel because i am very into computers. But i would have watched this in any case, because... SAWSAGES! P.S. I spelled sausages wrong on purpose, for emphasis. Saw-Sage-Es!

@bengoodwin2141 Ай бұрын

These comments are full of insane people

@tbird-z1r Ай бұрын

Brady heavily censors comments on his channel. Anything wrongthink will be removed, don't worry.

@ad_781 Ай бұрын

Oh no, it's Covid-24 !

@skytech2501 Ай бұрын

Why is he wearing masks?

@cocacoalabaer Ай бұрын

Maybe he's sick, maybe he doesn't want to get sick. Maybe he's in cancer treatment or had a transplant. Why do you even care?

@_zelatrix Ай бұрын

Because he's sensible. I'm actually thrilled to see it. Honestly I'd be annoyed to not see it in a video about Bayesian statistics in the middle of a pandemic that the world has been tricked into thinking has finished.

@Danny-hj2qg Ай бұрын

Probably recovering from the flu or pneumonia.

@bothrealting2598 Ай бұрын

@@_zelatrixbro what ! I don’t know anyone with Covid, you?

@hopperelec Ай бұрын

Someone else mentioned he had long COVID, which is more common than many people seem to think

@fruityfriend Ай бұрын

Whatever the reason for the 'face-diaper' as some other user called it: I hope you get better. Thank you for providing us with top notch educational content.

@misterhat5823 Ай бұрын

Face diaper beats pooping yourself like trump.

@UnderstandingCode Ай бұрын

Looks like its flu season across the pond as well

@niet9393 Ай бұрын

he has developed long covid. thats why

@quarteratom Ай бұрын

Stop wearing masks on video. Stop talking in masks on video. This video is unwatchable.

@hopperelec Ай бұрын

He probably has something like long COVID, or has an immune condition. The video is perfectly watchable, stop whining about someone trying to keep themself and others safe

@SDLearmonth Ай бұрын

❄️

@lbgstzockt8493 Ай бұрын

Cry about it.

@Yassinius Ай бұрын

Who cares?

@sanfransardine Ай бұрын

That's what I get from computerphile is watching the narrators beautiful lips. Without the mouths all the frisson is gone.

@JoãoPedroFavoretti Ай бұрын

Ahh is this guy wearing a mask?

@franklydoodle350 Ай бұрын

Maybe he gets easily sick

@djbehnkevideo Ай бұрын

Proof that you don't have to be smart to be a professor. Live in fear, doc!

@franklydoodle350 Ай бұрын

Maybe he has an autoimmune disorder, or maybe he is sick. Your comment is only proof that you jump to conclusions fairly quickly; not a great show of intelligence.

@ibrahiymmuhammad4773 Ай бұрын

hahahahah

@tbird-z1r Ай бұрын

For future videos, why not explain why he's wearing a mask? "I apologise for the mask, but I've recently been in hospital for some treatment." (Note the opsite dressing on the right forearm) We'd all understand if he said that, and even the most cooked of us wouldn't complain. But you let the poor guy look like he's some sort of nut job!

@misterhat5823 Ай бұрын

Why should he apologize?

@tbird-z1r Ай бұрын

@misterhat5823 Because it's weird, and it's harder for people to understand him. Acknowledging this and making your apologies is the polite and neurotypical thing to do in western English speaking countries.

@misterhat5823 Ай бұрын

@@tbird-z1r It's not hard to hear him. That's just trumpist BS.

@karenrobertsdottir4101 Ай бұрын

For future comments, why not explain your haircut? "I apologise for the haircut, but I recently had an accident with a blowtorch." (Note the singe marks on tbird-z1r's eyebrows) We'd all understand if he wrote that, and even the most cooked of us wouldn't complain. But you let the poor internet troll look as though he's some sort of a nut job!

@TRex-fu7bt Ай бұрын

Why not make someone explain why they are wearing glasses, hearing aid, a bandaid on their hand, etc?

@tupapau1241 Ай бұрын

WTH with the mask 😷. It's for painting... Not for talking about sausages 😂😂😂

@cocacoalabaer Ай бұрын

Maybe he's sick, maybe he doesn't want to get sick. Maybe he's in cancer treatment or had a transplant. Why do you even care?

@talideon Ай бұрын

He's ill and doesn't want to spread it.

@error.418 Ай бұрын

get over it

@rcmrcm3370 Ай бұрын

He heard you were coming and he found your body odor disgusting.

@tbird-z1r Ай бұрын

@@talideonHe should use a condom then

@JavierSalcedoC Ай бұрын

😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂😂

@timothybradford8216 Ай бұрын

Oh, a computerphile video! Oh, a mask.

@BarbarianGod Ай бұрын

oh a covid denier

@ldfgbnghcfhgvjbkhn Ай бұрын

Thank you patient zero.

@3rgoproxxy Ай бұрын

If only he could use logic to realise that his mask is doing nothing except cause more harm to himself.

@Matey8 Ай бұрын

His nose sounds a bit stuffed so he might be wearing it to hopefully reduce the chance of spreading to others, not protecting himself

@error.418 Ай бұрын

get over it

@tbird-z1r Ай бұрын

@@Matey8Those N95s block up a bit and if you're sick probably increase your risk of getting a secondary chest infection from the effect on breathing.

@misterhat5823 Ай бұрын

@@tbird-z1r Complete 💩

@tbird-z1r Ай бұрын

@@misterhat5823 Also there's a lot of microplastics, both into the environment and the user's lungs. "It was found that over the period of 8 h of usage, all the types of commonly used masks exude microplastics in thousands of numbers." KZbin doesn't let me link to the papers, but keyword masks microplastics and you'll find quite a bit of research. Of course, let's not trust *that* science.