Full transcript
0:00platform AI for good advances solutions
0:03standards and skills.
0:07In today's AI [music] for good event, we
0:09are counting on you to use the live
0:11video wall feature to ask questions and
0:13post comments to help create an engaging
0:15session.
0:17Boost your [music] AI skills and elevate
0:19your career with the AI for good
0:21discovery badge. Receive verified
0:23digital credentials that recognize your
0:25engagement in [music] cuttingedge
0:26sessions. Exploring how AI is being used
0:29to tackle global challenges. To see if a
0:31session is eligible, [music] look for
0:33the blue badge banner on the right side
0:35of the video interface that activates at
0:37the end of each live session. Share your
0:39badge [music] on LinkedIn and across
0:41your professional profiles and check the
0:43AI for good events web page for upcoming
0:46eligible online sessions.
0:49Discovery badge sessions are part of the
0:51AI skills coalition, ITU's global
0:53initiative to strengthen credible AI
0:56learning pathways and expand access to
0:58highquality, trustworthy AI knowledge
1:00worldwide.
1:02We encourage you to stay until the end
1:04to chat, [music] connect, ask questions,
1:06and network with our distinguished
1:08facilitators and worldclass AI experts
1:11in the neural network. It is now time to
1:14kick off the session and welcome our
1:16first speaker. The floor is yours.
1:25>> Good afternoon everyone and welcome to
1:27this AI for good discovery session on
1:30machine learning for ICT infrastructure
1:32detection.
1:34Today nearly 2.2 2 billion people remain
1:38offline
1:40and children in low income and rural
1:43areas are disproportionately
1:45affected
1:46in large in large part because we simply
1:50do not have reliable up-to-date data on
1:54where schools are and what connectivity
1:57infrastructure already surrounds them.
2:01This gap is exactly what this initiative
2:04was created to close.
2:07Mapping points of interests, modeling
2:10the infrastructure around it, and using
2:12that evidence to plan and finance
2:15connectivity where it's needed most.
2:19Over the next hour, you'll see how ITU
2:23and the King Abdullah University of
2:25Science and Technology have turned that
2:27challenge into a scalable vision only
2:31pipeline,
2:32one that detects schools and cellular
2:34towers directly from satellite imagery
2:38and then uses ITU's own connectivity
2:41plan and platform to assess
2:45with engineering grade precision whether
2:48a school can realistically be connected
2:50to a nearby tower.
2:53We'll start with a s with a short
2:55introduction from Dr. Slim
2:58setting the stage for the partnership
3:00and the problem we're solving. From
3:03there, Zachary will walk us through the
3:06school detection side of the pipeline.
3:09How we go from raw satellite tiles to
3:12precisely located schools using transfer
3:15learning even in data scarce
3:18environments.
3:20Then Shandor will then take us through
3:23the tower detection side and importantly
3:27ITU CPP contribution
3:29the terrainaware line of sight model
3:32that moves us beyond simple distance
3:35estimates to a realistic assessment of
3:38connectivity.
3:40Paris will close out the technical
3:43presentations by showing how these two
3:45results come together. The consolidation
3:49step that turns two detection pipelines
3:51into an actionable prioritized map of
3:56likely connected and hard to connect
3:58schools demonstrated on real data from
4:02lysuto.
4:03We'll then come back to myself for a few
4:07closing thoughts and to open things up
4:09for your questions. So please do keep
4:12them coming throughout.
4:15With that over to you Dr. Muslim.
4:28>> Dr. Slim, please if you could come in,
4:32turn off turn on your camera and then
4:34>> Yeah, sorry. Sorry, I forgot to unmute
4:36and myself and remove the kind of mask
4:39for the the camera. So, thank you Wid
4:42for this very nice introduction. My name
4:45is Muhammad Slim Alwini. I'm professor
4:47of electrical and computer engineering
4:49at King Abdullah University of Science
4:51Technology Kos in Saudi Arabia. My area
4:54of expertise is wireless communication
4:56and satellite communication. Over the
4:58last 10 years uh uh a lot of my research
5:02efforts and the research effort of my
5:03group have been focusing on you know the
5:07general topic of connect the unconnected
5:09and actually we established few years
5:11ago UNESCO chair uh at CAST in my
5:15research group that kind of address or
5:17kind of conduct research in this general
5:19area of connected and connected and uh
5:22essentially as Wed mentioned we live in
5:25a world where we still have about 2.2
5:28billion people offline. Actually uh uh
5:32other kind of statistics are telling us
5:35that 3 million uh schools are
5:38unconnected and that means that about
5:40500 million kids are offline during
5:43study time. Uh so all of this has
5:47motivate us at kst to collaborate with
5:50ITU in particular with the team led by
5:52Dr. to lead to try to kind of joint
5:55effort and complement expertise to uh
5:59come up with an approach or kind of
6:01holistic approach to detect and
6:04connected school uh in the best possible
6:07way. So we are happy to uh that these
6:11efforts led to this kind of uh
6:14interesting uh uh joint publication and
6:17uh uh it's our pleasure uh during this
6:20uh uh you know hour uh to present
6:24collectively cast researcher ITU
6:27researcher our uh joint solution. Thank
6:31you again wid for giving us the
6:33opportunity to work on this very
6:35relevant problem and I hope that uh the
6:38presentation that we will share with the
6:39audience will be beneficial for
6:41everyone. Back to you.
