How Machine Learning Makes Our Decisions Smarter
6:46
94,718
SciShow
Reddit
Thanks to IBM Z for sponsoring this episode. To learn more about the annual student contest, go to: ibm.co/34Ii9IT. Find out more about IBM Z Machine Learning capabilities here: ibm.co/2OVPP0i
Whether you're picking a place to eat or something to watch, machine learning helps us make smarter decisions in our daily lives.
Hosted by: Hank Green
SciShow has a spinoff podcast! It's called SciShow Tangents. Check it out at www.scishowtangents.org
----------
Support SciShow by becoming a patron on Patreon: www.patreon.com/scishow
----------
Huge thanks go to the following Patreon supporters for helping us keep SciShow free for everyone forever:
Kevin Carpentier, Eric Jensen, Matt Curls, Sam Buck, Christopher R Boucher, Avi Yashchin, Adam Brainard, Greg, Alex Hackman, Sam Lutfi, D.A. Noe, Piya Shedden, Scott Satovsky Jr, Charles Southerland, Patrick D. Ashmore, charles george, Kevin Bealer, Chris Peters
----------
Looking for SciShow elsewhere on the internet?
Facebook: facebook.com/scishow
Twitter: twitter.com/scishow
Tumblr: scishow.tumblr.com
Instagram: instagram.com/thescishow
----------
Sources
machinelearning.apple.com/2017/10/01/hey-siri.html
www.theverge.com/2018/5/9/17307156/google-waymo-driverless-cars-deep-learning-neural-net-interview
www.cs.princeton.edu/courses/archive/spr07/cos424/papers/mitchell-dectrees.pdf
www.cs.cmu.edu/~aarti/Class/10701_Spring14/slides/decision_trees.pdf
subs.emis.de/LNI/Proceedings/Proceedings165/170.pdf
www.irjet.net/archives/V3/i8/IRJET-V3I8393.pdf
news.mit.edu/2018/machine-learning-financial-credit-card-fraud-0920\
Image Sources:
www.istockphoto.com/vector/watercolor-wooden-background-gm497864388-79359543
www.videoblocks.com/video/friends-checking-out-a-match-on-a-dating-app-is-she-hot-or-not-hddy3f-1mj9s6sklv
www.istockphoto.com/photo/upset-woman-using-online-banking-service-problem-with-credit-card-gm1137821415-303553920
www.istockphoto.com/photo/stack-of-multicolored-credit-cards-close-up-view-with-selective-focus-gm1137281183-303217572
www.istockphoto.com/photo/apple-tv-gm459014981-23854275

Comments
  • Sophia Astatine
    Sophia Astatine

    I guess unicycles and tricycles aren't vehicles :(

  • Paul Moran
    Paul Moran

    Sorry folks, this is just another mainframe add. Not impressed.

  • Fernando JV
    Fernando JV

    IBM, your software sucks

  • Obelisk
    Obelisk

    "This episode of SciShow is brought to you by IBM Z" Me: imgur.com/LJllYs4

  • Delivery McGee
    Delivery McGee

    IBM has to advertise big iron on youtube? Not complaining that they're throwing money at SciShow, just seems to be an odd place to advertise that specific product.

  • Nick
    Nick

    fancy if statements

  • Justin Cullen
    Justin Cullen

    First question "what is the purpose of the vehicle?"

  • P4riah1
    P4riah1

    reminder: ai, machine learning, neural networks, etc, still contain the flaws and biases of the people who create them. the things they value, the inputs they provide. indeed, theres research happening now that is coming up with fairly good evidence that these systems actually amplify such biases, making them worse. 'they reason just like us, but faster!' = they make decisions full of racism, sexism, classism, etc, but faster! woo. these things arent making our decisions smarter. theyre deepening the biases in our decision making, while making us THINK theyre making our decisions smarter.

  • Sagacious Eagle
    Sagacious Eagle

    Machine learning vs. cold reading. There are similarities in how they ask questions.

  • Daniel Shao
    Daniel Shao

    MIT's algorithm may have reduced error rate by 54% by being more lenient but what about its false positive rate? How sensitive was it to detecting actual fraudulent purchases?

  • Katie T
    Katie T

    I had a bank that started flagging all transactions as fraud unless the chip was used, but the chips in their cards were so bad they degraded to the point of being unusable within 6 months....and they charged for replacement cards. I quit that bank within a few weeks of them doing that.

