 the webinar series of the IT Journal on Future and Evolving Technologies. My name is Alessia Magniarditi from ITU, the International Telecommunication Union. ITU is the United Nations Specialized Agency for Information and Communication Technologies. ITU allocates frequencies to the services that make use of the radio communication spectrum, it develops standards and assists developing countries in setting up their information and communication infrastructure. ITU and academia share a commitment to the public interest and this commitment is embodied by the IT Journal which offers complete coverage of communications and networking paradigms free of charge for both readers and authors. Our journal welcomes submissions at any time on any topic within its scope and we believe that this new webinar series will inspire more contributions from researchers around the world. It is my pleasure to open today the webinar with Dr. Xi Jinping from Queen Mary University of London who will speak about cementing communications transmitting beyond bits. We count on your support to make this webinar an interesting experience. Please submit your questions via the Q&A channel. We will address them to our speaker during the Q&A session. After the talk and the Q&A, please stay online. We have something very special for you. The wisdom corner, live life lessons. Dr. King agreed to a very personal chapter. She will share with us some lessons learned over the years that might perhaps be useful for some of you. Now it's my pleasure to give the floor and to introduce Professor Ian Achilditz, editor-in-chief of the IT Journal on Future and Evolving Technologies and founder and president of Truba from the United States. With Professor Achilditz in August 2020, we launch this new scientific journal and after a year and a half we are already probably moving towards impact factor. Professor Achilditz established many research centers worldwide in the last two decades. He is editor-in-chief emeritus of impact factor journals, highly cited and at the top of the most prestigious international rankings is visiting distinguished professors in many universities around the world. His current research interests include 6G and 7G wireless communication systems, semantic communication, hologram communication, molecular communication, terahertz, internet of things in challenged environments, nano networks and many other topics. Professor Achilditz, the floor is yours to introduce our speaker and then to moderate the Q&A session. Thank you. Thank you Alessia. Again very nice introduction. Good morning, good afternoon and good evening all around the world. We have many people in the webinar as I see. I again welcome you all to our ITU Journal for Future and Evolving Technologies webinar series. I have the great pleasure to introduce you our webinar speaker, Dr. Qin Zixin from Queen Mary University of London. Zixin received her bachelor's degree from prestigious university in China. It's called BUPT, Beijing University of Post and Telecommunications almost a decade ago in 2012 and the PhD degree from Queen Mary University in Electronic Engineering in 2016. She was a postdoc at Imperial College London from 16 to 17 then a lecturer at Lancaster University in 17 and 18. It looks like she really loves you know the kingdom. So since 2018 she has been a lecturer. It's kind of like associate prof in the US with the School of Electronic Engineering and Computer Science at Queen Mary University of London and her Google scholar is age index is 31 and total number of citations is 4,800 which is really excellent for her generation and for her seniority. Don't forget she just received PhD in 2016 and I followed her work with keen interest. She did very good research on NOMA, Non-Ortagonal Multiple Access. She has very interesting and pioneering papers on that topic but lately also on semantic communications and again I really congratulate her for all pioneering work on this semantic communications topic. Now everybody is following her work and Zhichin is very active in the service to our research community. She gave again some keynotes and tutorials at some major conferences such as IEEE PMRC or in jargon it's called PIMRIC. IEEE VTC and IEEE Globe come in recent years so her career is really shooting up extremely fast and she's serving as an editor for IEEE JSEC and IEEE Transaction Communications, Communication Letters and Transactions on Congressional Communications Networking. She also served as the track chair for many conferences like Globe come 20 and 21, VTC in 2019, Globe come 2020 and her research interests include deep learning, enabled semantic communication and to end communication deep compressive sensing, intelligent resource allocation, various applications such as UAV, communications and famous reconferential intelligence services. Again we thank you Zhilin for accepting our invitation. We look forward to your presentation. Thanks again. Flora is yours. Thank you very much Professor Akides. Yeah you I literally before I started the talk yeah I want to mention a little bit since like when I did the PhD I read lots of your papers many many of them because I you can see I worked on the comprehensive sensing for the spectrum sensing so your your your nice surveys the technical papers are cognitive radio spectrum sensing since yeah that's really amazing and also use the master in our era I guess I will when you attended the talk knows you well and know your work well yeah thank you very much for guiding us in the society and thank you very much for inviting me it's really a great pleasure to get the invitation yeah thank you okay let me first can you say my slides yes okay brilliant yeah thank you very much everyone yeah good morning good evening good afternoon today I'm going to introduce a little bit about our work the initial work literally on the symmetric communications at the at the very beginning I would talk a little bit about the general ideas of the symmetric communication then we introduce our first work developed by my one of my first PhD students literally Huixiang who is also attending the talk today that's about the deep learning in medieval the community the symmetric communications then I will talk about the several variants based on the deep learning based on this work we call that deep passing so the we have developed quite a few new words in the area that's also mainly contributed by my PhD students and visiting students okay at the end of the talk I will leave some open questions actually the research on symmetric communication is at the very beginning so I think we got a lot of problems to solve in the future right so let's see the rules of the symmetric communication now we all of us have seen that okay the research on the 6G is on the way we are dealing with different types of the intelligent networks either on the different applications for the like the machine learning in different perspectives of the communication system well one of the key rule in the 6G or the beyond is like as you could see from the figure we are dealing with the different types of the data massive amount of the data and also the data could be multi-modal so one of the rules of the in the 6G and beyond is we are trying not just trying to transmit the symbol correctly we are trying to support the intelligence transmission so for such a task symmetric communication is one of the enabling techniques to achieve that and we could we could ask image that a potential application for the symmetric communication could start from the machine to machine communication as we did in the 5G but the difference is okay in symmetric communication the devices will be more intelligent to deal with intelligent tasks and after that we would probably be able to support the human-to-machine communications and even more human-to-human communications well we could also see that in the industry and the research council are also quite interested