6:44>> Thank you very much Dr. Slim. Uh now uh
6:47we're going to uh move on to our first
6:49uh speaker. Uh so uh Zachary
6:54um let's suppose we are in a new region
6:56where no data is available to train a
6:59model. How could we apply apply your
7:02approach in this kind of a setting or
7:04kind of like a problem framework?
7:08Over to you Zachary.
7:10>> Thank you Wid. Thank you everyone for
7:12being here. So we're going to start to
7:16answer this question by first looking at
7:19the first foundation
7:22of our school detection model.
7:26So in this whole pipeline of school
7:30connectivity detection, first we need to
7:33know where the schools are. So we
7:36develop a whole pipeline, a whole
7:38computer vision models to know where the
7:41schools are.
7:44And basically our objective here in this
7:47components of the of our approach is to
7:52build a detector which means a computer
7:55vision model which is able to detect
7:58where school where schools are and to be
8:02able to distinguish schools from other
8:05buildings like here in the in this
8:08image. But the problem to be able to
8:11train such a model we need labeled
8:14images which means satellite images with
8:20containing schools and these satellite
8:22images that are containing schools we
8:24need a bounding box over the schools.
8:28We need thousands of image like like
8:30this. But the problem that we faced is
8:33that such data is not available. All
8:36what we have is just single latitude
8:40longitude points per school which means
8:43we have a big CSV file that's containing
8:46a lot of schools with their positions
8:49longitude and latitude but we don't have
8:52bounding boxes around the schools or the
8:55images.
8:57So this is a a missing piece in in our
9:00pipeline. One can say okay let's get
9:05this positions and extract the satellite
9:10images the corresponding satellite
9:12images from these positions and then
9:15label them manually I mean drawing
9:18bounding box around each school for all
9:20the satellite images okay this is
9:23possible but the problem is that this is
9:26very costly and it's slow and the
9:29problem also if in the ja projects
9:32We aim to connect every school in the
9:34world. We need to do this for many
9:36regions in the world. So it's almost
9:39impossible to do such a work.
9:43So this push us to think about an
9:46automatic way to label and create our
9:51first database that we can train our
9:53models on. And the idea basically it's
9:57very simple. So we know the the school
10:01position. We can extract a first image
10:06a first subate image of the school
10:08position. After this we use a model
10:11called leng sam which is trained on many
10:14images and its principle is simple like
10:18you give the model the link some model
10:21uh an image and also you give him a
10:25prompt which is like an object that we
10:28want to detect in this image and this
10:31link some model returns masks of the
10:35corresponding objects in the image that
10:37you give
10:39And basically this is what we used. We
10:44take our image, we give it to Lam model.
10:48We give him also some prompts like
10:51building, roof, school, etc. to get
10:55candidates, objects, masks. And then
10:58once the model give us the masks, we
11:02filter them using some metrics that we
11:06have in our algorithm.
11:08And then we keep the closest mask to our
11:12school position because we already know
11:15the school position. And then once we
11:18have the selected mask, we just draw the
11:22bounding boxes and it's done. So doing
11:27this work for all the images that we
11:30have in an automatic way enables us to
11:34create our large autole data sets with
11:38zero manual annotation. And actually
11:41this is very useful in the context of
11:44the JGA project where scalability is a
11:47huge concern.
11:50But the problem with the the data set
11:53that we just created is that it contain
11:56some noise because the process the whole
11:58process is automatic. So we have some
12:02noise in the segmentation we may have
12:04some noise in the in the filtering. So
12:08we cannot treat this auto label data set
12:12as a ground truth.
12:14But how do we still end up with a
12:17nitrates thrust detector?
12:20The answer is very simple.
12:23We take our auto automatically labelled
12:28data set. We train a first detector like
12:32yolo for example in our case on this
12:34data set. And the objective of this
12:37first step or of this pre-training is to
12:41learn a representation of how schools
12:44look like.
12:46And then we take this pre-trained model
12:49and we create a small data set manually
12:53labeled like we manually label just a
12:55hundred of images. For example, in our
12:57case, we take images and we draw bounded
13:02boxes around the schools in the images
13:04that contain schools. And then we take
13:08our pre-trained model and we fine-tune
13:10it on this small label data set. And by
13:15doing this two-stage training, we get a
13:18very good model, a very good detector,
13:21which is really production detector with
13:24very strong results. And with just a
13:28minor efforts in labeling, we just
13:31labeled 100 images. And the results, for
13:34example, in the case of the US data set
13:39with only 100 labeled images, we were
13:42able to have 86 per 8.8% 8% of detection
13:48accuracy and 85.5%
13:54of recall which are very strong results
13:58in the in the context of object
14:01detection
14:03and also the gain of our method in
14:06comparison to when we train only on
14:09automatically labelled data set we get
14:12only
14:1346%
14:15of detection accuracy which means that
14:17we have a gain of this 40 points of
14:20detection accuracy just by this by by
14:24our approach which use only 100 labeled
14:28image and also
14:31in comparison with when training on only
14:36manually labelled data set we have a
14:38gain of plus 55%
14:42of detection accuracy
14:45So basically the conclusion here is
14:48neither ingredients is allow is enough
14:52alone. Scale without correction or
14:55correction without scale without scale
14:57they run 100 they turn 100 image into a
15:02production grade detector.
