Act, Confirm, or Stop? Smarter behavior for AI assistants, wearables & robots — Amit Desai, Roku

[music]

Hi everyone. How's it going? Hey

Patricia, how are you?

>> Uh so last presentation of the day, so

let's make it count. Um,

all right. Let's, uh, let me start with

a little bit of background on myself.

And, um, my background, I'm a voice

subject matter expert. I've been working

in voice AI for a long time across

different surfaces, devices, and um,

both at Alexa, at at Roku, at my own

startups, you know, in the app store.

And my perspective is a little different

from a lot of other voice AI

practitioners. I think it's a

combination of um a deep um voice user

interface expertise and intu intuition

mixed in with new technical approaches

uh that I think can produce really

magical experiences. So I think it's

both sides and I think that's especially

true in this new area that we're in with

frontier tech where the human interface

is basically being redefined. So let me

start with uh I'll just blast through

the first couple of slides then get to

the premise. I think everybody knows

that voice has incredible potential.

There's the power of voice I think

across everywhere. It's the most natural

interface. Humans love talking. And uh

the problem is the other half is the

pain of voice. So it's the power and the

pain. Voice is errorprone. And I think

those errors are going to continue for a

while. And I think the cost or

consequence of those errors is going to

grow, especially as we go fromational

AI bots to embodied AI where rather than

just giving answers that might be

erroneous,

we're going to have AI systems take

physical actions or digital actions

where, you know, if the robot throws

your watch out with the trash, it's a

lot worse than playing the wrong song.

So I do think that a new approach is

definitely needed and here's the TLDDR

of the premise we're going to walk

through today. Um there are two ways to

improve customer or user satisfaction of

a voice AI assistant and that is by

increasing accuracy which people know

about I mean technically accuracy and

the other is a different knob that we

have that we are not using adequately

and I'll call that a system decision

which we will define which is orthogonal

which is different from accuracy and I

believe This approach which I have used

in several different environments and

seen some success I think is a promising

area that we should consider developing.

Um let me walk through this with a

simple smart speaker example and we'll

go step by step with this approach but

it is a scalable approach that I think

uh can apply across different surfaces

and devices. So let's get started. So

suppose we all you know are making a

smart speaker coincidentally called uh

Alexa and Alexa is very simple. It just

allows you to you know ask for music and

it'll play a song and of course it will

play either the song you wanted or a

different song. So it'll be right or

it'll be wrong. This isn't that

different from what you've seen out

there. Um now let's to first talk about

accuracy. Accuracy. Let's say we define

it as we you know take a thousand spoken

requests. We observe the input and the

output. We label it and we look at this.

This is the map of a thousand points and

79% of the time 790 dots here were

actually the correct song. This is let's

say human annotated 20% 21% wrong song.

So that's the accuracy. Now, like I

said, knob one is to spend a lot of time

working on improving the accuracy, you

know, um, percentage point by percentage

point at any layer in the stack.

There's, if it's a cascaded system, you

know, there's a perhaps a wakeword layer

and a speech ASR layer and a NLU layer

which might have intent classification,

entity extraction, a lot of different

layers, VAD, etc. And any of those can

contribute to errors. So we spent time

we might be able to reduce that 210 to a

smaller number that is I think a known

area that we're tackling but I think

knob 2 which is what I was talking about

is what we'll go through here which is

keeping the accuracy exactly the same.

So 79% what could we do in conditions of

uncertainty to improve user satisfaction

apparent and I I think we can do a lot.

So let's start first with the original

system is just acting like I said user

says something system plays a song it's

either the right song or the wrong song

immediately I think just common sense

tells us that we could introduce at

least one system behavior to stop or

rather to reject the hypothesis and do

nothing. So uh there is now one more

option to decide the system may decide

and say sorry I didn't get that or sorry

could you repeat that? Uh the challenge

of course is how how when do we decide

to stop and I mean quantitatively. Um

here's one approach to kind of

visualizing this because if we don't

we'll just take probably some swag like

some guesstimate and I'll prove that if

we just took a guesstimate we would end

up with a worse situation than a more

rigorous approach. So let's just assume

I took those thousand data points and

like I said they've been annotated and

we assign a confidence score a single

confidence score to the hypothesis that

was generated by the system you know

between zero and one and let's say it's

reasonably calibrated. This is a

simplification of if it's a cascaded

system there are multiple layers and

multiple you know confidence scores but

let's just assume that for now. Whoops.