in the topic I've witnessed the the proposals of the successful grants founded by like the the UK UKI the UK research fund council and also the funds on the joint research project between UK and the US and mainly on the statics statitions like candies from the from Stanford University who is also working on this on the information pursuit for the multi-modal data so also from the companies we've got funds from Huawei and we also also people from Nokia China Mobile working very actively in this area okay I noted that at the very beginning someone asked us to give a note for what is a symmetric communication so let's have a look at the difference between the conventional communication and symmetric communication so before for the current system we are treating the communication system as a tube for transmitting the data we don't really care about the content of the data right so what we care is the successful transmission of the symbols so this is actually the first level of the communication that was categorized by Shannon and we were and also we could see in the in the past eight decades most of the researchers has made a great efforts in this area to and we can see that in 5G or I mean for the even in 5G we are actually approaching the channel limitation while to support those like some huge amount of huge amount of data provided by the huge number of devices symmetric communication actually is the second the second level of the communication which aims to support the successful transmission of symmetric information so we as a level two we are actually trying to transmit the symbols that to make sure they actually convey the design the meanings from the south right and by doing so when we we know that's many of the communication we have the corresponding tasks as a receiver to carry out right so with the symmetric communication all many of the researchers also call it as a goal oriented communication we are actually transmitting the symmetric features relevant to the task only so there are some informations that was transmitted in the conventional communication system but not actually they are not highly related to a special specific task so in symmetric communication or in the goal oriented communication we want to transmit those features by doing so we could reduce the size of the data to be transmitted and improve the transmission efficiency significantly here I provide a simple example for the symmetric communication so typically if we want to transmit an image in the conventional communication we actually convert the image to the two bits we process the image pixel by pixel right and we perform then once we got the bits we perform the south coding channel coding modulation etc and then we transmit the symbols as a receiver what we normally do is try to recover the image or estimate the image pixel by pixel right so that's at that level we are focusing on the bit sequence transmission but with the symmetric communication the transmitter and receiver should have stronger or more powerful should be much more stronger or powerful to deal with the to interpret or to understand the the south and also to understand what's the task the other receiver with with the symmetric communication so for example at the transmitter the symmetric transmitter could okay try to understand the the content of the image that's actually describing that people is riding a bicycle right so at the receiver we may not care about each of the pixel but we care about the content so in this case we only need to generate a new new image based on the content we received and which actually shares the same symmetric information as a transmitter if your task is to understand the understand the image of course there are i noticed that there are people using more advanced techniques to try to understand the or try to describe the image like since graph generation but that's all kind of detail apply the tools to support the goal-oriented symmetric communications right now let's have a look the the developments of the symmetric communication as i mentioned the symmetric communication is not like a new concept that's actually been mentioned eight decades ago but in the past we have shown some there's some works on symmetric communication but many of them are trying to follow the past of the channel that that developed for the information theory so the researchers tried to normally they tried to use the logic probability and then develop the model to to to provide kind of the limitations for symmetric communication and typically like the work i mentioned here one or two they are mainly focusing on the text transmission because if you think about the symmetric communication then in the natural language processing they have developed some language models that could help us to that's that's kind of straightforward for us to think about the symmetric informations from the text right but we noted that most of the works they are still deal with the things at the world level and also they cannot they can't fully understand the meaning behind the also space and also the the limited applications of those model also narrows the the further development here i want to i want to just highlight this one of the definition for the symmetric capacity and usually that's provided in the work too that for a discrete and memoryless channel we could or they actually derive the symmetric capacity in this way okay if you observe the equation carefully we could see like okay the the key difference is the third part here right and actually 88 sorry let me actually this part is used to measure the average logic information of the received message which is actually representing the ability to interpret the received message so but here unfortunately they also only provided such general expression and we we didn't have for the for the closed format to to support or to get the symmetric capacity but i'd like to highlight this okay that's actually one of the one of the developments in the past well we could see like the limitation of the previous of the existing work on symmetric communication is mainly because of the lack of the general mathematical model like people trying to use either logic probability on some or some other rules but they are actually trying to follow the the way that's generally defined for the information theory but we we note that's okay symmetric information is actually quite hard to quantify so but to avoid requiring such a general mathematical model we are inspired by the wide application of the deep learning and also we notice that the deep learning actually okay we actually kind of find a close for closed format expression to to express the sense but we could use the like deep learning or deep neural network to represent to to model to model or to capture the features right so that's why we used the deep learning enabled symmetric communication to improve the system performance by using the deep learning to design symmetric communication we are facing three challenges but the first is okay how could we define the meaning behind the big sequence right and second is if we could understand or we could define the symmetric information how could we design the performance matrix for the symmetric communication system because in the conventional communication system we are using beta error rate symbol error rate to define and to measure the system but now the transmission the focus of the communication system is not the beta error rate or symbol error rate right we are trying to guarantee the successful transmission of symmetric information so how could we how could we design the corresponding performance matrix to measure the system as one of the core challenge and the third challenge is okay if we could have such a system sorry if we could have such we could understand the symmetric information and know how to measure the system how could we design it as the communication system at the symmetric level right