ITU AI/ML in 5G Challenge Grand Challenge Finale (Final Conference)
- Shared screen with speaker view

06:15:12
Hi all

06:15:37
Hello everyone, Dr Naranmandakh T, Greetings from the Communications Regulatory Commission of Mongolia,

06:16:31
Welcome :)

06:17:08
Hellow Vishnue;

06:19:04
Reinhard Geneva

06:19:06
Hello everyone from Nairobi, Kenya

06:19:09
Geneva, Switzerland

06:19:13
Mumbai,India

06:19:13
New Delhi, India

06:19:13
Kind regards from Germany

06:19:15
Berlin, Germany

06:19:15
Hello from Brazil, Belem

06:19:16
Barcelona, Spain

06:19:19
Tokyo Japan

06:19:21
Bangalore, India

06:19:22
Vannes, France

06:19:22
Munich Germany

06:19:23
Belem, Brazil

06:19:23
Bangalore, India

06:19:24
Dia Duit (Hello) from Donegal, Ireland :)

06:19:25
Johannesburg South Africa

06:19:25
Hello everyone! Linköping, Sweden

06:19:25
Kingston, Canada

06:19:26
London, UK

06:19:27
Hi, attending from Tokyo, Japan:-)

06:19:29
Barcelona, Spain

06:19:31
Riyadh, Saudi Arabia

06:19:34
Pittsburgh, Pennsylvania USA

06:19:38
Recife, Brazil

06:19:46
Shagufta, LYIT, Ireland

06:19:46
Maryland USA

06:19:59
Hi, attending from Osaka, Japan

06:20:02
Hello from Cairo, Egypt.

06:20:05
Geneva

06:20:06
Mohammad Abid, stc, Riyadh, Saudi Arabia

06:20:09
Okinawa, Japan

06:20:23
Paris

06:20:27
Willy from Indonesia

06:20:28
Greeting to all -MP Singhal New Delhi IndiIa

06:20:29
Beijing, China

06:20:32
Barcelona, Spain

06:20:37
Hello from Okinawa Japan.

06:20:59
Geneva, Switzerland ;)

06:21:40
Guizhou, China. :)

06:22:37
London

06:23:41
china

06:24:04
Greetings form Istanbul

06:24:09
St.-Petersburg, Russia

06:27:25
https://www.itu.int/en/ITU-T/AI/challenge/2020/Pages/PROGRAMME.aspx

06:37:00
With which product did you compare yours ? do have any results?

06:41:02
I have a follow on question to vishunu

06:58:57
Do you have any justification why MLP is giving you the worst performance among all the predictors used? could it be due to manual feature refinement and reduction carried out?

07:00:40
why Tree model work better than MLP and SVM in this problem?

07:04:43
p0[

07:14:55
Any reason why the recall of your trained model is low?

07:16:03
the GCN model specifically

07:19:26
why GCN performs better than other models, can you explain based on the problem

07:43:17
how we can get the slides of this event.

07:44:25
The slides and information can be found on the website

07:56:16
What kind of environment did you use to test your solution?.. simulation?.. real network deployment?

07:59:29
Also when you say traffic prediction.. do you mean the overall traffic on the network?

08:14:09
What is the type of nodes in your scenario here (in the context of 5G)? and is it practically easy to introduce these changes to the topology?

08:15:02
that was wonderful , for 44.5% improvement simulation , what was the duration of data what was taken into consideration?

08:39:51
questions: 1. how much % energy saving will be achieved from this approach?. 2. any sensor input considered as data source?. 3. what all data inputs expected for this approach?

08:53:40
answer for question 1:We don't know how much energy saving because we just have the data of everyday's energy-saving status.

08:56:16
answer for question 2:We have no access to real-time data source. If sensor input is offline, we should have considered.

08:57:35
answer for question 3:Historical base station performance data, historical energy saving status and base station configuration information.

08:59:38
Mr. Ming: is there a relation with standards?

09:11:01
What kind of algorithms are used for the orchestration?

09:18:38
@Thomas when will we know results of public voting

09:20:35
The results will be announced on Thursday

09:20:52
thanks

09:26:22
Did you consider evaluating the cost of the communication overhead added by applying federated learning?and also from the aspect of practicality, will these distributed devices be able to carry out the local training tasks? as in most of the federated learning cases the training task is simpler than the one in your work!

09:31:41
no problem!

09:32:19
Thank you Ahmad

09:34:05
For NU Huskies: Do U have results for infusing two out of three of the modalities (image, lidar, coordinates)? How would a network infusing only LIDAR and coordinate data work? How can you be sure that your proposed detection and background removal on image data is actually the reason for improvements in the final result when 3 modalities are infused? Do U have an ablation study to show contribution of your proposed detection and background removal on the final performance given the fact that it increases the complexity of NN significantly?

