Need a VPN that I've gone to great lengths to check works in Antarctica? Bam: www.privateinternetaccess.com/standupmaths Want to go to Antarctica and double check it for yourself? Bam: standupmaths.com/antarctica Sick of sponsor messages in pinned comments? Me as well, but this was a real expensive trip and it didn't feel right to make my patreon supporters pay for it. So here we are. Thanks for reading all the way down this message. I doubt many people will. If you're reading this: comment with a bad whale pun.
Just gotta appreciate the context here: a guy who just started off as a regular maths teacher at a completely normal English school is now going on all these crazy expeditions to spread the gospel of maths.
Yes, but Horatio knows that every dramatic sunglass removal needs to be accompanied by a cheesy pun, unlike Matth... erm, nevermind. Forget what I was saying.
Great job naming the 2 animals in the water. I have become an expert at naming trees. The 6 trees in my yard are named Fred, Rich, Bob, Neeko, Nakko, and Nocko.
To this day I still take a picture and sometimes think, "this would look nice in a spreadsheet" and so I head on over to Matt's website and let it generate my picture as an Excel file with all the conditional formatting
@@leptok3736 the captions say "tales of whales" which is just very interesting all around. It's almost like it was only a verbal typo, but then it's spelled like the story, so maybe there's layers of mistakes here?
You should name the whale "Parker". After all, if it seems you've discovered a new whale, but it's actually an extant whale, wouldn't that be an example of a Parker Whale?
I love that the video involves going on a grand adventure to Antarctica, but the headline used to pull in viewers is "I found a whale using maths" - which could potentially have been done from anywhere if provided the relevant data set - and I love that it worked and I and many others really are here for the maths with the globetrotting just being a bonus.
For edge cases, they really ought to compare where it was found vs how far it could feasibly have traveled in the time between sightings. For instance, it wouldn’t make any sense to mark a whale as being seen in Antarctica, and then seen a day later in the North Pole.
Unsolicited Excel tips: 1) Instead of =IF(A>0,A,0) you can use =MAX(0,A). Given A is very long in your formula this would save a lot of repetition. 2) Excel has built in matrix multiplication in the MMULT function.
I was expecting a call to action for commenting when you were saying "if you think you have a better whale name" so just telling people to get out there and find their own whale is the ultimate TH-camr call to action
This is so cool. There's a popular app called Merlin that does something similar for birds. But instead of identifying the unique individual you're looking at, it's able to tell you what species you're looking at (most of the time). I wonder if it used a similar neural network strategy.
Excellent video Matt, one small nit I'd add is the convolution doesn't exist to reduce number of *nodes* it exists to reduce the number of *weights*. Functionally you can think of a convolutional net as a fully connected one except you use the same weight for many of the edges, reducing number of parameters by a ton
Plot twist: The whales are on to us and are busy painting each others tails with squid ink. You think you found Bob the whale but it is George in disguise.
I'm well-acquainted with convolutional neural networks, but I've never seen them visualized in a spreadsheet like this before! This is a really fun way to explain the concept!
Surprised you never named it "Bowl of petunias". Also I wonder how long that spreadsheet would take on a Raspberry Pi 2? Will it ever be completed? Who knows.
@@Abigail-hu5wf Exactly. He saw a humpback. The sperm whale came into existence along with the plants. Hence the long name. Or call it "Oh no. Not again". Or even "42",
I'm taking a Computer Vision course in uni right now so for once I understood all the maths before even starting the video lol Very cool video, thanks Matt!
I love the technique of putting the image on a spreadsheet. It really goes to show how far technology has come. The program we normally use to do accounting and such has image processing power that would have been the envy of 1980s computer scientist, especially for a home machine.
Great, spontaneously diplomatic answer from Zoe about whether this new method is more desirable than the old, "I think it's one of the best uses of technology we have." It's natural to recognize the value of the new technology while having nostalgia for the old way.
I didn't even recognise that as a diplomatic answer. In the context of the question asked, i took it to mean Zoe was not only agreeing, but very much so.