okay here I list a few works that I'm going to mention today so first is the DPCS the deep learning enabled symmetric communication system that's the work and literally we are focused in this work we are focusing on the text transmission and based on the frame structure proposed in this work we further developed a few variants to support the speech transmission the multimodal data and the multimodal data transmission and also we developed the light mode for the DPCS to make it affordable for the iot devices right let me first introduce the basic structure of the DPCS so here as you can see from Shui in the figure we have a transmitter and the receiver the core components in the transmitter is the symmetric encoder and channel encoder of course we could further extend that to like later discuss that we could further include the modulation or et cetera but here we are treating the system as an end-to-end system so we don't separate those blocks and by at the transmitter and the receiver we are using neural networks to represent them respectively in the here in this system the input is just the the sentence and after it pass through the symmetric encoder and channel encoder the information are actually facing two types of the channels one is the physical channel that we are actually facing in the current communication system that's in the physical channel right and the other one is the symmetric channel so in symmetric channel we need to deal with the symmetric channel noise and here the symmetric channel noise are actually referring to the misunderstanding about during the interpretation of the disturbance in the estimated information so the the concept for symmetric noise would be quite wide and it could be quite different for different sources so for example here if we consider the text we could consider the with the symmetric noise the source would be corrupted so the sentence here let's say that's the original input that's correct or the ground the ground choose of the input sentence with the symmetric channel sorry symmetric noise it may be corrupted like some of the some of the letters or even words in the sentence could be modified or could be changed or attacked that could be one of the symmetric noise but of course there would be more different there there are more formats or reverence and also it's still not very clear how could we have a general model to to represent the symmetric noise like we did for the physical channel like we model that has an AWG and is an AWG channel right okay anyway based on the basic structure we we introduced and we used the transformer based symmetric coding so here the transformer is literally a very powerful neural network proposed by google back to about four four four four to five years ago so the the key idea for the trans transformer is okay it's actually used in attention so let's take this picture as an example if the sentence is the monkey eats that banana because it was too hungry right if we are using the transformer the neural network will be able to understand the the correlations all the connections between eight and the corresponding words and after the by using the attention the neural network will be able to learn okay the the word eight has the highest probability to to representing the monkey okay if we use a human judge we could say okay that that's actually correct right and it's the probability the probability for eight to representing banana is relatively lower and for the other words that it's lower and lower right so by using such a structure we could use the transformer to capture all the symmetric features from the text to understand the the meanings behind the text and also very recently we noted that the wider application of the transformers so later i will introduce another work for based on this right so for the deep assay the detail of the design is first we actually use a loss function and the function and the loss function literally includes two parts one is about the cross entropy that's used to if we could reduce the cross entropy the neural network could learn the the meanings are behind the text right as i just introduced earlier and the other part of in the loss function is the mutual information so here we're trying to minimize the mutual information and here's a lambda is just a parameter to weight the weight parameter right so with such a loss function when we start to train the neural network we will take this we will first take the second part of the second part of the loss function the mutual information part so we will try to maximize the mutual information part so that we just train the channel coding and decoding channel encoder and decoder right and then phase two we will train the whole model all together here in the following i will show you the performance obtained by using such a loss function and and the training process but what i want to highlight is the gain that we can actually get from the the deep assay actually from two parts one is mainly introduced by the symmetric encoder and another is end-to-end wireless communication system design so maybe you've noted as many of the researchers that's worked on the end-to-end communication system and compared to the typical block structure based communication system we could obtain a certain gain all right with such a design we try to measure the system to see if that works well always that works as we as it was expected so first we take the blue skull as the performance matrix literally blue skull is a typical or let's say a popular matrix used in the natural language processing and here the key idea for the blue skull is is trying to compare the difference between the words in two sentences so if you notice that's the the characteristic of the using the blue skull so we are still uh compares the the two sentences by words but if you choose one group uh one one gram blue skull like you you are kind of comparing the two sentences word by word every one word right if but also you could choose two grams three grams or four grams that's the popular sense in the when using the blue skull and we could see that's okay if for example for the foreground that's actually you are comparing the the two sentences every four words right but we notice that okay the such matrix is still dealing with things at a word level so here in this work we also propose to uh you propose to define the sentence similarity which actually use the the bird model here the bird model is also uh is a model uh printing model provided by google which was streamed by a massive number of the text and it's also based on the transformer and we could see with such a printing model the powerful model we can we can map the the sentence actually this model i find a symmetric vector space by using it we can map the sentence from the text the word by word into the symmetric vector space and by doing so okay as is the transmitted sentence and as had its estimate say the sentence at the receiver so by rather than compare them word by word we will calculate their distance in the symmetric vector space and if their space are close enough we think they are sharing the same or similar symmetric information right that's the second way to measure the system okay based on that here i show some quickies simulation results as you can see at the left hand side we're showing we just show the one gram blue skull and we are comparing our methods the DPSC with a typical communication system like here we adopt a Hoffman coding and a tube code and also we compare it with deep learning enabled join the south channel coding we could see that when the SNR is 12 dB the blue skull could be improved like in the typical methods it's about 0.1 0 to 0.15 and then with the DPSC we could achieve like kind of 0.9 then we can see that we could achieve quite significant performance gain especially when the SNR is relatively lower of course here i didn't show the example in a complete way so if we further