Have telcos got the right data in the right places to harness AI?

Summary

Mobile Europe's recent presentation delves into the transformative potential of AI in the telecommunications sector, exploring if telcos have the 'right data in the right places' to harness AI's full potential. The conversation is centered around the dramatic shifts induced by AI advancements, particularly generative AI, which pose significant opportunities and challenges for telcos. Key discussion points include the significance of preparing and utilizing data effectively, collaborations with educational and tech institutions, and leveraging AI for optimized customer experiences and network efficiency. The session also addresses concerns about the pace of AI integration and the evolving roles of telcos into tech-oriented entities.

Highlights

• Telcos are transitioning into more tech-oriented organizations, dubbed as 'techcos'. 💡
• AI is revolutionizing customer service through advanced chatbots like Swisscom's SAM. 🤖
• Telcos have massive amounts of data, enabling better network optimization and energy efficiency. 📊
• Knowledge graphs help telcos in infrastructure management and anomaly detection. ⚙️
• There's a massive potential in predictive maintenance and cybersecurity through AI in telecom. 🔒

Key Takeaways

• Telcos need to harness the power of AI with the right data, or risk becoming obsolete. 🤖
• There's a strong push towards transforming telcos into techcos to stay relevant. 📈
• AI can drastically improve network optimization and customer service in telecom. 📶
• Synthetic data generation is crucial for maintaining privacy while enabling research collaboration. 🔍
• The future of AI models is promising; telcos must be ready with sorted data to benefit fully. 🚀

Overview

In a revealing session by Mobile Europe, the focus sharply turns to how telcos can effectively integrate AI to stay ahead in the rapidly evolving tech landscape. The main inquiry challenges whether these companies have access to the correct data sets needed to leverage AI's full capabilities. This transformation is not just about technology but a shift towards achieving operational excellence and superior customer satisfaction.

The discussion sheds light on how Swisscom, a leading telecom operator, is pioneering the use of AI to reconstruct their traditional operations. By collaborating with universities and tech giants like Nvidia, Swisscom is not just implementing but also innovating with AI-driven customer service, network operations, and cybersecurity tools. These efforts position them at the cutting edge of what telecom services can achieve.

However, the conversation doesn't shy away from cautioning the industry: the data needs to be organized and accessible to fully benefit from future AI advancements. With the promise of more sophisticated AI models on the horizon, it’s a call to action for telcos to structure their data practices effectively to ride the upcoming AI wave.

