so over the last couple of lectures

artemy has been telling you about

neural networks and how how many how

these major advances have been coming

about as a result of this new technology

in this in this lesson we're more going

to talk about the caveats of machine

learning and things to think about

before you deploy these new systems so

one of the major piece of advice that we

have for you is that data and questions

come before algorithms so that is a lot

of machine learning problems don't need

advanced techniques like deep neural

networks with LST M's and rel use and

all these fill these other buzzers right

a lot of questions can be solved by

having a really clear question that

you're answering and a really good data

set that you have to answer that

question and so the first things to

think about are what what do I want to

know and if I can't know what I want to

know is there something that will be

useful anyway and secondly do I have the

data to answer that question that if I

don't have that data how can I go about

and get it and so neural networks

oftentimes really need really large data

sets in order to work properly and for

smaller data sets there are other

different algorithms that work better

with small data and so this is one thing

to think about is trying these other

algorithms before jumping on the neural

network high train and again to

emphasize think about data and questions

first before we think about algorithms

so on that point about clean data though

another point we want to make is that

it's not always obvious how to make data

clean and so in particular certain

things are obvious right you can

represent age as a number you can

represent zip code as a number but other

things aren't so obvious so let's

imagine we want to build a product

recommendation system and you've seen

that someone in the past has built has

bought a tripod they've bought a

microphone well how do we take that data

and make new recommendations right one

way you could do it is to literally list

the product number that they bought but

that might be really confusing maybe you

want to represent that as they bought a

tripod so maybe if you see they've

bought a tripod and they bought a

microphone they may also need a camera

but the same time if you see that

they've been looking at six different

kinds of tripod maybe you do want those

as exactly the tripod that they are so

you can make a different recommend

for another tripod for instance so

another thing is that natural language

data can be really difficult to

represent and so you might take each

individual word and make it a number but

a lot of times these words combine in

difficult ways now hypothetically you

could take an entire sentence and

represent it as a number but that's not

going to be very useful to use because

most sentences aren't used all that many

times so again it's really tricky and

requires a lot of research to figure out

how it's a map something like natural

language data and have it say oh this

person likes me or doesn't like me so

the other caveat we want to mention is

that is this concept of hardware so

these neural networks and other machine

learning algorithms can be oftentimes

need to be trained on large data sets

and they need to run through this data a

lot of times and so the thing is is that

it often takes a lot of hardware in

order to practically train these models

and so alphago which we talked about a

couple of lessons ago that actually ran

on something called a tensor processing

unit which is a specially designed piece

of hardware in order to run fast and so

you don't necessarily need that much

customized specialized Hardware but it

is the case that oftentimes you need

lots of graphics cards or access to

something like the Amazon Cloud in order

to practically train these algorithms so

it's another thing to think keep in mind

of depending on how long you have to

train there can be different trade-offs

and machine learning algorithms a third

thing though to talk about which i think

is big is this idea of transparency and

robustness so many of you may have seen

projectile motion either an introductory

physics class and let's think about we

have paratroopers that are going to jump

out of an airplane and we're gonna take

Newton's Newton's laws of motion with

projectiles that say well the velocity

with which they hit the ground is a

square root of two times the height they

jumped times the acceleration due to

gravity and if we run those numbers we

find that they're gonna hit the ground

at a hundred and twenty miles an hour

that's a really fast speed so they

probably shouldn't be jumping right

wellnot right so we know Newton's laws

and with projectiles ignore air

resistance and these paratroopers have

parachutes which and operate entirely

under air resistance so we know that

this

is going to be wrong because of air

resistance being around ok so here's

another example there's a neural network

that classified this Pando with 50 7.7%

and it classified this very close image

of a panda as a given with 99.3% so why

is it that the neural network made this

mistake well it's really not clear right

with and so this is one of the major

benefits to say scientific discoveries

over pure machine learning in science we

often know what the assumptions are and

when those assumptions don't no longer

hold well with machine learning

algorithms and particular the more

advanced machine learning algorithms we

don't really know what core assumptions

are operating under and therefore when

we would expect them to fail so this is

something to keep in mind before you

deploy or use a machine learning

algorithm is to understand what you need

to know in terms of what kinds of

mistakes that it's going to make or not

in this particular example was made with

a very particular kind of thing called

an adversarial input where they knew how

the neural network behaved and they took

that original Panda image and found a

very small change in the image that's an

imperceptible change to humans but to

the neural network made a very big

difference and so there are kind of two

messages to come here one of which is

that while it seems like machine