I'm a little embarrassed for how long I thought Matt "found a whale that uses maths" until I got the correct interpretation of the title and the point of this adventure... 😄
Matt! The length of this video is perfect! 22:22 is neat number! The number 2 is already my favorite and lucky number, but 2222 is even cooler! Twenty-two is two 2's, but two thousand twenty-two is two 22's! Yessss!
I'm sorry Matt, my dyslexia lit up the word "indidividuals" like a huge red flag even if I just saw it for a slit second and now you can't miss it either. Again I'm sorry. Thank you for a fascinating video.
Biology is basically just fancy chemistry, chemistry is just physics and physics is just applied maths. So Matt is only three degrees of science removed from being a biologist.
You should make a follow up video going over the loss function that used. I’m assuming they’re used triplet loss, which is commonly used in facial detection, and I think it’s super interesting.
Again, I love your content, Matt! It’s you and Fluffy for me when I am looking for a good laugh. But more than just comic appeal, you always have something very interesting to teach me about. Thank you!
[ Matt Orca from Swim-up Maths ] I found a human using maths! I recently discovered that using clever mathematics, you can algorithmically identify the face of a human from a photograph... ;-)
I need to know more about this underwater world, is Matt Orca an orca? Is he a whale whose ancestors farmed dolphins? What numeric base does Matt Orca use to count? Does he run his algorithms on a fluketop computer or a reeftop?
You did a machine learning demo using match boxes to learn tic tac toe a while ago. That was a really eye-opening visualization. I have since pursued a very machine learning focused course at university and participated in shared tasks to solve other peoples problems (more academic than industry). And it's always fun to compare your failed approach to what other people came up with. So thank you for that video back than.
"I Found a Whale [that's] Using Maths" ≠ "I Found a Whale [by] Using Maths" Still gonna watch it, though; after the worst of the dejection has worn off.
Great ideas in here! I would love to learn more about 1.) how to convert image files into csv files and 2.) about the math behind the filtering! (Excel cells CQ4:CT7) (By the way: In the excel file "spreadsheets_and_tails.xlsx" offered for download (thanks a lot!) there are some absolute/relative reference errors in the if clauses. To correct it, make sure that in cell CW1 all references to columns CQ, CR, CS and CT have two $-signs in them; then expand the formula to the lower right corner of the image.)
As a bit of an Excel nerd, the formula Matt starts creating at around 9:44 could be done much easier like this: =MAX(0,SUM(OFFSET(A1,,,4,4)*OFFSET($CQ$4,,,4,4))) Still needs to be dragged to the dimensions of the original "image"
Matt, I've watched and rewatched every video on your channel and I just have to tell you, I love your humor. Its not laugh out loud haha funny, but a sort of "ha, I like that" kinda funny. It really puts me at ease and ready to enjoy whatever fantabulous escapade you have endeavored this time. That csi miami nod was a good one. Keep it up, I love every one.
would SUMPRODUCT have worked on the 2D array? edit: the answer is yes! I checked on the spreadsheet in the description, =IF(SUMPRODUCT(A1:D4,$CQ$4:$CT$7)>0,SUMPRODUCT(A1:D4,$CQ$4:$CT$7),0) does the same thing
... and min(1,max(0,sumproduct(…))) to speed it up a bit more. But Matt’s original did the job, and it’s not as though we’re saving a month of processing time here.
Using 3Blue1Brown's video to explain Neural Networks has such a reaction channel vibes, but is undoubtedly very effective. Funny we don't see more of that
CNNs are like a sort of blend of machine learning data analysis and classic computer vision. That first convolution layer you created looked a lot like photoshop's "find edges" filter, and it is the same algorithm! Those masks or kernels are used to find areas of interest in an image by calculating which pixels have significant contrast going in a certain direction. There's some interesting calculus there but because pixels are discrete data structures, the implementation of that math becomes just a collection of very simple addition/multiplication. It's been a while since I've done any CV work and I kinda miss it, it's really fun math
I vaguely remember that kernel stuff from an image recognition class and I think what you did here was generating the derivative of the image. If there are big changes in an area of an image, the derivative is also high, and if there are no big changes in the image, the derivative is low. It's like the derivative in maths which gives you the slope of a function, ie.e how strong the change is. In this case the slope is the change of the individual pixel values. It blew my mind, because something that was purely mathematical and abstract like derivatives could be applied to identify arbitrary features in images.