15:05So here this approach this two-stage
15:08approach works well in the context of US
15:13here for example where we have a lot of
15:16initial points initial data to start
15:19with.
15:21But what about an underserved region
15:25where we barely have some local data to
15:29start with? For example, in the context
15:32of Leoto where we don't have that much
15:36data, we just have few hundreds of of
15:39data. We cannot build with that uh uh an
15:44automatically labelled data set. How to
15:46do in this context? And the question is
15:49can we adapt the model that we trained
15:52in the context of US to this new region
15:58with only small amounts of data. The
16:01response is yes and this is via transfer
16:04learning. So basically here in the
16:07context of leoto we only have few images
16:12and few few hundred of of images. We
16:17take them we create a small manual label
16:20data set as as we did in the second step
16:23of the US data set and then we take our
16:28model which is trained on the US and we
16:31finetune it to this new context. We
16:33finetune it on this manual label data
16:35set that we created for Leoto and
16:38basically doing this
16:40transfer all the knowledge that the US
16:43model has to the new context which is
16:46Leoto and basically we have very strong
16:50results also for risoto by doing the
16:53transfer learning 91% detection accuracy
16:57and 95%
16:59of recall.
17:02So basically a detector which is trained
17:05on uh a richly available imagery become
17:10a strong usable starting point for any
17:13new country. No need to relable
17:17thousands of imagentions each time. And
17:21basically all this this pipeline for
17:23school detection that we developed is
17:27scalable and it requires minimal efforts
17:31in labeling just as you as you've seen
17:35100 images to to labels to to label and
17:39it delivers strong results making it
17:43particularly well suited to the
17:46context where efficient scaling is
17:48crucial.
17:51the you wanted.
17:53>> Thank you very much, Zachary, for giving
17:56us a sneak peek under the hood uh on how
18:00uh you could use uh the raw data uh to
18:03locate the school leveraging transfer
18:06learning. Uh now moving on to um our
18:10next uh pipeline which is the tower
18:12detection. Uh, Shandor, um, how could
18:17you apply your tower detection model if
18:20you were to, uh, use it in a different
18:23country, for example? Uh, over to you,
18:26Sandor, please.
18:32>> Thank you, Wid.
18:34Uh, thank you everyone being here. So
18:42I will start with uh uh with with our
18:45base model we created uh uh before
18:51uh we choose uh
18:54YOLO as uh as our framework uh to work
18:58with which is really good u Python based
19:03and uh you can include it in your
19:06scripts and uh you can use it on on
19:10uh cloud environment and it has a
19:14special feature uh which is oriented
19:17bounding box where uh you can do the
19:22your labeling different than uh uh
19:27uh you use it uh normally
19:30please uh next slide.
19:33So this oriented bonding box uh is
19:38different from regular bonding box which
19:40is uh the the side of a bonding box is
19:44actually parallel uh to the axis. And
19:48with this you can uh use more narrow and
19:52uh tighter bounding box for your
19:54objects. And uh this results uh less
19:59distraction background for the patterns
20:03uh of uh what uh the model has to start
20:08uh uh calculate and stores weight. So uh
20:13we found that uh this is very useful for
20:16uh for uh
20:19detecting cell towers
20:22and uh after that we had first uh u
20:27train model which uh was started with
20:30with a yellow uh pre-trained model. Uh
20:35unfortunately in that model there that
20:38wasn't uh any uh class for cell towers
20:42but at least those uh models were
20:44trained on aerial images.
20:47Uh
20:49we started with that uh with a set of uh
20:52uh images and u
20:55ground truth data from uh Mozambique
21:00and uh built a model
21:03and uh in that model we used uh two
21:07types of cell towers and uh their
21:10shadows because cell towers are uh
21:14thrust like a very thin objects. uh
21:17protruding perpendicular from the ground
21:20and uh very often those are not visible
21:24uh uh from the uh photos taken from up
21:30and uh we hope that uh
21:34the shadows they cast uh can be useful
21:38for detecting those cellars.
21:42Uh
21:43we we had a really good uh measures in
21:46in u
21:49in Mozambique
21:51and uh [clears throat]
21:54we moved uh to a different country uh
21:58Leoto which is relatively close but uh
22:02we found that uh there's no cellar type
22:06two so we skipped that part and uh
22:10retrained our model with a new images
22:14and fine-tuned the model to fit to uh uh
22:19leoto environment. Leto was very
22:22different even if it's uh really uh
22:27relatively close to Mosmbique
22:30at least in the same region uh the
22:34background the vegetation is very
22:37different which uh even with
22:41smaller bounding boxes uh can'ts so
22:47uh we have to retrain our model this is
22:50the result of the train. Uh in in uh
22:56in computer vision you have the same uh
23:00metrics but you have in general uh in
23:06um in machine learning uh other machine
23:11learning models but uh you have a
23:14special uh uh metric which is different.
23:17So actually you have a precision recall
23:22you have the F1 score for the uh by the
23:26end of the training.