Have telcos got the right data in the right places to harness AI? Transcription

• 00:00 - 00:30 so uh welcome everybody and I was really uh pleased and honored when I was asked to talk about to you today about AI transformation um and transforming Telos and especially on the question if Telos got the right data the right places in the right places to harness Ai and um I would like to start that with a little
• 00:30 - 01:00 uh picture that you see here which is inspired by the great um wave of kanaga which I guess you might see a little bit recognizing here but this is a geni Swiss skiing interpretation of that um so the great way for the tsunami that basically I think you know was hitting us all U globally um on the 30th of November 2022 so pretty much three in three days
• 01:00 - 01:30 uh two years ago basically uh raised a lot of questions about you know what especially AI gen AI large language models and how it will transform um our kind of business our world um our also private life and this is for somebody who has been very active in this area since quite some time not all the back that you from the history that you see here here but somebody who has seen at least
• 01:30 - 02:00 a little bit you know some of this waves and also of not only the you know the high sides of the weights but also a little bit of this winter part of the weights where basically everybody lost confidence in that this is this is something that is quite interesting to see how basically it's developed in the last years and also what we expecting uh to do so the question that drives me a
• 02:00 - 02:30 lot is will we be able to write this great wave that basically was hitting us pretty much now two years ago or will be somehow will be crashed by this wave and so the opportunities and risks that we have in front of us are enormous and they coming and you know influencing us in a high speed and with high energy and it's up to us what what we will do out
• 02:30 - 03:00 of that um when we go back to the question a little bit um I want to put it a little bit extreme and provoke a little bit with a [Music] questions are we the last generation to ask this question have Telos got the right data in the right place to harness Ai and the reason to ask that can be you know interpreted twofold way um because one could take the position that if we
• 03:00 - 03:30 are not maybe then Telos don't exist anymore as we basically know them today we might need to change or basically become so natural that you know this kind of question will get pretty much Irrelevant in the next uh generation and this is where I would also like to ask is the right the word Telco still the right word or are we not transforming more into to Tech Co and
• 03:30 - 04:00 then also the question is basically at least to be rephrased in the future and the three parts of this question that I was asked was basically to talk a little bit about you know um the raw data material um is it ready to to be used and what needs to happen to make it fully usable um I think this is I try to answer that a little bit in which areas
• 04:00 - 04:30 is the data the most advanced and which areas require more work I try to show this with some examples and also on the other side we are we not going too fast are we trying to run before we can walk and let me basically try to start answering that from three different perspectives that are very close um to me and my career and I will introduce myself a little bit more in the next three slides um on one one side
• 04:30 - 05:00 basically trying to answer this from the academic sector um fundamental research also from the public sector basically you know International uh regulation on AI it's a big Topic in that but also from the private sector and this is uh where my interest on Telos and this is where basically how this AI will play in the private sector plays out so my name is Daniel Doos I'm currently the research director of
• 05:00 - 05:30 swisscom I'm leading the swisscom digital lab together with one of my co-directors which does the AI research for Telos and the Tech Co swisscom since 2016 at epfl the University in laan um we're collaborating with other research organizations in Switzerland epfl eth but also have an outpost in Palo Alto in Stanford we do also quite a lot of Applied research implementation of state-of-the-art fundamental research
• 05:30 - 06:00 in AI but for the other perspectives I had the pleasure and honor to work with United Nations on strategic foresight especially on this topics of data and Ai and the future of it what can we expect there but coming originally from a very technical background where I was at particle physicist at CERN where my main day-to-day job was to deal with huge data bandwidths of something like 80
• 06:00 - 06:30 terabytes per second um something that even you know Telos for a long time had basically you know much less bandwidth um for for many years we had basically more bandwidth than the entire bandwidth of the internet now it's changed as you all know since uh video streaming quite dramatically but still the goal was always with this huge amounts of data to have ai and Advanced Data decisions very very fast real time you know we're
• 06:30 - 07:00 talking about making something like you know 40 to 200 million AI decisions so-called inferences per second um Quantum machine learning is also one part that I basically worked on um so this is basically I try to look at this question that was ask you know how basically we as Telco dealing with that from the three different perspectives and we started that basically this kind of Journey U with Swiss come with a partnership um very
• 07:00 - 07:30 very concretely in 2016 uh together with epfl where we basically set up the digital lab um this was built on collaborations in Switzerland with AI research which go back more than 30 years for example there was an um research lab in edab was called in Martini where since 30 years we are one of the partners and then basically we started to put 2018 2019
• 07:30 - 08:00 all eyes on this new technologies that basically is today the Transformers that basically where the you know the ground research work to allow all what we see today with chat GPT and others um basically you know this is sorry this is what I mentioned the research organization in in Martini in Switzerland EDF where we have been involved in this AI research since more than 30 years and since especially last two three years we looked more and more
• 08:00 - 08:30 how can we do together research applied research for Telco um with universities but then also with strategic Partnerships like with Nvidia that we announced with JY in the AI house in Davos um to answer basically you know where we are using AI this is on you know all kind of uh state-of the art uh technologies that we use in a
• 08:30 - 09:00 techco a Telco with you know all kind of techniques like natural language processing since a long time automated speech recognition uh distributed machine learning graph neural networks I will show you an example um animaly detection I will show you a little bit an example and then what we are using it for is really in all kind of areas to provide to to our customers a much
• 09:00 - 09:30 better user experience to understand our customers to basically operate and maintain our Network um to be able to use chatbots to communicate better with our customers personalize the customer experience but also then for our Workforce to do Advanced workforce management and do this in a cyber security and most secure Trust trusted