Channels with interesting content and understandable by such a wide range of backgrounds are so rare. We're having as much fun viewing as you did for writing! Thanks 🎉
Human visual processing works partly by different parts of the brain being good at identifying specific features, like vertical lines or horizontal lines, so it's not too surprising (but it is interesting) that a visual processing algorithm for machine learning uses the same strategy.
This is an amazing video using the ubiquitous Microsoft Excel in visualising the convolution (cross-correlation) operation between the filter and input image. Really motivates anyone to give it a try modifying the kernel as they whish and stacking the layers,
I read the title wrong here, I thought it was "I found (a whale using maths)" rather than "(I found a whale) using maths". I was quite far into the video when I finally realized there wasn't going to be a twist where whales use maths...
"If you change random parts of your code until you find something that works, thats called bad programming. But if you do it fast enough its called machine learning"
You need to add a pooling layer to effectively reduce the size of the data, computing a max or an average over a box that may be the same size of your kernel but with a stride > 1
Neural networks are fun. I don’t know much about them from the computing side, but when I worked in a marine biology lab in undergrad, we built a biological one out of sea slug brains that could distinguish between synechoccus and things that aren’t synechoccus. It was super cool!
Cool video! I would like to make one small clarification on CNNs. At 7:45 in the video you mentioned that is too many nodes for the NN. However, I think you are mixing up nodes and weights here. The kernels reduce the amount of weights required for one layer, but they do not reduce the amount of nodes and will instead increase it. You did show that some nodes are lost around the edges of the image, however usually people keep the 2nd layer image size the same as the first, and just do some zero padding or something when the kernel goes out of bounds. Instead, what actually happens is you usually end up with *more* nodes on the 2nd layer based on the number of different kernels you have, 93x60x(number of kernels). What then happens to reduce the nodes later down the line is something like max-pooling, or applying the kernel with some "stride" (skip every other pixel). One other reduction in pixel count that is introduced is by the cropping process. However, you might imagine that not all images of whales will be cropped to exactly 93x60 pixels, or maybe you want to automate the detection from a larger image. This would be another process, and there are many strategies on how to handle this (ie: A naive strategy is dragging your CNN across the entire image and using it like a kernel). And one tiny other thing that may interest you: while a 4x4 kernel is possible most CNNs use kernels with an odd number (3x3, 5x5, etc.). While an odd numbered kernel might intuitively make more sense since now there is an obvious "middle" pixel where the kernel is operating on, the actual CNN doesn't really care. It's just for implementation simplicity in the code, especially around the edges of the image. A cool math video to explore would be to show how many weights a CNN has compared to a fully connected NN, and how strides or max-pooling reduces the number of nodes and weights. It's a common question used in interviews for new college grads going into a job involving NNs.
Incorrectly regarded as a goof: Matt did not mess up the CSI gag when he whipped off his sunglasses; he deliberately parodied the action because David Caruso owns a copyright for "slowly putting on sunglasses while delivering pithy one-liner."
@@richardpike8748 hahaha ok, so, "your favorite copyright lawyer" might say you can't copyright individual moves, but you can copywrite a set coriography (music zoom move move) so blah funny reference anyway 😆
I've done this off the coast of southern California. Very cool that AI is able to assist in ecological conservation efforts. I also enjoyed how they updated me the next time the whale was spotted.