23:28Uh F1 score is the harmonic means of the
23:32precision and recall making balance uh
23:36balanced assessment of the model put
23:39into one uh uh metric.
23:44And next next please the this matrix
23:48what I'm uh talking about here is a
23:51intersection over union
23:54which uh measured that uh quantifies the
23:58overlap between uh predicting bounding
24:01box and the ground bounding box and it's
24:05very important uh of the object
24:08localization.
24:10So it's uh
24:12it's a special metric that we use for uh
24:16uh use in
24:18object detection and uh this metric can
24:23be used for uh calculating mean average
24:26precision at different thresholds and uh
24:30and uh correct localization
24:32uh
24:34for example uh threshold from 50 to 95
24:39five. And uh uh that uh does uh
24:44important uh
24:49information about your model uh
24:52performance.
24:55Yeah, next slide please.
24:58Yeah, we we had to uh and uh next slide.
25:02So uh we have our model and what we see
25:05here is uh the
25:08overall flow or uh
25:14overall workflow uh of the school
25:17detection and uh cell detection that we
25:23have schools input. Uh we want to find
25:27uh cell towers around schools and uh as
25:33a final step we have to check whether or
25:36not uh the schools can be connected to
25:39those cell towers that we found.
25:42Next slide please.
25:45Here's the input our case study uh prof.
25:53So on the left side you see the scores
25:57which are uh the confidence threshold uh
26:01greater than u uh 60%.
26:05And uh to save resources
26:09uh we set up a
26:12boundary
26:14uh initial uh
26:17constraint of 1 kilometer
26:20uh
26:22uh boundary for searching for cell
26:24towers. So what what you see here is uh
26:29a buffer around those uh schools
26:33school candidates and uh and uh and the
26:38generated grid that uh that is the
26:41outline of the um actual uh images what
26:47we have to uh search for uh cell towers.
26:52So those images are 15 uh cm per uh per
26:59pixel uh resolution
27:02and uh 1,24
27:05by 1,00 by 24 uh uh size
27:11and we uh run our model and uh detect
27:16cell towers uh uh on those images. The
27:20outcome is uh the result is uh row uh
27:26data, the row bounding boxes of those uh
27:31findings uh possibly satires and shadows
27:35of our classes. And uh in in the in the
27:41detection uh script uh we turn those
27:45bounding boxes into
27:48um into vector layer. And on that vector
27:53layer we perform uh post validation and
27:57u and um u some work uh with uh uh
28:03combining these bounding boxes into uh
28:07one point. So when we find a cell tower
28:11and uh the same time or type of shadow
28:16close enough to each other then we
28:19combine those and uh keep it that point
28:23as a result and with the result uh we
28:28have uh
28:30confidence and uh we can use auto
28:34validation uh for cells with they have
28:38uh those uh records we they have really
28:42high uh confidence value. We can
28:45autovalidate it as set to hours. those
28:48records we which are below some
28:51threshold we can throw it away and we
28:54have to check that manually the
28:55remaining setters but
28:58we are in in a GS uh software now
29:03because uh the row pixel values uh uh
29:07were transformed into uh GS coordinates
29:11coordinates and we can uh easily zoom to
29:14those places and manually validate uh
29:18our cell towers and in this case the
29:21result uh is uh this uh five cell towers
29:26in the buffer zone.
29:28Uh next slide please.
29:32Now we have uh to do uh the uh
29:36visibility check and um
29:40uh
29:42we have have to be aware of uh uh
29:47obstruction because we want to uh get
29:50the clean line of sight between the our
29:53schools and cell towers.
29:56Next slide please.
29:59So uh luckily ITO has a product for this
30:04and uh
30:06this is u sorry
30:11uh
30:13this product is uh
30:17connectivity planning platform.
30:20Uh and uh it has a web interface where
30:24you can upload your uh point of interest
30:27data and you uh and uh infra your
30:30infrastructure data and um run uh
30:35proximity analysis on that. Uh actually
30:38you can uh do a lot more in uh CPP. You
30:44can uh do other connectivity uh
30:49analysis uh for example uh fiber path
30:53analysis along road sides and you can
30:56even put uh cost uh data
31:00opex and copex uh data if you have and
31:05uh get the result of of costs and u
31:12for for now we just used the simple uh
31:16uh pointto-point analysis uh for this uh
31:19research
31:21and what you see here the result please
31:26the next slide
31:31uh the result shows the distances uh uh
31:36for the for each school to the closest
31:39cell tower The
31:43the very good thing here is uh that CPP
31:47takes account uh into account the
31:50elevation model. So what you get here
31:53the rear line of sight between these two
31:56points and uh uh uh and and and uh in
32:02this case I checked it's very
32:04interesting I checked uh just the
32:07distances
32:09uh without uh terrain model and it
32:13turned out that school number seven is
32:16closer to cell tower number 26 than the
32:21actual result
32:2421 and uh I checked that uh the profile
32:29between those two and it turned out that
32:32there is a blockage uh between those two
32:35points and the real uh clear line of
32:39sight is is the the final result what we
32:42get got here.
32:45Next slide please.
32:47here uh this uh slide just shows what uh
32:52pattern uh we have to find with uh with
32:56this model. So a very delicate uh find
33:00pattern what we have to locate. So it's
33:04not really an easy task for us.