  • Joshua Holland
    Joshua Holland

    Look, I get the IBM is probably sponsoring you, and as IBM is desperate to keep mainframes relevant you probably have to talk them up, but if you’re going to discus mainframes and make them sound fantastic, you should also discus the market that’s killing them, at least when it come to classical mainframes. Let’s see an episode on virtualization. Let’s see an episode about client server and how virtualization can allow you to run hundreds of servers on a single piece of hardware. Most importantly let’s discus how a server and a mainframe differ.

  • Bjorn P. Munch
    Bjorn P. Munch

    I should refrain from using my credit card online when at the office, at least shortly before or after using it from home. At home, I'm always in Norway, but at the office I'm behind a corporate firewall and from other sites it looks like I'm in the UK. Or The Netherlands. Or Denmark. Or Romania. Or maybe some other random country in Europe.

  • 2MANYWWWWWWWWWWWWS4U
    2MANYWWWWWWWWWWWWS4U

    First question should never be does it have wheels. You ask how many wheels it has right off the bat. You wasted one of your questions by doing that.

  • Jere Lull
    Jere Lull

    "Master the Mainframe". Funny to hear about that, and IBM's new mainframe, in an arena where Micro- and mini-computers rule. I got along well programing anything BUT Big Blue for 35 years. I categorically refused to learn COBOL though that meant I was excluded from some large projects that essentially were lifetime jobs due to the errors that the environment generated. COBOL so programmer-unfriendly that it is tough to write clean code that can by properly debugged and maintained. One project I was involved in was developed in FORTRAN on a DEC mainframe (Yes, DEC built a few) For political reasons, it was re-written in COBOL on IBM. What had been a stable system that could be maintained and upgraded by a single programmer became an unstable mess that their dozens of coders could never get to work properly, despite having all the (working) code in hand.

  • Russian Rarities
    Russian Rarities

    Great job making the general public aware that AI != Machine Learning != Neural Networks (I am a Data Scientist).

  • Kwan Lowe
    Kwan Lowe

    The problem is that Netflix' algorithms are so poor. Maybe they are using decision trees for speed, but it would be better if they optimized for quality.

  • Joko Mandiri
    Joko Mandiri

    Oh wait until i run my dumb dumb codes with neuromorphic chips

  • Booglium Moo
    Booglium Moo

    Logic and emotion are directly linked. Consequence from poor logic bring's emotion that changes how we interact with the world. How could a machine possibly fathom the nurturing experience which is brought forth through making mistakes? To have a machine telling us how to think will only bring about emotion's that satellite around free will and it's true meaning. You cannot make a mistake or you will be bound to a machine. Is "God" a machine? Is this good for humanity if "mistake" can be leveraged through "bondage" via a very smart machine? Lingering [Transcendence] emotion's. You know...the movie. See...I have a very big issue with human being's referenced to as something akin to or something similar to robot's or slaves. So. To __leverage__ a person's privacy, profile, or soul because of past mistake allow's for legal passage to harvest, steal, and circumvent the good's of future production. It's a human farming machine.....your PC. *I GOT AN IDEA* = [$]

  • Feynmanfan
    Feynmanfan

    Decision trees do not reason like humans. If I were uncharitable I would suggest the sponsor is influencing you in this video. Introspection by humans is not a good guide for what is actually going on inside your head.

  • Sean Carroll
    Sean Carroll

    I beg to differ about machine learning being entirely useful, and we are on the platform right now. Yes, USsofts, I'm looking at you.

  • Mark Zuckerburg
    Mark Zuckerburg

    Machine learning is just statistics. Read elements of statistical learning, the machine learning bible

  • mtlicq
    mtlicq

    it can gfitselff

  • xMisariex
    xMisariex

    Just in time ML finals this semester. Thanks SciShow! :)

  • Alpha Delta
    Alpha Delta

    Learn. Linear. Algebra.

  • Raymond K Petry
    Raymond K Petry

    *_...BANK AI: 'Oh' it's Hank-Hank buys lots of frogs: coat-fasteners, belt-swordholder, flowerholder lumps, stringholders for his violin, French-dipped-fries, voice-garblers, horseshoe-shims, slimy toads-wait! You said, Sly-Me-Todes, the dead-metal-music-band?-DECLINED...!_*

  • TheFinalChapters
    TheFinalChapters

    Should've just asked "how many wheels does it have?" since apparently it's not just yes/no.

  • Thessalin
    Thessalin

    Hey! Y'all got a sponsor! Like a real adult not internet business sponsor! I know it sounds silly, but it's pretty interesting to see IBM sponsoring y'all. Neat!

  • Josue Corella
    Josue Corella

    Why haven't we found a bio machine that can connect and upload info directly into the brain? I need this like asap

  • Dennis Haupt
    Dennis Haupt

    "how many" is an invalid question. then you could just directly ask "what's the vehicle"

    • James Gubbins
      James Gubbins

      My thoughts, too-aren't decision trees usually limited to queries that can only be answered in a binary (i.e. yes or no) way?