I actually thought you observed a whale, that was using maths 😅 i actually waited for that part of the video until you said thanks for watching and was really confused haha
At first glance, I though that Matt had found a whale that is using maths! How cool is that?!! But upon reflection, it’s probably Matt using maths to find a whale. Oh well, as long as it involves spreadsheets. /sigh
Need a VPN that I've gone to great lengths to check works in Antarctica? Bam: www.privateinternetaccess.com/standupmaths
Want to go to Antarctica and double check it for yourself? Bam: standupmaths.com/antarctica
Sick of sponsor messages in pinned comments? Me as well, but this was a real expensive trip and it didn't feel right to make my patreon supporters pay for it. So here we are. Thanks for reading all the way down this message. I doubt many people will. If you're reading this: comment with a bad whale pun.
hehe tweet
Also I hope you're doing whale, Matt! Your videos are always so whale done
I sorta wish you’d named your whale, Hex Hal. 😅
This video wasn't overwhaleming
Had a whale of a time whale watching your video!
There are a lot of good whale names out there, but I like Krill Bill too.
(That was the joke, but I also like Baleen Dion)
16:00 "Here we have another creature. Please identify this for my continued cooperation."
- Zoey Walker
🤣
15:40 Matt: You can't put whales in alphabetical order.
Zoey: Not so whale, no.
Perfect amount of sass.
What flavour is it ?
Actually just a Captcha.
Came to comments to be sure this was praised 👍
Just gotta appreciate the context here: a guy who just started off as a regular maths teacher at a completely normal English school is now going on all these crazy expeditions to spread the gospel of maths.
To the penquins.
Hi, do you have a moment to speak about our lord and saviour, Maths?
@@alphazero924 no come tomorrow
Mathelogy
spreadSHEET the gospels of math **
I'm sorry. I'll show myself out.
Putting on the sun glasses just to take them off is 100% something they would do in CSI.
true he looked like he does this every day. it was awesome though.
YeaaaAAAAAAAHHHHH!
@@JackBarlowStudios buuuuuum bum buuuum
@@JackBarlowStudios my dad side fam watches csi like 70% of the time someone's home lol
Yes, but Horatio knows that every dramatic sunglass removal needs to be accompanied by a cheesy pun, unlike Matth... erm, nevermind. Forget what I was saying.
16:00 "Please identify this [creature] for my continued cooperation"
I love this person XD
Great job naming the 2 animals in the water. I have become an expert at naming trees. The 6 trees in my yard are named Fred, Rich, Bob, Neeko, Nakko, and Nocko.
To this day I still take a picture and sometimes think, "this would look nice in a spreadsheet" and so I head on over to Matt's website and let it generate my picture as an Excel file with all the conditional formatting
How fun!
Wait what how, i got erroe
how do you make it one number like his? I dont want it a bunch of numbers I just want it being grayscale and stuff
The demonstration of a convolutional layer is really good! Not only is it intuitive, it is also interactive!
"for my continued participation please identify this creature" - what an awesome science person on your cruise...
And here I was thinking that Matt found an intelligent whale using math in its daily oceanic life.
So did I... Huge disappointment.
"Please identify this for my continued cooperation." LOL
I knew after "yep, we are off to antarctica" that you'll follow that up with the " we're gonna have a whale of a time" joke😂
Same here. Such an Aussie thing 🤣
"Whale of a book" also did not go unnoticed.
So I booked a ticket to Japan
I literally said it at the same time as him.
So corny never change
I love that at 0:23 you said tourists love photographing "whales of tails"
And later he says "You flick through the whale of the book" lol
Took me about five minutes to understand what that meant when I watched it lol
I took it as "whales of tales"
@@leptok3736 the captions say "tales of whales" which is just very interesting all around. It's almost like it was only a verbal typo, but then it's spelled like the story, so maybe there's layers of mistakes here?
It killed me! XD
You should name the whale "Parker". After all, if it seems you've discovered a new whale, but it's actually an extant whale, wouldn't that be an example of a Parker Whale?
That penguin floating by was a Parker whale.
If it's a whale using maths, it should Definitely be named Parker.