33:07Next slide please.
33:11So uh just uh uh
33:15we Actually
33:18it it can be used for uh very well we
33:23found that it it can be used for uh for
33:27having an insight of of gap analysis and
33:31uh the ability to connect to uh those
33:34points together and uh connecting to the
33:37network. So next slide please.
33:42Uh we think uh that uh this approach is
33:46very good and uh
33:49uh we can apply this but uh but we think
33:54that we have still have to uh uh keep it
34:00the the end of the process in hand and
34:03uh check our results uh uh and uh
34:07validate our results because uh if if
34:11you just put the threshold uh too high
34:14and the precision too high uh then we
34:18are missing uh actual uh valid cells.
34:23So because of this we chose uh to keep
34:27it low. I mean uh keep uh uh find
34:31everything and uh keep the recall
34:35focusing on the recall and uh just throw
34:38away the thresh uh later.
34:44So
34:46lastly answering your question uh
34:49sometimes we hoping that uh we can build
34:53a
34:55a model with uh more robust model but we
34:59can use out of the box but uh for that
35:03we need lots of data to put in and
35:06training.
35:08Thank you.
35:12Uh thank you very much uh Shandor for
35:14giving us this uh uh technical deep dive
35:17on uh on the model and uh particularly
35:20also like how we could use the
35:22connectivity planning platform uh to uh
35:25um kind of plan and design around uh
35:28networks and how we we can extend uh and
35:32apply these models to extend uh
35:34connectivity. Uh now um uh on to FYS uh
35:40where uh hopefully like we can uh Fris
35:44could walk us through like how we could
35:45bring all this together. Uh how can we
35:49consolidate all these uh steps in for
35:52these two detection pipelines into
35:54actionable uh um
35:58actionable outcomes. So, um, Ferris, um,
36:03can you, um, kind of like, um, take the
36:06time to discuss and summarize the big
36:08picture here, uh, and point us at the
36:11limitations and the future directions,
36:13uh, for how we could use these models
36:16along with CPP. Over to you, Fis.
36:20>> Thanks a lot, Wade. Uh I think Zakaria
36:23and Sandor already covered the technical
36:25parts really well. So I'll just uh take
36:28a couple of minutes to summarize the
36:30bigger picture. Uh yeah this was a
36:32fruitful collaboration and the joint
36:34effort with between the ITU and K. This
36:38research started with a very simple
36:41question but an important one which is
36:44can we locate schools accurately from
36:46satellite images without needing a huge
36:49amount of labelled data. And fortunately
36:52the answer was yes. And we were able to
36:54train school detectors with a few
36:57labels. Even more importantly we found
37:00that those models generalize easily with
37:03transfer learning and so on to different
37:05countries with very limited uh
37:07additional labor.
37:09So the follow-up work to this went a
37:13step further. uh instead of only asking
37:16where are the schools we also started
37:18asking what infrastructure is around
37:20those schools basically the part
37:22discussed by sand. So we are now not
37:26only locating the schools but also
37:29locating the nearby cellular towers and
37:31combining that uh information
37:35to actually have a complete and
37:37practical view of the connectivity
37:39situation. So for me these two works are
37:43complimentary and actually they are
37:46really strong beginning. So basically
37:49it's a a very strong step into uh
37:52bridging and uh achieving better
37:55connectivity in remote areas. However
37:58the real problem is not only to find the
38:01unconnected schools for statistical
38:03purposes. So not just just getting these
38:06just to know how many schools are
38:08unconnected. I think the bigger picture
38:10and the vision is to eventually be able
38:13to connect those schools in the future
38:16hopefully.
38:17So of course there are still many things
38:20we can improve uh to achieve that and
38:23even in the current methodology there
38:25are still some uh forms of limitations
38:28that could be addressed in future work.
38:30For example, some choices in the current
38:32pipeline uh are fixed by design.
38:35basically the 1 kilometer radius to look
38:38for nearby towers. This was like a
38:40simplification that it could depend on
38:44the country, the operator, the terrain
38:46and many other factors. So if we later
38:49could or we could collect more data on
38:52which schools were actually connected or
38:54not and under what conditions then some
38:58of these choices could be actually um
39:00more grounded and could be potentially
39:02calibrated or even could be learned by
39:06some uh better machine learning
39:08algorithms. And of course another point
39:12is that cellular cellular towers are
39:14only one option for to provide
39:17connectivity for those school. So for
39:20some schools it could be possible that
39:21satellite communication or maybe fiber
39:24or other combination of technologies
39:27might make it more sense for those uh
39:29remote schools. So the question becomes
39:32actually uh broader not just is this
39:36school uh close enough to a cellular
39:38tower or uh is this connected or not
39:42connected. It will actually grow into a
39:44broader question on what is the best
39:48possible way or the cheapest way or the
39:50most efficient way to connect uh a
39:53certain school. And once we ask this
39:56kind of questions, we naturally go into
39:58another problem uh building on top of
40:01what we achieved so far basically by uh
40:05prioritizing and optimizing these kind
40:08of policies and this can be
40:11a problem of sequential optimization
40:13potentially reinforcement learning which
40:15could become an a followup of those two
40:18works. So yes uh there is still a lot of
40:22work ahead but I think these works are
40:24good streets toward the real objective
40:27connecting under service schools
40:30especially in remote regions and
40:31hopefully improving the education and
40:34creating more equal opportunities.