09:39:02
For Participants in The ITU AI/ML in 5G Challenge, please respond to this survey: https://forms.office.com/Pages/ResponsePage.aspx?id=12TkI-YEh0uRPCS9iSGf0-yqkfLCoQ9IpTbc_XELf95URExJWlM2OExXVEQ5VFhFSTUzWUU3SDU1WS4u

09:39:13
maybe some training packet could be dropped to reduce communication overhead because ML training load distribution has long tail. Removing some redundant training data may not impact the final performance of ML training.

09:41:03
The Grand Challenge Finale (Keynotes + Awards) will take place on 17 Dec. Please register here: https://itu.zoom.us/webinar/register/WN_bm4oJ_OeRH-Ilnb4GY_gMQ

09:41:18
For NU Huskies: Do U have results for infusing two out of three of the modalities (image, lidar, coordinates)? How would a network infusing only LIDAR and coordinate data work? How can you be sure that your proposed detection and background removal on image data is actually the reason for improvements in the final result when 3 modalities are infused? Do U have an ablation study to show contribution of your proposed detection and background removal on the final performance given the fact that it increases the complexity of NN significantly?

09:50:48
If fc layer can be replaced by other opertaion such as average pool or 1*1 conv to reduce complexity of NN ? because fc(fully connection layer) may have so many parameters. My opinion..

09:52:53
When is the next challenge?

09:54:07
Please tune-in (join) in 17 Dec to get information about what happens next: https://itu.zoom.us/webinar/register/WN_bm4oJ_OeRH-Ilnb4GY_gMQ

09:54:38
For BeamSoup: What is beta in your loss function? Did you analyse the impact of beta by some ablation? It seems like such loss should force your outputs to two different distributions (one hot encoded best beam vs. signal power), can you please elaborate on that a bit?

09:55:13
Beam soup: The way the baseline pre-process LIDAR data at the front end includes embedding the coordinate data into the LIDAR input. I am worried that separately infusing coordinate data may be adding unnecessary complexity to the NN architecture and even destructive in non line of sight (NLOS) cases?

10:04:55
Congratulations again to all participants of the PS-012 [beam-selection]! Great job!

10:06:39
Of course, great job!

10:06:57
Thank you Aldebaro and your team for providing the Raymobtime dataset.

10:09:27
An lso for posing this Nice idea in the first place Aldebaro. Would be nice to collaborate to publish an overview kind of paper on this topic putting together achievements from all teams on the competition and your results of course.

10:11:10
Tomorrow 15:30- 15:35 (CET), Call for papers: Ian Akyildiz, Editor-in-Chief, Special issue of ITU Journal on Future and Evolving Technologies (ITU J-FET)): “AI/ML Solutions in 5G and Future Networks”

10:12:10
17 Dec 15:30- 15:35 (CET), Call for papers: Ian Akyildiz, Editor-in-Chief, Special issue of ITU Journal on Future and Evolving Technologies (ITU J-FET)): “AI/ML Solutions in 5G and Future Networks”

10:14:46
as far as I understand you are using some parameters to predict current throughput value,if so, did you consider predicting future throughput values and not the current ones based on historical data? (time series forecasting I mean)

10:21:46
Are the presentations available?

10:22:46
The presentations are available on the website: https://www.itu.int/en/ITU-T/AI/challenge/2020/Pages/PROGRAMME.aspx

10:30:13
Nice talk! May I ask what kind of aggregations you used in graph convolutions?

10:30:21
what is the range of throughput values here?to get a sense of the RMSE!

10:33:22
I cannot speak right now, but what is the computational complexity of GCN? compared to CNN?

10:41:04
Regarding Loïck's question, I leave here an interesting paper that includes a comprehensive evaluation of different aggregation functions in GNN/GCN: https://arxiv.org/pdf/2004.05718.pdf

10:42:49
We will continue tomorrow with presentations from various problem statements: https://www.itu.int/en/ITU-T/AI/challenge/2020/Pages/PROGRAMME.aspx

10:43:02
For Participants in The ITU AI/ML in 5G Challenge, please respond to this survey:https://forms.office.com/Pages/ResponsePage.aspx?id=12TkI-YEh0uRPCS9iSGf0-yqkfLCoQ9IpTbc_XELf95URExJWlM2OExXVEQ5VFhFSTUzWUU3SDU1WS4uThe Grand Challenge Finale (Keynotes + Awards) will take place on 17 Dec.Please register here: https://itu.zoom.us/webinar/register/WN_bm4oJ_OeRH-Ilnb4GY_gMQ

10:48:05
R

10:48:22
will the recording be shared by any chance?

10:48:34
I vote to last team

10:48:39
thank you!

10:48:49
Yes, recording will be available.

10:48:53
bye

10:49:03
bye bye

10:49:14
OK great.. thanks Vishnu! :))

10:49:15
Looking forward to receiving the recording. Thank you

10:49:16
bye