I mean if it actually is a whale, it should e named Parker's Dolphin
This is a whale of a Parker.
I believe this is a joke based on the "Parker Squares" episode he starred in on Numberphile. Looking it up, because it was funny.
'Projection Joke,' honestly made me laugh out loud. Well played.
I thought he was going to call it out as a Flat Earth joke. 😂
"Please identify this for my continued cooperation"
I love that the video involves going on a grand adventure to Antarctica, but the headline used to pull in viewers is "I found a whale using maths" - which could potentially have been done from anywhere if provided the relevant data set - and I love that it worked and I and many others really are here for the maths with the globetrotting just being a bonus.
Zoe was hilarious. I love charismatic scientists.
Thank you thank you! That is so kind 😄
@@zoewalker7690 it's the legend herself! 😱😁
For edge cases, they really ought to compare where it was found vs how far it could feasibly have traveled in the time between sightings. For instance, it wouldn’t make any sense to mark a whale as being seen in Antarctica, and then seen a day later in the North Pole.
That would be a nice filter to narrow down the possibilities. I wonder if they thought of that.
Well, they do check manually afterwards, I guess they just do it then.
About 20 seconds in, I guessed that there would be a "whale of time" joke at some point. I was not, however, prepared for that ending. 🐬🐋
Let's get honest here: Krill Bill is an excellent name, Matt!! 🐳
Unsolicited Excel tips:
1) Instead of =IF(A>0,A,0) you can use =MAX(0,A). Given A is very long in your formula this would save a lot of repetition.
2) Excel has built in matrix multiplication in the MMULT function.
I was expecting a call to action for commenting when you were saying "if you think you have a better whale name" so just telling people to get out there and find their own whale is the ultimate TH-camr call to action
This is so cool. There's a popular app called Merlin that does something similar for birds. But instead of identifying the unique individual you're looking at, it's able to tell you what species you're looking at (most of the time). I wonder if it used a similar neural network strategy.
Probably does!
finally, someone using spreadsheets for its intended purpose.
Took me a decent amount of time to realise there weren't gonna be any whales doing math in this video.
No, and some of them are even totally new to science, so they need an introductory course to get started.
Excellent video Matt, one small nit I'd add is the convolution doesn't exist to reduce number of *nodes* it exists to reduce the number of *weights*. Functionally you can think of a convolutional net as a fully connected one except you use the same weight for many of the edges, reducing number of parameters by a ton
The look on Matt's face for being right on his animal identifications while interviewing Zoe Walker--real joy.
Plot twist: The whales are on to us and are busy painting each others tails with squid ink. You think you found Bob the whale but it is George in disguise.
I'm losing track now; we have the parker square, the parker prime, and now the parker porpoise ?
Ty for yet another marvellous vid.
0:22 Photograph the "whales of tails" ?? hahah
"Whales of tails"*, I believe.
Also "indidividuals" on screen ~4:30
I'm well-acquainted with convolutional neural networks, but I've never seen them visualized in a spreadsheet like this before! This is a really fun way to explain the concept!
I think it deserves it's own video to disect the phrase 'the west of Antarctica'
I liked that bit of you "tweaking the knobs" - It really helped me understand image recognition a lot better than I did.
When zoe did the "ding! ding!" It immediately reminded me of the problem squared podcast :)
Love watching absurd uses of excel.
Surprised you never named it "Bowl of petunias". Also I wonder how long that spreadsheet would take on a Raspberry Pi 2? Will it ever be completed? Who knows.
IMHO he missed a great opportunity to introduce the world to the Parker Whale. Or the Parker Tail. Parker Whale Tail? Something like that.
It would need to be a sperm whale for that!
_Never_ named it? Did he have multiple opportunities?
Tempted to try the spreadsheet algorithm on my RPi 3B. A little more powerful I guess, but I can slow it down XD
@@Abigail-hu5wf Exactly. He saw a humpback. The sperm whale came into existence along with the plants. Hence the long name. Or call it "Oh no. Not again". Or even "42",
I'm taking a Computer Vision course in uni right now so for once I understood all the maths before even starting the video lol
Very cool video, thanks Matt!