So we're going to have 790 points 200

that are correct 210 wrong. Each one has

a confidence score and we're going to

plot it, you know, plot the

distributions. Uh on the x-axis, I've

just converted from 0ero to one to

percentages. And the question is how do

we choose a threshold t such that

whatever that percentage is um to the

left of it meaning if when the system um

forms a hypothesis if the confidence

score c is less than that t stop and say

sorry otherwise play question is how do

we choose a t so far everything I'm

saying is fairly common sensical but

this is where um intuition will fail us

we might say something like okay I don't

know let's do 65%. It seems you know gut

feeling like okay it's kind of confident

that's probably when we should speak. Um

now here's where we start coming out

with some sophistication.

Any tea we choose is producing bad

outcomes. Bad in the in two fields. One

is obviously on the left side anytime

you stop it's bad. The user doesn't want

it to stop. He wants to they want to

hear their song. The other bad is if you

do play a wrong song, of course that's

bad as well. So these are two two kinds

of bad outcomes. But here's the

important part. Now I've like elaborated

on the um tree diagram on the right hand

side. The bad outcomes are not equally

bad. They're not the same thing from a

user perspective. And obviously let's

let's think about it. If the wrong song

plays, you said play kiss and it starts

playing kiss by Chris Brown instead of

the one by Prince. That's going to be um

the highest user cost. Now I'm defining

user cost from the user's perspective.

First I have to like hear music and

realize that is not Prince. Then I have

to shout over my Alexa and um you know

get it to stop and then I have to

re-request. All of that is a lot of

effort. that is definitely a worse

outcome than the system stopping and

saying sorry I didn't understand that

however we should go further and try to

quantify that relative badness and there

many ways to do it and I think this is

an area to be explored for now let's

just consider this a heristic of if that

outcome happens how many more seconds

additional seconds will it take for the

user to get back to success which is to

play the song they wanted kiss by Prince

and I'm I just put down some numbers

here. Let's say in the case of a bad

song, it's 10 seconds if you add up all

the things I got to do. And if it's a I

didn't understand you, it's 4 seconds

because that's how long it would take

you to respe and and the extra latency.

And now here's where we can start

utilizing that. If we go back to our

distribution curve on trying to find out

where is T. Now we've basically turned

this into a problem of minimizing a cost

function. It's a user cost function. It

is the number of bad acts wherever that

whatever the t causes times 10 because

that was a unit cost we gave plus the

number of stops times four because

that's the the unit cost we gave. By the

way, one thing I should have elaborated

because I work in voice and we like

language and we like puns. So this whole

thing is called an outcome user cost

heruristic. So that spells the word ouch

and that is some expression of pain.

Yes, we are you know language nerds. So

these kinds of things amuse us. Um so

now let's consider that is the cost

function is to minimize the ouch. And

now um that let's see if uh I'm going to

bring up a tool. Let's see if this

works.

Where I have actually gotten or with one

of my coding assistants gotten uh an

interactive

um graph where we have actually plotted

those thousand points and as we vary the

threshold t you can see that the total

user cost here which is that function of

you know x * y + a * b actually changes.

So let's in the very beginning when we

said the system was just playing

the the cost across those thousand

points was 2100 or divided by a,000 is

2.1 ouch points per turn. Then we said

okay let's insert a stop behavior and

let's like wing it and say 65%. That's

when I want the threshold. If we brought

this up to 65 yeah that's better. Now

it's 1904 or 1.9 per turn, but it's not

optimal. As it turns out, if we do

actually um ask for the AI to solve the

uh the problem across this curve, it

turns out 43%. So I'll drag it now to 43

is in fact

the optimal

optimal point of t. This minimizes the

cost function. You can see it's the

lowest point on this graph down here to

1

27. So effectively we haven't changed

the accuracy at all. The system is not

any smarter in that sense. But with some

clever system behavior, conversational

behavior is what we'd call it and some

optimization and a cost function called

ouch. Um we have from the user's

perspective produced a more satisfactory

assistant. And this is not a trivial you

know accomplishment. Okay. Now, let me

go back to this. [clears throat] Let me

see if I can get this. Oh, great. Okay,

let's continue this. Let's continue this

with by now adding one more behavior.