  • Matthew Trzcinski
    Matthew Trzcinski

    They could reduce credit card fraud by comparing where my phone is against the transaction.

  • AxeLond
    AxeLond

    Why do we need to recreate how we reason, when we can recreate a system that can learn to reason like us?

  • John Opalko
    John Opalko

    So, what I hear you saying is that we are abnegating our responsibility to make informed decisions and turning them over to deterministic algorithms. That could explain a lot.

  • Aidan Or
    Aidan Or

    At 1:28 wouldn't the question "How many wheels does it have?" be better? The answer could have been 0.

  • General Durandal
    General Durandal

    Will Mainframe be fast enough to deal with the Neural Network? Find out next time, on IBM-Z!

  • Edward Varby
    Edward Varby

    I first encountered decision trees of this type in a simple example, for the Ripira Soviet programming language, found in the Andrey Ershov archive. Even though I've played with AI a bit, I didn't think they could be used so much. cheers!

  • Frank Em
    Frank Em

    Skynet

  • makaan1932
    makaan1932

    low resolution Hank is low resolution

  • IANF126
    IANF126

    the whole "2 states within an hour" purchase also gets thrown out the window when you deal with online orders. Just because you can purchase x or y item from the comfort of your own home doesn't mean the purchase will look like that to the bank, or at least it might not be formatted the same way as other online purchases. I'm not really sure how it looks to the bank but i can speak from second hand experience here, my mom bought something online and even though she does this a lot, the bank locked her account because they looked at it and thought the purchase was made in like italy because that's where the company was located. This was sorted out with a couple phone calls and all was well, genuine misunderstanding, but it just makes it that much harder to create a system that will catch fraud without being a burden, and i don't think that will ever happen. when you try to make something idiot proof there will always be a bigger idiot, similar logic applies.

  • Borden Fleetwood
    Borden Fleetwood

    Unless you screw up the answers by constantly ducking out of the decision tree halfway through, or maintaining a certain amount of cash-on-hand for all of those transactions that *aren't* buying pizza or fuel at that one corner store. Not that I do this as a rule, of course. That would be silly.

  • James Bcaster
    James Bcaster

    But Machine Learning can also make our decisions stupider. Our brain is designed to do inference based on raw data but if we receive preprocessed data, which is what recommendations are, we can easily push our generalisation to overdrive. Once I watched couple of Jordan Peterson videos and for some time after that I was bombarded with right wing propaganda. If we receive data at random we can still form somewhat balanced view but if the information we receive is already calibrated for your own preference you easily become an extremist. This is true for both left and right. I hope SciShow will do a video on that too. Knowing cons of science is just as important as knowing pros

  • BloatLers
    BloatLers

    Loved the videos but I don't understand why IBM would sponsor the videos (yes I know they are relevant subjects). Like no layman would buy a mainframe.

  • science
    science

    Way better mainframe ad than the first video a few days ago.

  • Jeann van Rooyen
    Jeann van Rooyen

    Great. Now they are "learning" how to annoy me...

  • CyanideSun94
    CyanideSun94

    Should i study? Yes. Can this count as studying? Also yes. Study two courses atm, one in decision support systems and one on legal aspects of AI (& ML)

  • spiaal
    spiaal

    why is machine learning not used before, why is this "new" ?

    • Hyrum Erdman
      Hyrum Erdman

      In many ways, processing power I believe. There are probably more reasons but considering how much faster and more efficient our computers get every year, not too long ago our machines just couldn't handle machine learning _on this scale_. I'm not an expert, just throwing a theory out there.

  • Isnamereally thatimportant
    Isnamereally thatimportant

    In the next 50 years : “How Machine Learning Makes Our Decisions”

    • Akshat Joshi
      Akshat Joshi

      15 years tops

    • Jerry Rupprecht
      Jerry Rupprecht

      Sooner

  • The heir of Slytherin
    The heir of Slytherin

    Shout-out to his ex at 2:48

  • sander j
    sander j

    It would be cool if you did a bit about random forest algorithms, it builds on the concept of decision trees but can be more powerful.

  • Kram1032
    Kram1032

    LOL that sponsor. I kinda doubt a whole lot of people will be interested in that. Like, for personal use.