This deserves to be viewed on the projector.
I love the technique of putting the image on a spreadsheet. It really goes to show how far technology has come. The program we normally use to do accounting and such has image processing power that would have been the envy of 1980s computer scientist, especially for a home machine.
Great, spontaneously diplomatic answer from Zoe about whether this new method is more desirable than the old, "I think it's one of the best uses of technology we have."
It's natural to recognize the value of the new technology while having nostalgia for the old way.
I didn't even recognise that as a diplomatic answer. In the context of the question asked, i took it to mean Zoe was not only agreeing, but very much so.
I'm a little embarrassed for how long I thought Matt "found a whale that uses maths" until I got the correct interpretation of the title and the point of this adventure... 😄
Matt! The length of this video is perfect! 22:22 is neat number! The number 2 is already my favorite and lucky number, but 2222 is even cooler! Twenty-two is two 2's, but two thousand twenty-two is two 22's! Yessss!
What is surprise when your upload contained more spreadsheet math, amazing!
Convolution Is one of those thing that is never really taught or explained in a way that makes me understand it
It also means slightly different things in different contexts.
yeah it really is quite... convoluted
there are a bunch of explanations on TH-cam, if you find a favorite I'd love to hear you come back and comment!
I'm sure Krill Bill is (or would be) very happy with his new name
I'm sorry Matt, my dyslexia lit up the word "indidividuals" like a huge red flag even if I just saw it for a slit second and now you can't miss it either. Again I'm sorry. Thank you for a fascinating video.
Biology is basically just fancy chemistry, chemistry is just physics and physics is just applied maths. So Matt is only three degrees of science removed from being a biologist.
You should make a follow up video going over the loss function that used. I’m assuming they’re used triplet loss, which is commonly used in facial detection, and I think it’s super interesting.
I just noticed the education poster in the background. "When all the parts are working." 🤣🤣🤣🤣 It's a very Parkeresque poster.
Got unreasonably excited when you started talking about werewhales being spotted, before realising I'd probably misheard
Again, I love your content, Matt! It’s you and Fluffy for me when I am looking for a good laugh. But more than just comic appeal, you always have something very interesting to teach me about. Thank you!
This whole concept, from machine learning to integrating the photos, is so COOL
[ Matt Orca from Swim-up Maths ] I found a human using maths! I recently discovered that using clever mathematics, you can algorithmically identify the face of a human from a photograph... ;-)
Lol 😂
I need to know more about this underwater world, is Matt Orca an orca? Is he a whale whose ancestors farmed dolphins? What numeric base does Matt Orca use to count? Does he run his algorithms on a fluketop computer or a reeftop?
@@Milamberinx Someone needs to write this fanfiction
@@miriamrosemary9110 I nominate my high school maths teacher, who's a marine biologist.
@@hughcaldwell1034 Perfect
You did a machine learning demo using match boxes to learn tic tac toe a while ago. That was a really eye-opening visualization.
I have since pursued a very machine learning focused course at university and participated in shared tasks to solve other peoples problems (more academic than industry). And it's always fun to compare your failed approach to what other people came up with.
So thank you for that video back than.
"I Found a Whale [that's] Using Maths"
≠
"I Found a Whale [by] Using Maths"
Still gonna watch it, though; after the worst of the dejection has worn off.
I can't stop watching the bit at the end, it's so clean. 10/10 pun and setup.
Great ideas in here! I would love to learn more about 1.) how to convert image files into csv files and 2.) about the math behind the filtering! (Excel cells CQ4:CT7)
(By the way: In the excel file "spreadsheets_and_tails.xlsx" offered for download (thanks a lot!) there are some absolute/relative reference errors in the if clauses.
To correct it, make sure that in cell CW1 all references to columns CQ, CR, CS and CT have two $-signs in them; then expand the formula to the lower right corner of the image.)