increase the SNR the results as a performance of the typical methods could be increased but we we are showing that okay actually when the SNR is relatively lower the performance as a DPSC is quite robust right and similarly we show under the same condition we show the sentence similarity so here we you can see that when if we take the SNR as 12 dB again the sentence the calculated sentence similarity is actually almost zero very close to zero that is because if we are using the the typical methods like the blue skull is about 0.1 if we receive the sentence with this kind of quality we are actually the people are actually not able to understand the meaning of the text because it is full of the arrows so here you can see the sentence similarity actually it's a good performance matrix to to reflect such a similarity in the semantic okay again when we are using the DPSC we could improve we could achieve very good performance that's almost to one okay apart from the initial results we further apply we further adopt the transfer learning to make the DPSC adaptive to the dynamic environment here's the dynamic environment I'm referring to two different two different types of the environment one is about the different background knowledge like if you could recall at the very beginning for the system introduction I mentioned that okay the the background knowledge is actually the key to enable the symmetric symmetric encoder and another difference is another different environment is about the channel condition so we'll know that in the wireless communication the channel condition can be very dynamic and various and by using the transfer learning here we have show that with the transfer learning as shown in the black curve the performance the model could be could be could converge to the state stature stable stature so after a few epochs compared to the case as shown in the red that is not using the transfer learning but meanwhile we didn't specify any performance okay that's for the DPSC and in the following I will introduce a few variants for the DPSC so first is about the speech transmission well here like we are actually I just highlight the the key challenges and our contribution and mention a little bit about the performance we've got if you are interested in the detailed work you're highly recommended to check the corresponding papers but here let's see for the speech transmission yeah we actually adopt a similar structure as as a DPSC we have the joint symmetric and channel coding right but for the speech transmission we actually categorize them into two different tasks one is okay the transmit the transmitter what sends the speech information while as a receiver literally the receiver either like if we are making teleconference sorry telephone call we we'd like to hear the voice from the transmitter right but sometimes we may only maybe we are only interested in the text the corresponding text information from the speech so that could be achieved by using the popular manual the popular speech recognition tools right so here we consider two cases and so we categorize the DPSC STS task orientation transmission and we separate here we separate the text the symmetric information from the speech signal what we actually did is okay as you here is the GUI that we developed for implement the DPSC ST for the speech recognition and the synthesize so at the transmitter I use the joint symmetric channel coding to get the corresponding to to perform the speech recognition and the corresponding coding and then I will get the text rather than transmitting the speech signals I only transmit the text the related symmetric features from the text recognized after performing the the speech recognition and after receiver we could either recover the sentence oh we could use the speech synthesis model to generate the speech signal and also we could give the each user a different idea and capture some and once they are registered we could capture the features corresponding to their voice so that at the receiver I can easily generate the speech signal that's I highly close to the speech voice transmit that's at the transmitter okay that's that's what we actually did so by doing so you can see I only need to transmit the text symmetric features right rather than compared to transmitting the speech features or the speech signals we could lower the network traffic or reduce the size of the data to be transmitted significantly so here actually this demo we we should have put it online yes I think we provided a link in this paper so if you are interested to play with it you could you could have a try right so before I mainly focused on the pointer to point transmission so in this work the MUD passing we are actually focusing on the multi-extended from the single user case to the multi-user case and also we are rather than transmitting the text speech only here we are caring about the multi-modal data transmission and also here we are trying to design a network sorry the commit symmetric communication system for serving some specific tasks so it we also call it as task-oriented communication symmetric communication system so here we choose the vision question answering as their task and you can see like okay for example we we just take the case with two users one of the user actually transmitting the text actually the text is the question that's related to the image that's transmitted by the other by the second user right so here each of the user they just transmit the symmetric features like either the text relative symmetric features or the image related symmetric features while as a receiver rather than recovers the image and text respectively and then carry out such a task we actually what we actually did is we use neural network by you by taking the symmetric features from the two users as an input and the output will be the answer relative to based on the question and the image right so if you take this as an example this is the image transmitted by one of the user and the question is are there any other things that are the same shape as the small red shiny object right and okay if we use develop the D-pass C-modal the receiver could understand could understand the image and the text and then generates the corresponding answers yeah if you decide that if you have a check that they can provide the correct answers but if you we use the typical one the like the g-pack ldpc etc they won't be able to provide provide the correct answers due to some of the impairments caused by the channel okay by doing so we actually we here we are actually just using the vqa as a example to show the to show the multi-user case and by doing so we could reduce the number of the symbols to be transmitted by about 17 percent and we also notice that's the number of the symbols sorry the the the number of the symbols to be transmitted for the image be reduced significantly right okay that's the third work and another extension of the extension of the D-pass C is called the light version of the D-pass C the ldpc and we could see that we actually we are using the iot devices and many of them are powered by battery so they are limited in terms of the the power and the storage the computing capability right so if we consider a system a network with massive number of iot devices so first there's maybe a cloud or the edge that could initialize or train the model and like train the D-pass C model and then they can broadcast that to to allow each of them or each of the devices to equip them with such a D-pass C model right and then once they have the data to transmit they can they can generate the symmetric features and send it over the channel and then the edge of the cloud could further perform the let's say the model update or the further data processing right so in this work here I mentioned here this work we actually used the model compression techniques to perform those non-essential model parameters so literally we