  • Irving Chies
    Irving Chies

    just out of curiotisy I went to akinator and it took him over 60 questions to guess I was thinking of you Hank

    • Irving Chies
      Irving Chies

      @Kateryna Shneidmillier well I didn't know him either lol

    • Kateryna Shneidmillier
      Kateryna Shneidmillier

      And Dave Green it does not know at all ;)

    • Kateryna Shneidmillier
      Kateryna Shneidmillier

      Irving Chies John Green Takes only 24 questions to guess

  • Efsan Simanjuntak
    Efsan Simanjuntak

    May I help you translate the CCs to Indonesian? This topic is very into me, and I'd be glad to spread this to the broad Indonesian community.

  • Logical Conservative
    Logical Conservative

    It's also being used to censor opposing viewpoints, and push often-false political narrative.

  • 3nertia
    3nertia

    The first time he mentioned the "decision tree" I said, "Hey, that's how I make decisions!" xD

  • Puc Kingery
    Puc Kingery

    Sorry, I'm not going to watch your paid sponsor portion of the video because I HAD TO watch an Apple ad before the video. What's the point of a paid sponsor when you still have ads?

  • Tampa Tom Fishing
    Tampa Tom Fishing

    That’s Jarvis

  • Financial Education 101
    Financial Education 101

    Reduced by 54% ??? Did it reduce false positives or negatives or both ? Did it reduce one and increase the other ?

  • Azael
    Azael

    Hey USsofts!! Maybe you should use that open source algorithm that netflix uses, god knows your algorithm sucks. BTW, Netflix does not use a decision tree, they use collaborative filtering. AND decision trees, while they do fall into the broader category of machine intelligence, is not machine learning. In decision trees all of the questions and possible answers are all pre-programmed. It is really an advanced form of flow chart that is done on a computer. Neural nets are the real work horse of AI. As a stand alone model with supervised learning (10,000+ examples with answers) is a super pattern recognition system. Unsupervised learning requires a goal to achieve and something like reinforcement learning to tell the neural net when it's doing good or bad. For a totally different approach there is something called Q Learning. But you cant really use this on open world problems bcs it collects too large amount of data so it can only be used in small applications with limited variables. But in those limited cases it is remarkably fast.

    • Azael
      Azael

      @Kenny Olsen I stand corrected Kenny. Thanks for that. I dont work in data mining but these have apparently been around for decades, I'm surprised they haven't popped up on my radar for something.

    • Kenny Olsen
      Kenny Olsen

      Decision trees are a thing in machine learning, when they are trained from observations using splitting criterions like gini impurity or information gain, which are easily optimized

    • Azael
      Azael

      @Daniel Wimmer collaborative filtering is in netflix case where 200,000 other people hit like to same 16 movies you have. All of those 200,000 also liked some movies that you haven't watched yet. Like John Wick 3 and Sleepless in Seattle. So collaborative filtering would recommend those movies with a very high potential. The fewer of those people who hit like the lower its score. I think netflix probably uses other algorithms as well, for example bcs you have never watched a romantic comedy the Sleepless in Seattle score goes down. There are probably other rankings that can add or subtract from the final score but the movies with say the top best scores end up in your list. Q Learning keeps a database of records that include 1) the target/goal. 2) current readings/inputs 3) what is done to make the inputs match the goal. 4) number of times this has been tried. 5) success rate. An easy example might be a thermostat. Say target temp of 72 and when activated the heater stays on for 10 minutes. So it has a range of records starting at about 60° to about 80° (yes it does get hotter in the summer). So you need about 20 records and after a bunch of tries it's got down the best results and pretty much just sticks to them. But say you want a different goal or target temp at night of 64°. Now you need another range of records for the inputs and new target. But because the 10° range is too big at night you add a five minute heating in addition to the ten minute burn you had. Double the records again. Now add AC and double the records again. Add 3 zone controller and quadruple your record count. So even at this stage we can handle this on a pc or a phone, but try to put it on a cheap $5 microcontroller and it's way too much memory. Whereas an arduino can accommodate over 200 neurons on it. (Flatworms only need 197 neurons to find food, water, reproduce and wiggle around)

    • Daniel Wimmer
      Daniel Wimmer

      Will you please briefly explain collaborative filtering and Q learning or else provide good sources to research them? Thank you.

  • A H
    A H

    I disagree with the title. Decisions can't be "smart" or "dumb" a priori. So nothing a computer does up front can change whether a decision would have been "smart" or "dumb." ML only limits outcomes, and unpredictable outcomes are eliminated. ML can make a decision simpler, and I mean less complex not easier, but not smarter.

  • lornlynx
    lornlynx

    Decision Trees just can no longer keep up with deep learning, the amount of data available now helps the latter.

    • 3nertia
      3nertia

      You underestimate the ingenuity of humans to optimize :)

Next