40:36>> Uh thank you.
40:46Thank you very much uh uh Faris for uh
40:49giving us like the overall picture and
40:51bringing it uh together. Uh now um uh I
40:57would like to thank our speakers for all
40:59this technical uh uh uh presentations
41:02and um I would like to open the floor
41:05for questions. I think we have few
41:07questions um uh sent by the uh by the uh
41:12audience. Uh I could probably start with
41:15the first one and uh see who from among
41:19our panelists could uh wanna uh um want
41:22to provide answers. So uh Tuesday asks
41:27uh uh how do you secure
41:30uh this kind of pipelines from bad
41:32actors? Um anyone um from our esteemed
41:37speakers want to take that?
41:47Maybe Professor Slim if you want to come
41:49uh forward on how we could protect this.
41:54I I I honestly didn't understand exactly
41:57what you mean protect what the data you
42:00mean or the like the procedure or the
42:02approach. I'm not sure what what what is
42:06meant here. Uh no, I think my
42:08understanding of the question is is
42:10probably and uh I could be stand
42:12corrected. Uh how um can you use these
42:17models? Can you make sure that the use
42:19the use of these models is not uh
42:22malicious?
42:23>> So that's what you mean. So yeah. Okay.
42:26So you know I will have to be honest in
42:29a way a generic answer to this question.
42:31Uh anything that we develop may have
42:33dual or triple use here. Obviously we
42:36are in this context in AI for good not
42:40only in this conference but actually in
42:42our research activity. Yes our objective
42:44is to de school maybe hospitals actually
42:47that's another activity that we thought
42:49about uh but of course uh you know
42:52people can focus on other type of target
42:55or other kind of things they want to
42:56detect. But uh yeah here the effort has
42:59been more to develop this approach at
43:03least from a from a detection
43:04perspective on looking at what kind of
43:08look like a school and u
43:12and that's that's where that's what we
43:14are focusing on in this paper and this
43:16research.
43:19>> Thank you very much Dr. Steam for those
43:21uh um clarifications.
43:24Uh on to the next question from Ursula.
43:28What rules uh do you use for uh machine
43:31learning for ICT infrastructure
43:34detection? Uh I think this is referring
43:37to the general rules um and guidelines
43:40for developing our models. Uh I don't
43:43know if u maybe Shandra or Zachary you
43:46want to take um you want to take that
43:48one.
43:54Uh okay. Uh basically uh what we are
43:57developing is computer vision models
44:00that are open and actually all our
44:02models that we trained are openly
44:05available for everyone to to use and to
44:08fine-tune and of course for the JGA uh
44:13projects. So, uh, basically it's it's
44:17open for for everyone and anyone can can
44:20can use them as, uh, as as they want.
44:25So, I don't know if it's
44:28if it's uh clarified more or
44:32needs more more clarification for the
44:35question.
44:35>> I think it's it's fairly clear. Um, so
44:38then that's a good segue to the next
44:40question. Uh how do you validate the
44:43results? Uh and how would you locate
44:45rural schools that may not be listed? Um
44:49are there on the ground efforts like
44:51boots on the ground or is it completely
44:54remote? Um meaning like um
44:58yeah
45:01I don't know fires maybe you have the
45:03overall picture here on like how the
45:05validation works and how um
45:10yes so definitely uh we can detect
45:13schools that are not on the list that
45:16are not detected. So the the approach is
45:18in at least in the inference time we can
45:20definitely use the method to detect
45:23schools or serial towers that we don't
45:26uh actually know or have data about.
45:30However, for the training
45:32uh we had like some ground truth to
45:35validate the approaches. For example,
45:37for school detection, we have a uh
45:40ground truth manually labeled school uh
45:44locations and images which for which we
45:47test the actual outcomes of these
45:49machine learning models. But this is for
45:51the to validate the method for the
45:54inference time. We expect these methods
45:56to generalize on other uh places even
46:00beyond the distribution of the of the
46:04data set that we just
46:08>> Thank you Fris. Um
46:11>> the next question is um about asking if
46:16it's is it globally or locally recorded
46:19and available data. Uh so maybe I can
46:23take that one. uh for the data uh I
46:26think we work with what what is
46:28available uh the base is satellite
46:32imagery so whatever we're able to uh
46:34extract any satellite imagery we're able
46:37to kind of extract the data
46:40um we are uh there's also another
46:43process where we kind of u uh uh scrape
46:47the web to uh find of like available uh
46:51geo tagged uh data where we can extract
46:55further information. There's a process
46:57around that. Uh usually when we do that
46:59as ITU, we notify the owners of the data
47:02that a certain uh amount of data has
47:05been collected. Uh and then we we try to
47:08specify uh the the goal and the and um
47:13and the um uh and the aim behind uh
47:17collecting such data. uh and then uh
47:20that exchange in official letters uh and
47:23then the owner of the data can specify
47:25how that data can be used uh can it be
47:29exposed publicly uh should I should it
47:31be restricted etc. Um so uh there are
47:36different ways to collect global and
47:39local data. Uh we work also in very
47:42close partnership with the with the
47:44regulators and with the administrations
47:46and with the operators uh in close
47:49collaboration to uh understand uh what
47:52we can collect and how we how we can uh
47:56the modality of how we could use uh what
47:58we collected. Um yeah so that's on the
48:03data. Uh there is another question uh
48:05from say here. Can we extend the
48:08outcomes to further evaluate post-
48:11disaster impact assessment and advise
48:14concretely on enhancing resilience of
48:17infrastructure?