after we did the simulation we observed that if we specify 90 percent of the parameters the model parameters the performance won't be degrade we could even further decrease the the the size of the model by 99 percent but we just we just need to sacrifice the performance a little bit and then for those essential model parameters we will need to quantize them so typically we use 32 bits in the current system right and but after the simulation we found that we only we actually only need eight bits to quantize those essential parameters in the L-D-pass C without any performance degression but by doing so we could compress such a D-pass C model by 40 times and the the power consumption at the devices can also be reduced significantly okay that's a light mode for the D-pass C for the for the affordable application right that's the technical works I'd like to share today so in brief we we develop a deep we're using the deep learning to empower to powers the the design of the symmetric communication and also we develop the different turns to support the text speech multi-modal data transmission and also we consider from the single user case to multi-user case and among all of them different all those different design we could see like in general that the symmetric communication system show a good robustness at the low SNR region and also it could reduce the size of the data to be transmitted significantly so that we can improve the transmission efficiency right last but not least I'd like to share and share a little bit thinking in the area so some problems I think that's important but probably not solved yet so first is okay as in the past eight decades as I mentioned as a beginning people are trying to find the like the all develop kind of theory let's say symmetric theory but so far we haven't got such a general general framework for the theory but but as far as I know there are many many series has been developed actually we review those different theories you know in our latest survey paper so if you are interested you could have a look at but they are they all those theory all have different limitations like they are mainly focusing on a specific area so probably one of the most fundamental and significant challenge is if is there such a limitation that's also something I'd like to discuss with you guys and if we could find a way to quantify the symmetric information yeah and well but for for me or for my group we are more focusing we are focusing more on the application-oriented symmetric communication like first how could we construct a knowledge base to enable in the design of the symmetric encoder to support the different tasks and different sources right and maybe literally we need to find a trade-off between the the performance the size of the data to be transmitted and the generalization of the transceiver right and the third challenge I want to investigate is about the symmetric aware network management so before we we normally use like the energy efficiency spectrum efficiency we're trying to maximize the the data rates maximize the the lowest data rate to guarantee the user fairness etc so we all those objects all the optimization formulation are based on the bit right we are processing the bit frequency in the current system but how could we have for the symmetric aware network we we need a kind of new formulation how could we how could we reflect like like the kind of if there is such a symmetric unit how could we can find the way to represent to maybe we we are we we are thinking a way to represent or define the symmetric spectrum efficiency or symmetric energy efficiency or etc that's that's more related to the symmetric part in in the system design and also by doing so we need to think about the resource allocation not just the the resource in the transmission during the transmission but also the resources are from the symmetric part like we could compress or we could throw many of the symmetric informations but what is a proper proper design to how much symmetric information should be reserved for supporting the transmission and the task performance task exactification as a receiver so that's that's a problem I think we would like to investigate okay here I list a few like a few more works as I mentioned this is a way we prepared very recently and some other related works I didn't have a chance to share today but if you are interested feel free to have a look at them okay that's all for my talk thank you very much thanks a lot very nice excellent talk I really enjoyed it personally and there are many questions let me start with Monsieur Starr is a tool he makes some comments about overall Shannon and semantic communication but his question is is there a theoretic limitation for semantic communication systems sorry I didn't sorry I repeat is there a theoretic limitation kind of like bounds I assume uh-huh communication systems that's the question yeah like it's kind of like performance measure right mm-hmm yeah thank you very much for the question yeah actually that's a very good and very important question we'd like to know the we'd like to know the if there's a limitation for the semantic communication system but there's an answer today I can give is we don't know if there's such a limitation because we are actually still far away from the way to design a symmetric communication system so yeah but maybe in the future some of the the you guys or maybe asked could find such a way to live on to quantify or define the the semantic communication limits yeah that's what I could answer at the point of today yeah same is a too sorry is asking can you use visual semantic embedding to assess image semantic similarity uh there are papers on those images and like evaluating them from semantic communication perspective but please go ahead so there are papers is that true okay go ahead yes I think uh for the semantic uh semantic uh the image semantic similarity uh for the viewing semantic embedding I'm not really sure if you are talking about the same thing as like I actually I have students working on while we are trying to use a symmetric embedding to to find a way to construct the image yeah I think well the answer for this question I think should be should be yes to assess the image similarity if we we could find a way to to quantify sorry to we could find a symmetric embedding space we will be able to construct the image like we could treat them as a base and then we can assess the image semantic similarity based on that okay so I jump to another question Sridhar Iyer he is in India sorry his name is Sridhar Iyer so he makes first comment uh I'm sure you heard that before whether semantic communication for wireless networks is just and another machine learning improved joint source and channel coding auto encoder or something deeper is not particularly clear yet maybe it's for him it's not clear but anyhow so one has to figure out if most gains come from improving what happens at the application layer or if the adaptation to the physical wireless channel is also important it is plausible that the most important aspect is to determine how to divide the application between transmit and receiver so overall he's you know I think you should focus on the first part saying it's kind of like is this yet another joint source and channel coding with beefed up with machine learning uh you know I don't agree but please uh your opinion yeah thank you for the question yeah actually many people ask about this question wow well my answer for this is what we have done now is uh deep learning enabled the joint symmetric channel coding but it's not just machine learning improved uh source channel coding yeah uh if okay like as I mentioned the one the one of the performance uh part of the performance gain is from the design of the symmetric encoder and part of the gain is from from the end-to-end communication system design but what what we could do more in the future is well