I remember participating in Humpback whale competition on Kaggle. Glad to see these solutions getting deployed in real world.
this is my new favorite video, this is literally machine learning by hand in excel
0:22 "...the whales of tails..."
As a bit of an Excel nerd, the formula Matt starts creating at around 9:44 could be done much easier like this:
=MAX(0,SUM(OFFSET(A1,,,4,4)*OFFSET($CQ$4,,,4,4)))
Still needs to be dragged to the dimensions of the original "image"
Matt, I've watched and rewatched every video on your channel and I just have to tell you, I love your humor. Its not laugh out loud haha funny, but a sort of "ha, I like that" kinda funny. It really puts me at ease and ready to enjoy whatever fantabulous escapade you have endeavored this time.
That csi miami nod was a good one. Keep it up, I love every one.
would SUMPRODUCT have worked on the 2D array?
edit: the answer is yes! I checked on the spreadsheet in the description, =IF(SUMPRODUCT(A1:D4,$CQ$4:$CT$7)>0,SUMPRODUCT(A1:D4,$CQ$4:$CT$7),0) does the same thing
... and min(1,max(0,sumproduct(…))) to speed it up a bit more. But Matt’s original did the job, and it’s not as though we’re saving a month of processing time here.
@@andrewkepert923 no, we did that last week
This is such a great video! Love spreadsheets, whale, maths, Antartica, terrible puns. Fantastic!
Using 3Blue1Brown's video to explain Neural Networks has such a reaction channel vibes, but is undoubtedly very effective. Funny we don't see more of that
CNNs are like a sort of blend of machine learning data analysis and classic computer vision. That first convolution layer you created looked a lot like photoshop's "find edges" filter, and it is the same algorithm! Those masks or kernels are used to find areas of interest in an image by calculating which pixels have significant contrast going in a certain direction. There's some interesting calculus there but because pixels are discrete data structures, the implementation of that math becomes just a collection of very simple addition/multiplication. It's been a while since I've done any CV work and I kinda miss it, it's really fun math
I vaguely remember that kernel stuff from an image recognition class and I think what you did here was generating the derivative of the image. If there are big changes in an area of an image, the derivative is also high, and if there are no big changes in the image, the derivative is low. It's like the derivative in maths which gives you the slope of a function, ie.e how strong the change is. In this case the slope is the change of the individual pixel values. It blew my mind, because something that was purely mathematical and abstract like derivatives could be applied to identify arbitrary features in images.
This is so subtle, but I love how you refer to others as “they.” I don’t feel I need to elaborate, but I notice and appreciate it. :)
21:40: A preview of the next Pi Day video, in which Matt calculates pi using his nose.
Channels with interesting content and understandable by such a wide range of backgrounds are so rare. We're having as much fun viewing as you did for writing! Thanks 🎉
Human visual processing works partly by different parts of the brain being good at identifying specific features, like vertical lines or horizontal lines, so it's not too surprising (but it is interesting) that a visual processing algorithm for machine learning uses the same strategy.
I agree with Zoe Walker. This is some of the best use of technology and maths yet.
Only at the ad read and I can tell this is going to be a great video
I'll admit I was a little bit disappointed that this wasn't about a whale that was found to be using maths.
This is an amazing video using the ubiquitous Microsoft Excel in visualising the convolution (cross-correlation) operation between the filter and input image. Really motivates anyone to give it a try modifying the kernel as they whish and stacking the layers,
For some reason, Respighi's "Pines of Rome" were playing in my head for most of the video
And I thought it was a "whale using Maths" now that would have been fun.
17:43 this was a great bit
I'm always amazed by how much utility you can squeeze out of a spreadsheet.
I read the title wrong here, I thought it was "I found (a whale using maths)" rather than "(I found a whale) using maths". I was quite far into the video when I finally realized there wasn't going to be a twist where whales use maths...
