Good day everyone. Thank you for joiningFor 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

good morning from 🇧🇷

Hi :)

Dear all,"du supports the proposal to form the FG for autonomous networks, we think the deliverables are valuable for operators like du"

Greeting to All - M P Singhal India

Hello Singhal ji, thanks for joining :)

Greetings to all.

Is it possible to share the presentations after the event. Thanks.

how do you measure packet loss? at layer 2? layer 3 or application APIs?

how does the model work? offline or online?

Why choose video traffic?

Is it a possibility to view the slides?

can I ask a question pls?

Is it a possibility to view the slides after the presentation?

Program is here (includes slides): https://www.itu.int/en/ITU-T/AI/challenge/2020/Pages/PROGRAMME.aspx

My question to Mr. Yuusuke : 1. Could please tell us how did you choose the percentage of packet loss that you show it your presentation ? Does these value are the required values to meet the QoS. ?

Thanks for the question.The dataset we used in this project was not generated by ourselves, but by NEC. We are using the dataset provided by NEC to train and test our model. Therefore, the packet loss rate is determined by NEC.

Ok. thank you . The second question to Mr Yuusuke. about the bandwidth. You aim to increase the bandwidth in your work? if yes. could you pls explain the mechanism that you follow it to increase it. ,

OK. The goal of this challenge is to estimate the network state.

Could pls tell us the objectives of your work ?

to mr >thomas

Hello yo everyone! My question is: what the technical requirements for the devices in this framework (AI/MEC)? Limitations?

I mean the person who is presenting right now.

ok thank you I got it

what is the relation between your work and the consumption energy . you mentioned it in the background in your presentation but I did not see what did yo do to make the power consumption low while the data transmission is increased.

What is the differece from that of Federate Learning?

Thanks a lot!

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

Comments for Mohammad:1. how to measure privacy?2. how to evaluate the overheard in transmitting the models and training the models at the devices?3. You may want to look at ITU's FGAI4H

Related to the question of Akihiro Nakao about private training of medical data: There is no reason not to use this work in real world applications. Such applications are in their infancy right now. That is why it is rare to find already deployed applications for now, but in the future such approaches will be in use for sure. Also the solutions presented here is not limited to the medical applications but can be extended to any distributed learning problem.

so far as I know, The technical solution and application requirements of FL are discussed in IEEE 3652.1 as

https://standards.ieee.org/standard/3652_1-2020.html

Sorry, weren't we supposed to vote in a poll after the break?

Thanks Vishnu, about Q 1 ) the patient-level privacy is measured by two Differential Privacy parameters, one is epsilon-D and another is delta-DP. delta-DP is the probability of failure which we set it to be 0.0001 and in general must be > 1/N, if N is the number of patients. As we had <10000 patients, we set it to be 0.0001. which is a very small probability of failure. The other one is epsilon which we have shown in the results. It depends on the amount of accuracy you needed. We tried to keep epsilon less than 5.

Burak thanks for your clarification.

Thanks to Vishnu for the comments. The differential privacy is quantifiable with well known techniques in literature, as Mohammad just explained above. For secure aggregation using homomorphic encryption, privacy is measurable by the computational hardness of solving the LWE problem on which homomorphic encryption is based on.

Thank you Nitish and Mohammad. We should talk offline :)

About Q2) how to evaluate the overheard in transmitting the models and training the models at the devices? That’s for sure important. We assume that patients trust their hospitals. And so, hospitals can afford the required computation to keep their patients private and also to collaborate with other hospitals in training a model. So, to compute the overhead we need to consider a real world setting which for sure is a very good direction for expanding this solution.

Thanks all. Please reach us offline :-)

There is a trade-off between the level of privacy from homomorphic encryption and the communication overhead, because the ciphertext size needs to be increased for a larger amount of privacy. However, state-of-art techniques like batching, which packs many values within one ciphertext, are able to reduce the communication and computational load by as much as 3-4 orders of magnitude, which we use in our implementation.

Great demo! I could see in your report that for the aggregation you use a concatenation of avg/sum/min/max/var. To what extent did it improve the accuracy compared to a simple sum?

And what is the cost during inference? What is the approximate cost over a middle-size nnetwork topology?

Question for Loick: how dynamic are the hidden parameters? (e.g. I see that you are using capacity as one of them, practically the capacity of the link could change if it is wireless).

Vishnu: we can basically put any feature we want to initialize hidden states, and these hidden states are "propagated" during inference using weights learned during training. Capacity is one feature available for links, the type of link could also be one feature. I hope this answers your question!

Greetings from Mongolia.

Did you consider using a one-hot encoding vector for the queue scheduling configuration instead of numerical values?

on 17 Dec (during the awards + keynotes)​15:30- 15:35 ​Call for papers: Special issue of ITU Journal on Future and Evolving Technologies (ITU J-FET)​): “AI/ML Solutions in 5G and Future Networks”Please register: https://itu.zoom.us/webinar/register/WN_bm4oJ_OeRH-Ilnb4GY_gMQ

Impressive results! I'm curious, how did you manage to embed node's features (QoS policies) into links' states?

Thank you for all RouteNer challenge presenters and everyone. Signing off for today.

@Dheeraj: we want to explore the generalization of the solutions for such dataset. Do you have suggestions? E.g. If we have a standard mechanism of representing the "pattern" that you mentioned, is it better to generalize across the data from Chennai to Istanbul?

An important aspect to achieve generalization is take network topology and geographical features (like forest, urban areas etc.) into consideration. In fact, we designed our solution for generalization as original problem statement said we will be evaluated on different topology. However, we need to test over multiple data-sets from different locations to validate above idea.

Thank you very much for very interesting presentation

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

Well done to everyone, excellent presentations. Thank you to our wonderful hosts Vishnu and Thomas, what a team :)

See you tomorrow!