literally I noted there's some other research has done but they they that's based on the that's based on the uh the end-to-end communication system we actually need to so so far we are we are actually uh simplified that we we haven't just think about the the the following structure we only care about the main part like the source channel coding part but if we uh further think about we need the modulation we need we need to design the waveform to build a system so here we if we want to have a symmetric communication we need to consider the further the the following parts what we could do like one of the simple ways okay we could add the current join the symmetric channel coding design to the to the to combine that with the existing modulation techniques but we could also combine that together by treating it as the end-to-end communication system what I've noted is that one of the work published probably several years ago they are trying to learn the waveform or let's say the the constellation diagram to support the multi-user transmission based on the end-to-end system so I think that's something definitely we need to further investigate so I don't think that's one of the idea for next it's not just about the source channel coding and also something more interesting to be honest that's that's what we could do like now we are using the deep learning but I think the the society the area I really look forward to a kind of theory that finds a way to to quantify symmetric information I think that's probably the one of the most important thing to do yeah okay thank you and there's there are many more questions there is one from Liu Chuang Hong he's from your alma mater BUPT he I think it's a he right or she or they or whatever right so he that person says it's a wonderful talk and he has two questions one is what does the different knowledge refer to in on the slide 17 I think different knowledge does this refer to different training data set that's the first question oh yes I think here if we if we are talking about yeah we actually refer to the different training data set but it could be a like a broader concepts like if you are changing changes switching from text to some other different sources the knowledge could be very different right for us to understand the image or understand the text so that's what we we know for different knowledge and also I noted that some of the researchers started to invest find a unique way or a general way to model such a knowledge base yeah so has another question which is really interesting in my opinion I think the best questions so far in my opinion how did we start to study the measurement of semantic information and also semantic channel capacity wow that's yeah that's a very very good and a very important question yeah well without without if we can't find a way to quantify the semantic information probably it's hard to define the the symmetric channel capacity like I mentioned is the simple equation but that's all they have like we don't have any further insights or results so let's back to the initial initial question how could we quantify the symmetric information and I I have noted like there are many researchers tried they developed their own frameworks or their own worth to to quantify the semantic information that could be from many different areas like even the foundry system they try to define kind of confirmation of degree but all those rules or theory they all end up with okay they actually don't know how to how to quantify the symmetric information so what they mentioned is okay the many of the cases those symmetric information all the so the so-called symmetric information in the system are designed manually by the experts so you could see like we are we don't know like or maybe many researchers from different areas they have tried to find a way to to quantify that yeah but it's hard we don't know if that's possible to quantify that and also that's why we use deep learning to to avoid to quantify that yeah that's what we can do now but yeah I would do really help the some genius could solve the problem yeah but if you're interested you could actually have a look at the survey I mentioned and also you could track the one of the project I mentioned that in the survey that's a joint project of contributed by many of the statisticians like from the Cambridge Stanford they said yeah they are working on this but some of the recent work I noted from the Macy they are also trying to use deep learning to to kind of solve the problems yeah that's my okay thank you one more question uh antero ante I call it we call him under why Neo is from Finland University of Helsinki was the amount of transmitted data measured for text transmission with deep sc in comparison to other methods sorry I didn't quite understand was the data measured or measured yes data or text you know for kind of yes yeah okay with deepest yeah yes so you compare with other methods uh yes I think so and to be honest here for in terms of the size of the data to be transmitted for the text we don't have much like if you check our paper for the multimodal data that the multi-user multimodal data I think we provide the results the comparison we don't have much reduction in terms of the size but what we could do for the text is we could improve the robustness at a low SNR region this is because yeah literally the size of the text is already quite small so we actually this not that are emerging to reduce it to a smaller size but for the other sources like speech or the image tag sorry videos we could reduce the size of the data significantly yeah thank you there is a question uh you know sometimes you can see from the question some of them are very deep into the subject some of them are really trying to get into subject I guess and this question is like I don't know uh so the question is from young Zhao Hui yes okay so you know maybe the person yes I know Zhao Hui yeah okay does the transmitter need to send the common knowledge over I don't know what you or he means traditional links before he over traditional links before semantic transmission maybe I guess I guess I guess he's thinking like if we need to share the the common knowledge or the background knowledge over the traditional communication as links is that before we started the trans symmetric transmission like we I guess like we he's talking about handshaking type thing or like you know like request to send clear design and then this I don't know that's what he wants to say that I don't know yeah I see or maybe that's not uh limited to the traditional links like you know in the cdma type you know I'm sending in this code yeah use that code to decode it right some I don't know that maybe yes I think uh the the transmitter receiver literally what we are currently we are doing now is we do a joint uh sorry the the join the training so actually they are sharing the those common knowledge yeah we we could share it by the traditional I don't know traditional links or some like we could design some protocols to support that okay so one more question from Han Tianxiao and again I'm really puzzled by the question the question is what is the difference between semantic communication and goal oriented communication I have to admit this is the first time I hear about goal oriented communication I don't know maybe he wants to say data-centric communication or contents or I don't know what wall oriented communication means I'm really sorry but yeah yeah yeah I have like I've heard some of the researchers like to use the goal oriented and the communication yeah I think semantic communication is more related to like link the level two communication we're trying to guarantee the semantic information transmission while the goal oriented communication they are focusing more on the tasks so I think this kind of more specific kind of specific communication system if we have if we have like relatively specific specified tasks we