learning algorithms are intelligent and

often reproducing human intelligence

it's certainly sure that they don't make

mistakes and the same way humans make

and the second thing to be a wary of is

this kind of security issue in that

there are these kinds of inputs that can

completely fool the system in ways that

are really unclear and so it's important

to keep these kinds of things in mind

before you just let them loose and let

them do whatever they want another major

caveat we want to talk about or what

we're calling feedback effects that is

when the neural network influences its

own predictions and so a simple example

of this is that t caste system that was

mentioned way back in Lesson one where

it was this sort of series of if

statements that told a pilot how to

behave in order to try to avoid two

pilots colliding one problem with the

original T caste system is it didn't

know about something called the ground

and so it oftentimes make bad

predictions to pilots say when they were

taking off and landing and so the system

wasn't really sophisticated enough to

deal with the full environment in which

it was deployed and so part of the

problem was that pilots now sort of

learned to not really trust the system

because it wasn't making the decisions

that they hoped it to make and so

because of this feedback loop the system

wasn't as effective as they hoped it

would be originally and had to upgrade

the system several times so let's

consider another example from

epidemiology where maybe we're gonna

model a disease outbreak and we're gonna

try to see how quickly it's gonna spread

through the population now these models

often assume that I'm gonna do what I

normally do so I'm gonna go to work I'm

gonna come home maybe I'll go to a

friend's house and that's gonna

influence who I interact with and

therefore who I can spread the disease

to however if I'm going to take this

epidemiology model and say oh my gosh

this is really bad and then announce to

everybody that we should be really

careful well now all the sudden maybe

I'm not going to be going to work maybe

I'm going to do something entirely

different instead and maybe in fact I

have the disease but I'm going to go to

an entirely different City to avoid it

thus spreading the disease much faster

than if I hadn't used the model so it's

important to make sure that the model

kind of isn't influencing how the system

that it's modeling is going to behave

and another kind of example of this

comes from something called recidivism

in parole decisions so deciding if you

should grant parole or not to

individuals and there are some issues

with fairness that are being researched

now in terms of how fair or say biased

are these decision-making algorithms and

that's one major thing to think about

and the second related thing to think

about is let's say we don't grant parole

to a certain class of individuals like

our algorithm just says never never

grant them parole well we're also if we

want to collect data in the future we're

never getting data on that category of

individuals and so we never learned that

our system maybe maybe was wrong and so

again if the machine learning algorithm

is influencing the data that then is

going to train the next iteration of the

machine learning algorithm you really

need to be careful so another thing to

think about is that often times problems

aren't even AI problems at all so there

are lots of

that involve humans getting together to

make decisions as collectives for

example political problems or to deal

with emotional or relationship issues

like social problems and in many cases

just computers aren't there to solve

these problems for us these are human

human problems that involve conflict

resolution and other very important

discussions that we just can't program a

computer to tell us the right solution

another example though is that there are

learning problems that aren't machine

learning problems so as mentioned

machine learning is generating new

predictions from data but a lot of

problems don't require data they require

insight and so in the middle here for

example is there's a text that was looks

like gibberish but in fact it was

generated according to something that's

called a Caesar cipher and so what a

Caesar cipher is is that for each letter

in the alphabet we replace it with a

different letter in the alphabet so for

example a becomes F and B becomes Q etc

and so if you have insight what you can

realize is that in English texts the

letter E occurs the most often so maybe

we can count up the letter that occurs

most often and make that in E and then

by doing this procedure you can

translate that back into its original

text but we see here right that this

required two levels of insight it

required that we understand that this

text isn't gibberish maybe it does

represent something real and the second

thing was how to go about this

translation process it didn't require

lots of data

it required insight to see that that's

there another example of this is the

code talkers during world war ii where

instead of taking english encrypting it

and sending it over the radio instead

they had native navajo speakers speaking

over the radio so if you're German and

have never heard Navajo before this

sounds like gibberish and maybe you

think that it's this text is somehow

encrypted well instead it's a different

language entirely and if you had the

insight to realize that it's Navajo

maybe you could do something about it

but in Germany there weren't very many

native Navajo speakers and so they never

noticed this so to summarise machine

learning is good and has made lots of

major advances on problems that

previously had been hard however there

are these really important caveats to

consider and so our advice

you is to have clean data and clear

questions before you start to think

about algorithms make sure you have what

you need to answer the question before

you think about ways of how to answer

the question and so and also then before

you deploy this system we encourage you

to think about all these caveats that

we've mentioned for example transparency

and different kinds of feedbacks that

the system might have with itself