Let's call it the confirm behavior. So,

there was play obviously, then stop,

confirm. Confirm is basically the system

after you said something saying uh kiss

play kiss by Prince or maybe play kiss

by Chris Brown. And uh you know the user

can either confirm like affirm it or

they can correct it. It is a different

kind of behavior and again this is kind

of how humans behave. Um that's

obviously the inspiration. Now if we go

back to our problem of optimization,

we have a third obviously um option

which is to confirm. And so this would

translate to two thresholds

um two thresholds which are separating

the distribution into three spaces of

stop, confirm and uh act. And the

question is now where are these T's? and

we have now given up on guesstimating

because we know it doesn't work. So

we're going to be a lot smarter and go

back to the concept of user outcome cost

and then you know use it go look for

some optimization in that graph. So

let's uh define what are the what are

all the possible bad outcomes that t1

and t2 um make for. So good you can see

my cursor. So uh of course any stops are

still bad. Then in the middle are

confirmations. Confirmations are bad

because they slow the user down. There

is a confirmation outcome called confirm

yes where they just affirmed it by

saying yeah or no where they had to

correct it. And going back to our

formula these outcomes are not equally

bad. And in fact, nobody will, I think,

argue here from a user's perspective.

Affirming, just saying yes is obviously

less painful than saying no and then

having to restate whatever it is that

you wanted in the first place. So now we

I've assigned values of two or six. And

again, I said it was a heristic. This

would be roughly the amount of time it

would take for the extra for the user to

get to the song they want. Saying

listening and then saying yes is like

two seconds. Um and then now

uh we restate the cost function for this

you know added behavior as this number

of you know bad type one times unit cost

bad type plus bad type two times unit

cost etc. And now we try to minimize

this user cost function and minimize the

ouch.

Yes, I'm going to keep doing that pun.

Um let's go back. So this is now the

interactive graph but

with

um the cost values the unit costs here

10264

and uh you know we're just going to ask

the AI to tell us here's the heat map

because it's now two dimensions saying

that the optimal values are 41 for the

the T1 and 49 for the T2 and if we

employed that then we would go to 1464.

Uh, by the way, whatever numbers I put

in here, like let's say I thought wrong

act was 20. It's really irritating and

painful and takes way longer to actually

correct it when you hear a wrong song.

That would change you know all these

numbers uh and the optim optimal point.

So again it is about how what is the

relative badness of these outcomes also

of course the distribution curves

naturally. Uh let's go back here. Okay.

So, um I'm gonna

speed up a little bit. Uh let's go back

here.

Presentation mode. Okay. So, what have

we shown that if we did the super naive

approach, it's 2.1 act and stop 1.9 then

1.27 then 1.26. We are able to bring

this with every added layer of

sophistication, adding more behaviors,

being smart about outcome, uh, user cost

and optimizing. Um, we have made a

tremendous difference without changing

the accuracy at all. Um, this was a

super simplified example. In real

systems, you're not going to have

obviously some offline decision

threshold or two. It's going to be a

real time, you know, learned decision

model. But the principle is the same.

And I believe this is uh scalable across

all voice AI surfaces. Obviously this is

a smart speaker but if we go across any

of these surfaces you will find the

equivalence. If we um we will find the

analogies with some differences but the

spirit and the I think the the gain will

be similar. So just for example in the

TV AI assistant space if you employ it

here it's going to you're going to have

the same thing when users express

intents like on TV it's you know open a

channel that's one of the most common

obviously um requests on a TV voice

assistant same thing you're going to

find you'll have exactly the same

approach but the difference will be

maybe in the the assignments of the user

outcomes because the UI and the

modalities are different when you have a

TV you have a multimodal interface where

choices can be shown. So instead of you

know asking did you mean ABC you know uh

news live by speech that you will the

system would display choices and not

just one it show ABC News live this that

would be the confirm step and if it's

visual and you can use your remote

control to select something it's less

pain so you would change some of these

values or if in fact launching the

channel would kick you out of your

current state then it would go in the

other direction than cost of you know a

bad act would go much higher. So it's

the same concept but in this new

modalities

um variables can change, values can

change, arguments can change but the

premise still holds and you can improve

from the user's perspective because

we're all about you know making humans

happy. Um you can make them happier and

this as I said in conclusion can be

applied across all surfaces. I did say

at the very beginning, just to recap for

us, that voice is great when it works,

bad when it doesn't. And as we get into

embodied AI, where these AI assistants

are taking actions, physical or even

digital, like making a phone call or

sending an email, it is getting more and

more difficult just to rely on accuracy

to improve user satisfaction. I believe

there's a whole knob the second knob

called smarter conversational behavior

under uncertainty

and um if we actually exploit that we

can uh very much help these AI systems

reach a acceptable user experience

otherwise I think this will continue to

be a bottleneck like a lot of things

will get better but if the voice

interface as experienced by user does

not improve it is going to be a a a

choke point. And um if you just remember

one word or two words from this whole um

presentation, it would be to minimize

the ouch of the experience. Um so thank

you. I'll stick around for questions if

you guys got any. Thanks a lot.

[applause]

>> [music]