could design the system like the the task could be just for image classification right so in this case the amount of the data to be transmitted could be largely reduced because if you have enough power at the transmitter you could even do all the processes at the transmitter and just transmit the classification results so I think the goal oriented communication is more like the level three like you care about once you have the semantic information transmitted how are you going to use that to carry out the following tasks or following application like whatever is required by the application that's that's my understanding is this yet another name for like I told you like the data-centric communication or context communication uh I don't know like maybe they have to work together right semantic and yeah well oriented but somehow you have these levels as you said and then on the upper layers you go for like you know data-centricity like you know there's no difference it's kind of a complementary but you know I see more and more new names for the old you know solutions that's the generation what can you do so Behnam Ojagi is asking about network slicing is very important as you know for 5G and 6G especially automatic network slicing and then he's asking is it does it need to be redefined for semantic aware resource allocation my opinion is you already mentioned that in your last slide right the network management and also resource allocations but please go ahead so if you want to add yeah thank you yeah thanks for the question yeah I think so like we need to kind of redesign the way to to perform the resource allocation but for the resource allocation I'm not just talk about the resource for the transmission for the for the you know like the network slicing we're also care about the semantic related resources like let's say the the the coding rate the semantic coding rate etc that's what I was I wanted to highlight in my last slides but I yeah I think we need to have a joint design for the those resource allocations so we need to update this corresponding yeah so the answer is yes it's really important and so it would be good that some people start to look into that so Tolga Giriji he's from Turkey he's asking and I can rephrase it but the exact question is how can we adapt the classical kind of like TCP IP protocol structure right bits and packets and error control routing congestion control etc to semantic communication so in other words the question is do we really need to modify the TCP IP protocol stack and consider these you know like we already have a right the the syntactic errors like this typical classical error corrections and then we have the B section about the semantic and the third one is the effectiveness so if you consider all of these three cases do we really redefine a new TCP IP protocol stack after you answer it I will add my opinion so please go ahead okay yeah yeah I've actually discussed this question with some some researchers on the network design yeah I think the symmetric communication at least the semantic communication I'm talking about today we could we could make it a compare compatible with the current current system so probably you just replace your your your physical layer and then you could even still use the higher layer as we are using but maybe that's not a very efficient way it could work but I've noticed that some of the researchers they've kind of developed a new network architecture that's kind of the architecture they actually simplified the actually they redefined each of the layers the tasks and how the how the data how the flow are connected with each within each layers I actually I can't remember exactly the design but I did review the a few of the the the frameworks well they are still at a very high level design not with very specific design so yes I think we could the the simple answer is yeah we could use like at least the symmetric the let's say deep passing well well connected to the existing network structure but if we want to further improve the efficiency for the data flow processing probably we need to redefine that but that's that's another debate like the compatibility or performance yeah you know I can give you my opinion to the question so 1990s people were attacking TCP IP protocol stack I remember even some people were saying we should do instead of bottom up top down so the application layer should be all the way down so all these crazy ideas were floating but then we moved on and now I'm sure many people especially in this forum they heard about networking 2030 and Richard Lee is pushing from Huawei for new IP you know if some of you don't know please take a look at it so they're they're very active trying to redefine the you know kind of like protocol stacks and all that you know my just for instead of like going very wide all these TCP protocol stack I think it's very important to talk this question yes we need new protocols algorithms when we look at all the ABC cases because most of the algorithms we have from routing congestion control for whatever you think of they're mostly about you based on the classical like syntactic errors right like you know the digital the bits bit errors and all that but now we have all these B and C sections so all the algorithms may not help us to satisfy those B and C conditions so now is it looks like you and many others are really focusing on the you know kind of like channel all and you know end to end with some source coding but I expect many computer science people will jump on they will write thousands of papers on routing congestive flow control etc so it's a good question tolga and also he's up to sar you know he made so many comments I can't read all of them but what I distill from his comments is exactly goes with my statements and your answers he is again I assume he's a he he says we need performance metrics you know yeah a clear definitions of performance metrics right so I think we agree on that so hopefully the researchers will look into those problems and I think the Q&A session is over I didn't I hope I did not miss anything I really thank you Juchin excellent I thank you talking to you really fantastic you are a great person and I have a super future in the next decade I can assure you that please continue thank you so I ask Alessia to take over and continue our thank you very much Ian for moderating this session and thank you thank you Dr Keena for this very informative presentation so now welcome to the wisdom corner live life lessons which is based upon the idea to give a special angle to this new webinar series adding a personal touch so successful researcher like our speaker today will guide students and young scholars in the field of current IZT research and they will also share some impactful life lessons we always say that success is not because we never fail success is because we never give up so we would like to ask you which is your hard-earned life lessons or failure that you would like to to share with us today that might perhaps help someone attending this webinar okay thanks very much for the question that that's yeah that's well I don't think I'm a successful I guess Professor Akute will have much more experience to share how to be a successful researcher but for the failure yeah I did have many well well I think if there's only thing that I could share I think it's about yeah if you are really need to do a PhD I think not just myself maybe many of you guys I guess many of the audience have found the PhD or is doing the PhD start so are we really like I think many of us have asked ourselves oh how can I how can I make any progress for my research how can I get a PhD why I'm here why I started why I chose to do the PhD but I think I also got those feedbacks from some of the students so but I could feel it's yes sometimes we we are kind of fresh we maybe we're done we we got the problems we can't solve that we just really can't solve it and