New SOTA Image Captioning: ClipCap

We’ve seen AI generate images from other images using GANs. Then, there were models able to generate questionable images using text. In early 2021, DALL-E was published, beating all previous attempts to generate images from text input using CLIP, a model that links images with text as a guide. A very similar task called image captioning may sound really simple but is, in fact, just as complex. It is the ability of a machine to generate a natural description of an image.

It’s easy to simply tag the objects you see in the image but it is quite another challenge to understand what’s happening in a single 2-dimensional picture, and this new model does it extremely well!

►Read the full article: https://www.louisbouchard.ai/clipcap/
►Paper: Mokady, R., Hertz, A. and Bermano, A.H., 2021. ClipCap: CLIP Prefix for Image Captioning. https://arxiv.org/abs/2111.09734
►Code: https://github.com/rmokady/CLIP_prefix_caption
►Colab Demo: https://colab.research.google.com/drive/1tuoAC5F4sC7qid56Z0ap-stR3rwdk0ZV?usp=sharing
►My Newsletter (A new AI application explained weekly to your emails!): https://www.louisbouchard.ai/newsletter/

00:00

we've seen ai generate images from other

00:02

images using guns then there were models

00:05

able to generate questionable images

00:07

using text in early 2021 dolly was

00:10

published beating all previous attempts

00:12

to generate images from text input using

00:14

clip a model that links images with text

00:16

as a guide a very similar task called

00:19

image captioning may sound really simple

00:21

but is in fact just as complex it's the

00:23

ability of a machine to generate a

00:25

natural description of an image indeed

00:28

it's almost as difficult as the machine

00:30

needs to understand the image and the

00:32

text it generates just like in text to

00:34

image synthesis it's easy to simply tag

00:36

the objects you see in the image this

00:38

can be done using a regular

00:39

classification model but it's quite

00:41

another challenge to understand what's

00:43

happening in a single two-dimensional

00:45

picture humans can do it quite easily

00:47

since we can interpolate from our past

00:49

experience and we can even put ourselves

00:51

in the place of the person in the

00:52

picture and quickly get what's going on

00:55

this is a whole other challenge for a

00:56

machine that only sees pixels yet these

00:59

researchers published an amazing new

01:01

model that does this extremely well in

01:03

order to publish such a great paper

01:05

about image captioning the researchers

01:07

needed to run many many experiments plus

01:10

their code is fully available on github

01:12

which means it is reproducible these are

01:14

two of the strong points of this episode

01:16

sponsor weights and biases if you want

01:18

to publish papers in big conferences or

01:20

journals and do not want to be part of

01:22

the 75 of the researchers that do not

01:25

share their code i'd strongly suggest

01:27

using weights and biases it changed my

01:29

life as a researcher and my work in my

01:31

company weights and biases will

01:32

automatically track each run the hyper

01:34

parameters the github version hardware

01:37

and osu's the python version packages

01:39

install and training script everything

01:41

you need for your code to be

01:42

reproducible without you even trying it

01:45

just needs a line of code to tell what

01:47

to track once and that's it please don't

01:50

be like most researchers that keep their

01:52

code a secret i assume mostly because it

01:54

is hardly reproducible and try out

01:56

weights and biases with the first link

01:58

below as the researchers explicitly said

02:00

image captioning is a fundamental task

02:02

in vision language understanding and i

02:05

entirely agree the results are fantastic

02:07

but what's even cooler is how it works

02:09

so let's dive into the model and it's in

02:12

our working a little before doing so

02:14

let's quickly review what image

02:16

captioning is image captioning is where

02:18

an algorithm will predict a textual

02:20

description of a scene inside an image

02:22

here it will be done by a machine and in

02:25

this case it will be a machine learning

02:27

algorithm this algorithm will only have

02:29

access to the image as input and will

02:31

need to output such a textual

02:33

description of what is happening in the

02:36

image in this case the researchers used

02:38

clip to achieve this task if you are not

02:40

familiar with how clip works or why it's

02:42

so amazing i'd strongly invite you to

02:45

watch one of the many videos i made

02:47

covering it in short clip links images

02:49

to text by encoding both types of data

02:52

into one similar representation where

02:54

they can be compared this is just like

02:56

comparing movies with books using a

02:58

short summary of the piece given only

03:00

such a summary you can tell what's it

03:02

about and compare both but you have no

03:04

idea whether it's a movie or a book in

03:07

this case the movies are images and the

03:09

books are text descriptions then clip

03:11

creates its own summary to allow simple

03:14

comparisons between both pieces using

03:16

distance calculation on bit differences

03:19

you can already see how clips seems

03:21

perfect for this task but it requires a

03:23

bit more work to fit our needs here clip

03:26

will simply be used as a tool to compare

03:28

text inputs with images inputs so we

03:30

still need to generate such a text that

03:32

could potentially describe the image

03:34

instead of comparing the text to images

03:36

using clips encoding they will simply

03:39

encode the image using clips network and

03:41

use the generated encoded information as

03:44

a way to guide a future text generation

03:46

process using another model such a task

03:49

can be performed by any language model

03:51

like gpt3 which could improve their

03:53

results but the researchers opted for

03:54

its predecessor gpd2 a smaller and more

03:57

intuitive version of the powerful openai

04:00

model they are basically conditioning

04:02

the text generation from gpt2 using

04:04

clips encoding so clips model is already

04:07

trained and they also used a pre-trained

04:10

version of gpd2 that they will further

04:12

train using the clips encoding as a

04:14

guide to orient the text generation it's

04:17

not that simple since they still need to

04:19

adapt the clips encoding to a

04:21

representation that gpt2 can understand

04:23

but it isn't that complicated either it

04:25

will simply learn to transfer the clips

04:27

encoding into multiple vectors with the

04:30

same dimensions as a typical word

04:32

embedding this step of learning how to

04:34

match clips outputs to gpd2's inputs is

04:37

the step that will be thought during

04:39

training as both gpt2 and clip are

04:41

already trained and they are powerful

04:43

models to do their respective tasks so

04:45

you can see this as a third model called

04:47

a mapping network with the sole

04:49

responsibility of translating one's

04:51

language into the other which is still a

04:53

challenging task if you are curious

04:55

about the actual architecture of such a

04:57

mapping network they tried with both a

04:59

simple multi-layer perceptron or mlp and

05:02

a transformer architecture confirming

05:04

that the latter is more powerful to

05:06

learn a meticulous set of embeddings

05:08

that will be more appropriate for the

05:09

task when using powerful pre-trained

05:11

language models if you are not familiar

05:14

with transformers you should take five

05:15

minutes to watch the video i made

05:17

covering them as you will only more

05:19

often stumble upon this type of network

05:21

in the near future the model is very

05:23

simple and extremely powerful just

05:26

imagine having clip merge with gpt3 in

05:28

such a way we could use such a model to

05:30

describe movies automatically or create

05:32

better applications for blind and

05:34

visually impaired people that's

05:36

extremely exciting for real world

05:38

applications of course this was just an

05:40

overview of this new model and you can

05:42

find more detail about the

05:44

implementation in the paper linked in

05:46

the description below i hope you enjoyed

05:48

the video and if so please take a second

05:51

to share it with a friend that could

05:52

find this interesting it will mean a lot

05:55

and help this channel grow thank you for

05:57

watching and stay tuned for my next

05:59

video the last one of the year and quite

06:01

an exciting one

06:06

[Music]