How Uber Uses AI to Improve Deliveries

How can Uber deliver food and always arrive on time or a few minutes before? How do they match riders to drivers so that you can always find a Uber? All that while also managing all the drivers?!

Well, we will answer exactly that in the video...

References

►Read the full article: https://www.louisbouchard.ai/uber-deepeta/
►Uber blog post: https://eng.uber.com/deepeta-how-uber-predicts-arrival-times/
►What are transformers:
►Linear Transformers: https://arxiv.org/pdf/2006.16236.pdf
►My Newsletter (A new AI application explained weekly to your emails!): https://www.louisbouchard.ai/newsletter/

Video transcript

       0:00

how can uber deliver food and always

0:02

arrive on time or a few minutes before

0:05

how do they match riders to drivers so

0:07

that you can always find a uber all that

0:10

while soon managing all the drivers we

0:12

will answer these questions in this

0:14

video with their arrival time prediction

0:16

algorithm deep eta deep eta is uber's

0:20

most advanced algorithm for estimating

0:22

arrival times using deep learning used

0:25

both for uber and uber eats deep eta can

0:28

magically organize everything in the

0:30

background so that riders drivers and

0:32

food are fluently going from point a to

0:34

point b as efficiently as possible many

0:37

different algorithms exist to estimate

0:40

travel on such road networks but i don't

0:42

think any are as optimized as uber's

0:45

previous arrival time prediction tools

0:47

including uber were built with what we

0:50

call shortest path algorithms which are

0:52

not well suited for real-world

0:54

predictions since they do not consider

0:56

real-time signals for several years uber

0:59

used xgboost a well-known gradient

1:02

boosted decision tree machine learning

1:04

library xjboost is extremely powerful

1:07

and used in many applications but was

1:09

limited in uber's case as the more it

1:11

grew the more latency it had they wanted

1:14

something faster more accurate and more

1:16

general to be used for drivers riders

1:18

and food delivery all orthogonal

1:20

challenges that are complex to solve

1:22

even for machine learning or ai

1:25

here comes deep eta a deep learning

1:28

model that improves upon xg boosts for

1:30

all of those oh and i almost forgot

1:33

here's the sponsor of this video

1:36

myself please take a minute to subscribe

1:39

if you like the content and leave a like

1:41

i'd also love to read your thoughts in

1:43

the comments or join the discord

1:45

community learn ai together to chat with

1:47

us let's get back to the video

1:49

deep eta is really powerful and

1:51

efficient because it doesn't simply take

1:53

data and generate a prediction there's a

1:56

whole preprocessing system to make this

1:58

data more digestible for the model this

2:00

makes it much easier for the model as it

2:02

can directly focus on optimized data

2:05

with much less noise and far smaller

2:07

inputs a first step in optimizing for

2:10

latency issues this pre-processing

2:12

module starts by taking map data and

2:14

real-time traffic measurements to

2:16

produce an initial estimated time of

2:18

arrival for any new customer request

2:21

then the model takes in these

2:23

transformed features with the spatial

2:25

origin and destination and time of the

2:27

request as a temporal feature but it

2:29

doesn't stop here it also takes more

2:32

information about real-time activities

2:34

like traffic weather or even the nature

2:36

of the request like delivery or ride

2:39

share pickup all this extra information

2:41

is necessary to improve from the

2:43

shortest path algorithms we mentioned

2:45

that are highly efficient but far from

2:47

intelligent are real world proof and

2:50

what kind of architecture does this

2:52

model use you guessed it a transformer

2:54

are you surprised because i'm definitely

2:56

not and this directly answers the first

2:59

challenge which was to make the model

3:01

more accurate i've already covered

3:03

transformers numerous times on my

3:04

channel so i won't go into how it works

3:07

in this video but i still wanted to

3:08

highlight a few specific features for

3:11

this one in particular first you must be

3:13

thinking but transformers are huge and

3:16

slow models how can it be of lower

3:18

latency than xg boost well you will be

3:21

right they've tried it and it was too

3:23

slow so they had to make some changes

3:26

the change with the biggest impact was

3:28

to use a linear transformer which scales

3:30

with the dimension of the input instead

3:33

of the input's length this means that if

3:35

the input is long transformers will be

3:38

very slow and this is often the case for

3:40

them with as much information as routing

3:42

data instead it scales with dimensions

3:45

something they can control that is much

3:47

smaller another great improvement in

3:49

speed is the discretization of inputs

3:52

meaning that they take continuous values

3:53

and make them much easier to compute by

3:56

clustering similar values together

3:58

discretization is regularly used in

4:00

prediction to speed up computation as

4:02

the speed it gives clearly outweighs the

4:04

error that duplicates values may bring

4:07

now there is one challenge left to cover

4:10

and by far the most interesting is how

4:13

they made it more general here is the

4:15

complete deep eta model to answer this

4:18

question there is the earlier

4:19

quantization of the data that are then

4:22

embedded and sent to the linear

4:24

transformer we just discussed then we

4:26

have the fully connected layer to make

4:28

our predictions and we have a final step

4:31

to make our model general the bias

4:33

adjustment decoder it will take in the

4:36

predictions and the type features we

4:38

mentioned at the beginning of the video

4:40

containing the reason the customer made

4:42

a request to uber to a render prediction

4:44

to a more appropriate value for a task

4:46

they periodically retrain and deploy

4:49

their model using their own platform

4:51

called michelangelo which i'd love to

4:53

cover next if you're interested if so

4:56

please let me know in the comments and

4:58

voila this is what uber currently use in

5:01

their system to deliver and give rides

5:03

to everyone as efficiently as possible

5:07

of course this was only an overview and

5:09

they used more techniques to improve the

5:11

architecture which you can find out in

5:13

their great blog post linked below if

5:16

you're curious i also just wanted to

5:18

note that this was just an overview of

5:20

their arrival time prediction algorithm

5:22

and i am in no way affiliated with uber

5:25

i hope you enjoyed this week's video

5:28

covering a model applied to the real

5:30

world instead of a new research paper

5:32

and if so please feel free to suggest

5:35

any interesting applications or tools to

5:37

cover next i'd love to read your ids

5:39

thank you for watching and i will see

5:41

you next week with another amazing paper

[Music]