when we start and but I think it's all about the the interest especially like what I'm doing now like the the symmetric communication that's yeah that's actually some say really interest to me so I don't think like I'm limited or I'm asking myself okay I have to to find the limit for symmetric communication I have to solve the problems that like I'm don't think about it okay if I need to publish a good paper I think like if for my PhD student Huichang did the first work at that time there's not many actually there's very limited work related so we start from the spread so I think what motivated me is okay that's some say really interest and for me so I so that we enjoy the research even though we met a lot of problems we got rejected for papers for the grounds for a lot of things but yeah we know that some say we'd like to work on yeah I think that's that's what I like to share thank you so much I have another question which which strengths you believe are capabilities that students of young researchers should be most focused on developing and how do you think you would suggest that they could accomplish in this oh for the researchers here you mean the capabilities yeah I think they should yeah if we talk about the the PhD at the PhD level I think oh not just PhD whenever we start a new topic I think though the ability to perform the the proper literature review to not just read so many papers summarize them really understand the key contribution from the different groups of different researchers and what problems has been solved what problem was to to be solved and if the are we redevelop or re-investigate some problems that's literally solved many times before yeah I think that's like find the the essential the key problems to solve and understand what the others have have done this very important and without understanding that we can't do any further work yeah thank you and which fields in particular and which topics would you recommend students to study today well that's quite quite hard maybe I'm I'm not senior enough to provide that I don't want misleading people but I think if we want to I just limited to start about if you want to be to get ready for research on semantic communication you need a good understanding about the at least the general mathematical background is a physical layer communication and some information theory things yeah that's I think of but it's very big and the hard question I just limited to a very narrow yeah no worries but can you tell us which is the difference in the educational system in China and UK where where you studied sorry what what is the the difference that you see that you have experience in the education systems in your country in China and in UK where you studied oh wow teaching yes yeah well if we talk about the study I think that like Professor Akut is that I do really like greater because wow well for the I know that I can the system like at least for the PhD education system in in China or in US so all the students need to have some to take some models or take some courses at the first one or two years but in Britain they don't need to take any course they just start to work on the research topic directly so I think that's very different that just they need to learn whatever they need when they face a problem so I think we will never be ready to be knowledgeable to start a PhD you will always find that there's a lot of new problems or new knowledge to to learn so I think the the system may I was joking with many of my friends I think oh the Britain system they really suit like the diversity could be with the Britain system the the outputs from the students could be quite diverse and you know some students they could get the quite good achievement but then in the US also in China I think that the average would be I mean because they got a relatively longer period to do the PhD they got a five or even six years but in UK they normally do finish that within three to four years so on average I would say yeah they they they they are more experienced but I don't well it's it's really dependent you can't judge a student achievement based on like the the the the lens of the study or etc so yeah but I have just the least point out the difference I will just comment more on the I think both system has the the good point that we got brilliant researchers educated by all different systems yeah sure yeah you you mentioned achievements I would like to ask you if you can share with us some of you the most tangible contributions that you have made you believe you're made in your career that had a direct impact on your professional life or your personal life as well well I think I do feel that education not just for PhD like for the undergraduate student that's really I mean I never I never feel that that's proud after I received the message from my student okay I've got this super nice offer thank you very much for the support yeah I would never think about that okay why was the students I can't feel that I can't understand how much that mean for a teacher or for for for a lecture so I think the most I feel very proud of the I saw the the my students they they start from the beginning and they grow up they generate good results like many of them shared today so I'm really proud of the the yeah and I'm proud of myself to get this to educate and work together with the students yeah and then I saw they grow up excellent thank you very much I have a last a last question for you if you want to share with us before we close this webinar like a workflow that you believe in an aphorism or a book a movie music that you want to let's say describes you or your professional path that you would like to share with us okay yeah I'm I'm thinking about that I think the movie the the quite famous movie called the forest to come maybe I would know that yeah I watched that I can't remember how many times but that's really inspiring me and yeah you if you haven't watched that before I really encouraged you to check it I think like people we have up down in the life I mean during the study I must do yeah maybe I can't talk much about the life experience but during the study the research yeah we met good people we met we got a good topic we had a good timing to start the new research but sometimes yeah we got rejected I mean your ideas your papers etc but yeah never never give up you never know what life will bring you tomorrow it's like a box of the chocolate right yeah that's what I like from the movie thank you that's nice iana yeah you wanted to be in a please again thanks a lot really fantastic I personally enjoyed it and I assume everybody else enjoyed it I thank you are actually on behalf of the entire ITU team we thank you for taking time and presenting us your excellent research results and hopefully we'll be in touch and hope to see you somewhere right yes sure yeah thank you very much for the invitation and thank you very much for joining us today and Ian I would like to remind our attendees that our next and final webinar for this first series will be held on a 22nd of June at the same time and it will be uh professor Joseph Jordanet from Northeastern University who will talk about ultra broadband communication and networking solutions to unleash the terahertz band and so I hope to see you again online with us thank you so much uh Dgine for for your contribution we really enjoyed this talk yeah thank you a pleasure yeah thank you bye yeah thanks bye and I also thank the participants for taking time and listening to our speaker today I hope you benefited from this webinar and we have the last speaker as Alessia mentioned in three weeks however I have good news uh our series will continue in the fall we have outstanding six more speakers they're really outstanding like this uh phase and hopefully uh you will join us also in the fall thank you and enjoy your day and night or evening whatever you are yeah so thank you bye bye thank you bye bye