Projection jokes are objectively the best jokes
"If you change random parts of your code until you find something that works, thats called bad programming. But if you do it fast enough its called machine learning"
You need to add a pooling layer to effectively reduce the size of the data, computing a max or an average over a box that may be the same size of your kernel but with a stride > 1
That excel demo is just beautiful
Neural networks are fun. I don’t know much about them from the computing side, but when I worked in a marine biology lab in undergrad, we built a biological one out of sea slug brains that could distinguish between synechoccus and things that aren’t synechoccus. It was super cool!
That does sound cool!
Correction: 4:22 small typo in the word individual
Cool video! I would like to make one small clarification on CNNs. At 7:45 in the video you mentioned that is too many nodes for the NN. However, I think you are mixing up nodes and weights here. The kernels reduce the amount of weights required for one layer, but they do not reduce the amount of nodes and will instead increase it. You did show that some nodes are lost around the edges of the image, however usually people keep the 2nd layer image size the same as the first, and just do some zero padding or something when the kernel goes out of bounds. Instead, what actually happens is you usually end up with *more* nodes on the 2nd layer based on the number of different kernels you have, 93x60x(number of kernels). What then happens to reduce the nodes later down the line is something like max-pooling, or applying the kernel with some "stride" (skip every other pixel).
One other reduction in pixel count that is introduced is by the cropping process. However, you might imagine that not all images of whales will be cropped to exactly 93x60 pixels, or maybe you want to automate the detection from a larger image. This would be another process, and there are many strategies on how to handle this (ie: A naive strategy is dragging your CNN across the entire image and using it like a kernel).
And one tiny other thing that may interest you: while a 4x4 kernel is possible most CNNs use kernels with an odd number (3x3, 5x5, etc.). While an odd numbered kernel might intuitively make more sense since now there is an obvious "middle" pixel where the kernel is operating on, the actual CNN doesn't really care. It's just for implementation simplicity in the code, especially around the edges of the image.
A cool math video to explore would be to show how many weights a CNN has compared to a fully connected NN, and how strides or max-pooling reduces the number of nodes and weights. It's a common question used in interviews for new college grads going into a job involving NNs.
Incorrectly regarded as a goof: Matt did not mess up the CSI gag when he whipped off his sunglasses; he deliberately parodied the action because David Caruso owns a copyright for "slowly putting on sunglasses while delivering pithy one-liner."
no way that action is copyrighted lmao
@@richardpike8748 Sounds like you're trying to tell me...
(⌐■_■) ( •_•)>⌐■-■ ( •_• )
...cetacean needed.
~YEEEEEEEEEEEEEEEEAAAAAAAAAAAH!!!~
@@richardpike8748 hahaha ok, so, "your favorite copyright lawyer" might say you can't copyright individual moves, but you can copywrite a set coriography (music zoom move move) so blah funny reference anyway 😆
@@generrosity Fair enough I suppose
I love that both the start and the end is the same cliche hollywood put on to take off the sunglasses
Thank you, this is amazing!
19:53 " If you think you have a better and or funnier whale name, find your own whale! "
You're a funny man Matt. 🤣🤣😂😂
I've done this off the coast of southern California. Very cool that AI is able to assist in ecological conservation efforts. I also enjoyed how they updated me the next time the whale was spotted.
I actually thought you observed a whale, that was using maths 😅 i actually waited for that part of the video until you said thanks for watching and was really confused haha
At 4:26 , I assume Humpback Whale indidividuals is supposed to say individuals, as individuals is an actual word. So close!
"I think I'm becoming a biologist"
Glorious pick up line!
The Pines of Rome humpback whale animation footage from Fantasia 2000 unlocked a core memory for me 😍
Yep, came here to say the same thing.
Now my life goal is to go to Antarctica just to find my own unique Whale
I was bingeing your videos and said to myself "wait, I don't remember this one" before realizing it's a new one 🤩
At first glance, I though that Matt had found a whale that is using maths! How cool is that?!! But upon reflection, it’s probably Matt using maths to find a whale. Oh well, as long as it involves spreadsheets. /sigh