48:19Um Dr. Sleim, would you like to take
48:21that one on resilience and maybe the
48:24disaster aspect of using uh these
48:27pipelines?
48:33Uh so uh
48:37I mean the I think this is maybe Shandor
48:41can answer the question in in a better
48:43way in terms of the uh how resilient is
48:47uh the approach he developed with the
48:50ITU team regarding identifying tower. So
48:53I I assume here the question is around
48:56if a disaster happens and the tower are
48:58broken or are damaged are we still able
49:00to recognize them? Uh this I'm not sure
49:03how resilient is algorithm to these kind
49:06of perturbations. So maybe Sandor can
49:09answer this question you know. So,
49:11Shannor, could you please come in and uh
49:14tell us more if we could use your models
49:17uh for disaster impact assessment and uh
49:21also for pointing at uh resilience of
49:25infrastructure or single point of fails
49:28etc. Uh over over to you Sando.
49:31>> Thank you. Uh I I think um
49:36eventually we we can use so maybe not at
49:41this po point because uh we we have to
49:45make uh create a more robust uh uh
49:48model. We want to uh include as many uh
49:53places uh
49:56all over the world. uh so to be able to
50:01use it uh out of the box but uh
50:04definitely we can use it for uh uh
50:08finding the gaps. So if if there is a if
50:13you can cannot find a cell tower
50:16somewhere where uh
50:20because of that cell is not uh visible
50:24because it's broken uh
50:28then definitely it shows a gap so for
50:32gap analysis uh we can use
50:36>> yes
50:36>> so yeah so well now I can pick it up if
50:39you don't mind.
50:41>> Yes, please, please.
50:42>> Yeah. So, going so yeah, I think Sandor
50:43is kind of the was expert on that
50:46particular point. But on the other
50:47issues that were raised, if I understand
50:50the question properly in terms of why
50:53not contacting the countries to get the
50:56data and also what kind of extra risk
50:59this leads to. So I think maybe we need
51:01to kind of reemphasize uh what we are
51:04trying uh to to to solve here. Uh and
51:07actually thank you. I think wed is the
51:09one who exposed me to this problem. So I
51:13mean people agree think that every
51:15single country has a perfect database of
51:17all their school with their location
51:19with the number of students. That's
51:21probably true for developed country very
51:24well organized administrations.
51:26Unfortunately this is not the case
51:27everywhere. There are big countries.
51:29There are countries that roughly have an
51:32idea where schools are but don't have
51:33the exact location for all schools. as
51:35strange as this seems but that's kind of
51:38a reality. So actually we are going to
51:40help these countries to basically
51:42identify their school with the tool that
51:44has been developed because actually they
51:46don't have that information themselves.
51:48Of course, if the information is
51:49available, that's kind of redundant in a
51:51way because you know if it's known,
51:54there is no extra information provided.
51:56But here, this tool is going to help
51:58countries where these database are not
52:00up to date or are basically kind of not
52:03very well organized. And this AI tech
52:06approach is going to help them identify
52:10school and then hopefully identify and
52:12connected school. Now, is there a risk
52:14to this? Again, I'm not sure if I
52:16understand the question.
52:18uh uh fully but uh you know we we we are
52:21just here accessing data making some
52:24estimate we may be right uh uh most of
52:28the time we may have some mistakes but
52:31uh I don't see any anything risky with
52:35that to be honest it's just like an
52:37assessment and uh helping countries
52:40administrations identifying location of
52:42schools and with that maybe allocating
52:44resources or have a better kind of
52:48idea where student and children and kids
52:52are located and maybe develop
52:55accordingly some other kind of followup
52:58or uh suitable program to support this.
53:01So uh we we we we are not talking to
53:04students, we are not interacting with
53:06the uh with the community there. Uh we
53:09are not affecting the communication
53:11infrastructure. We are just looking at
53:13images and based on the images we are
53:15trying to make these estimates.
53:18>> Uh thank you Dr. Slee and echoing uh
53:20your comment here. We are we are working
53:23in full transparency and actually we
53:24would like uh to increase the
53:27transparency about like infrastructure
53:29data so we can attract more uh players
53:33or uh more uh um uh uh partners in
53:37solving and tackling this uh uh
53:40universal and meaningful connectivity uh
53:42challenge. um uh some some of the
53:46countries um there are a tremendous
53:49number of schools like I can probably
53:51like mention the example of Brazil where
53:53they there are more than 400,000 schools
53:56across the the the national territory.
53:59Um most of those schools cannot be
54:02identified uh um uh by a human crew. it
54:08would the cost of of such an endeavor
54:10would be uh would be so high. So these
54:15tools that we're developing and these
54:17models would come in a handful handy uh
54:21for to tackle such such issues and
54:24particularly schools in remote areas
54:26that are very difficult to uh locate and
54:29to reach. Um and as Dr. uh uh Slim was
54:34mentioning uh we're we're not uh really
54:37um um uh impacting the content piece and
54:42like what gets delivered on this
54:44infrastructure in terms of content is
54:46another as a subject for another debate
54:48I think but in this case our problem is
54:51how we can uh extend the connectivity
54:55[clears throat] uh whatever comes on
54:56that connectivity whatever gets
54:58delivered on that connectivity is a a
55:00different uh it's a different issue
55:03Um and that's in u in in in answer to
55:06the question about like the risk uh for
55:09children. Um I don't think a
55:12connectivity by itself or the
55:14infrastructure of of the connectivity
55:17um presents any risk for for children or
55:20anyone. Um uh I think I will uh be
55:24asking um I know there are many other
55:27questions u on on the chat. uh um maybe
55:31we can take those offline but I would
55:33like to quickly give uh uh uh our
55:36speakers um uh a round of like uh maybe
55:40each one minute as we are uh short in
55:43time uh to uh to close and wrap up. So
55:47I'll start with you uh Dr. Slim for uh
55:49closing remarks.
55:53So thank you again wid for basically
55:57setting up this collaboration and for
55:58kind of formulating the problem and kind
56:00of exciting us as a team uh jointly
56:04working between kos and ITU to address
56:06this problem. We hope like mentioned in
56:08one of the last questions to generalize
56:11this approach to other type of useful uh
56:14uh essentially uh environment in
56:17particular uh probably hospitals uh or
56:22clinics that would be actually quite
56:24useful to to make sure that actually
56:26they [clears throat] are designated and
56:27they are not targeted for instance in in
56:29war type of situations. So uh uh we we
56:32we kind of uh believe that the approach
56:35we developed can be extended to other
56:38scenarios and we hope we'll have other
56:40opportunity to collaborate with ITU and
56:42generalize this approach to these kind
56:45of other useful scenario. Thank you
56:47again and back to you.
56:49>> Thank you very much Dr. Sle and we value
56:51very much uh our collaboration and
56:54looking forward in continuing to uh
56:56working on these uh challenges. uh uh
57:00Zachary if you want to uh provide quick
57:03uh closing u statement or remarks.
57:07>> Yeah, thank you Ali. Thank you professor
57:09slim for giving me this opportunity to
57:11work on uh this big project. Uh and
57:15thank you everyone for being here and
57:16for asking your questions. Um,
57:20you can still ask other questions uh and
57:24ask me personally or ask uh the other
57:27members. If you have other questions,
57:30don't hesitate to reach out to our
57:32papers and if you have any other
57:34questions, you're more than welcome.
57:36Thank you so much, everyone.
57:39>> Thank you, Zakaria. Uh, appreciate your
57:42technical uh presentation and insights.
57:45Uh, Shandor.
57:48>> Uh, thank you Balib. Uh, I got a few
57:52questions I think. Uh,
57:55it's just for me. Uh, and it was also
57:59your question. uh how how can we adapt
58:04this uh worldwide this model and how can
58:08to transfer to other countries and uh I
58:12think uh with in involving more more
58:16data we can uh do uh more robust model
58:20and uh we can use it out of the box uh
58:24worldwide. I hope we can do that.
58:27uh
58:29this is what what we are on now. So
58:33thank you.
58:34>> Thank you very much Shanor uh for those
58:38inspiring remarks. Uh and then uh Fyus
58:42uh
58:45>> yeah uh I would like to thank you all
58:47again. Thanks also for the audience for
58:50for their attention. So I would like to
58:53uh mention again that this is the first
58:56uh great step into the providing
58:59connectivity for schools. But I I
59:01believe from research perspective there
59:03are also many opportunities uh to f to
59:06first improve the methodology that we
59:08propose including uh optimizing the
59:11school detection or further uh improving
59:13the cellular or in general the
59:16infrastructure uh detection but also
59:19there is an opportunity to use AI to
59:21actually optimize the policies on how we
59:24will eventually connect those schools
59:26beyond the actual detection of those
59:29schools in the intersection. And for
59:31that uh thank you again uh everyone.
59:35>> All right. Uh so please join me in
59:37thanking all our esteemed uh speakers
59:40today. Um thank you very much. We will
59:43probably like address the remaining
59:45questions uh offline. Um and then uh
59:48with that um this closes our um uh
59:52infrastructure detection leveraging AI
59:55session. Thank you all for your
59:57attention and have a great rest of have
1:00:00a great rest of your day.
1:00:06>> Thank you for participating in today's
1:00:08AI for good session. We hope you've
1:00:10learned something new, innovative, and
1:00:13engaging in today's event. We now
1:00:15encourage you to continue the
1:00:17conversation on the live video wall in
1:00:19the neural network. Here you can ask
1:00:21questions, like and comment, share
1:00:24links, complete the poll, connect with
1:00:26interesting profiles, or speak
1:00:28one-on-one using the chat and video
1:00:30function. We invite you to explore the
1:00:32lobby, try the smart matching quiz,
1:00:35visit the virtual exhibits, poster
1:00:37boards, the eShop, and build your
1:00:40personalized AI for good program. Let's
1:00:43shape the future of AI for good.
1:02:28>> [music]