• 09:30 - 10:00 way um I'll show you most Pro most uh mostly today some um projects from the area where I try to answer this questions if we have the right data and if we are ready to use that uh from free focal areas with Mo mostly small smart operation and maintenance customer chatbot and then customer understanding and here you have a list um of some master fees works that we have been doing doing now uh one year ago together
• 10:00 - 10:30 with research organizations in Switzerland most prent maybe epfl in loan one of the top 20 universities worldwide but also with more applied University and to give you here an overview basically of different topics I'll leave it here for a second but you don't even have to read through because some of the highlights I will I will show you out of that and to get
• 10:30 - 11:00 basically a good understanding of of you know what kind of projects we are working on um I would highly recommend you in Switzerland a conference which is a biggest applied uh machine learning conference in Switzerland it's called applied machine learning days AML next Edition will be in February uh next year
• 11:00 - 11:30 and we will cover there mostly topics Ai and cyber security but also Ai and and connectivity and last year we have been and I will show you some of this kind of results that we achieved there to answer this questions that we that we started with is that we showed the power of data and revolutionizing mobile network optimization by two colleagues we looked into you know generative Ai and how it will revolutionalize Swiss come uh
• 11:30 - 12:00 customer care and the third one that I really will highlight today is a little bit so what are the what are the impacts of knowledge graphs and knowledge graph that are really huge and you know going above normal uh applications that exist in many other companies that are not techos and and Telos so one project that basically um starts with with answering if we have
• 12:00 - 12:30 the right data and to do better customer experience yes the amount of data that we as techos Telos really have goes beyond what most other companies organizations um have know even many cases way way too much and filtering and smart filtering compliant filtering understanding for which purposes we can
• 12:30 - 13:00 use which kind of data is is a key aspect and how it can be used basically for you know providing a better service is in the very simple case you know really understanding how good is our service in which areas of the countries how can we optimize the roll out of for example in this case mobile antennas with help of machine uh learning how can we really optimize this kind of coverage
• 13:00 - 13:30 maps that we have in a way that we constantly you know improve our Network and also move to the smartest Network which means also the most sustainable and energy uh saving kind of network how can we for example um optimize the use of cells in the night and maybe you know switch off parts that are not needed to save energy
• 13:30 - 14:00 um the second project that was highlighted there uh I would like to highlight here a little bit is about finding event sequences in large data streams and you can imagine that there are multiple different use cases from that this is purely from the network uh optimization um this is from outlier detection uh when one talks about security and cyber security so basically we see that in many different use cases
• 14:00 - 14:30 as something where we look at it first as a stream of data that maybe we don't even have to fully understand at the very beginning but we can start to look for sequences of events which maybe know a human would not even find as something that is obvious to look for but that basically are uh indications of certain
• 14:30 - 15:00 events this can be for predictive maintenance this can be for something where we can avoid all kind of instabilities in the network by changing components or replacing components early enough this can be in the area of cyber security where some special sequence and maybe even timely distributed sequences are indication of some issues or or potential threats or can can be really in you know optimizing our um customer
• 15:00 - 15:30 experience in the network and this is where basically this kind of loop architecture really helps us to on one side find this kind of events but always keep the experts in the loop um large language models I think was discussed a lot and uh also basically heard about uh retweet R Rex so ret augmented generation this is something where which is extremely
• 15:30 - 16:00 important us for customer care dialogues so how can we automate basically the customer interaction even more so that our customers have a good experience even if it's whatever Sunday morning at 5:00 or maybe especially if it's Sunday morning 5:00 um where basically they need our support in solving issues or challenges that they have and we made very
• 16:00 - 16:30 excellent experience with a llm um based chatbot name is Sam where we not only have seen that this helps us to automate many process that were very difficult to automate before but that it also dramatically um increased The NPS the the satisfaction of our customers of what we can do in self-service this is a double-sided sword um I also have to put maybe as a warning because once you
• 16:30 - 17:00 start using this Advanced AI then also the expectation of the customers are rising with basically every kind of um you know success story and then basically the expectation is also very fast big that this kind of um large language model based chatbots can then help them in any given situation um I mentioned the that knowledge graphs is
• 17:00 - 17:30 something that we are very interested in and and and exploring um extensively because with our huge infrastructures that you that you all know it is very difficult to understand you know what kind of issues our infrastructures might have and so making multilevel abstractions of this huge amount of um notes and edges so with dots and lines that you see here that
• 17:30 - 18:00 describe our infrastructure so that the human can understand was a topic that basically was presented and now really allows to understand much easier and much more in a human natural understanding how our infrastructure is built up or where we have anomalies in this kind of infrastructure so this is where graph similarity comes in so that we kind of group how our infrastructure is rolled out how it's looks up and then
• 18:00 - 18:30 identify if we have some kind of situations where we installed or operate our infrastructure in a different unusual way maybe in a more risky way or maybe in a in a way which uh requires more maintenance and to understand this in this huge amount of data this is where graph knowledge graphs and with understanding of similarity helps us a lot I also use maybe last uh few minutes to
• 18:30 - 19:00 show you some uh current research topics and cannot show you also the we try to answer this question that what was initially asked if you're ready and if you're not going too far um I really want to show you some state-of-the-art research but this also means that that I cannot show you the full results yet so you will see some of the pictures still blurred but in the next weeks you know can tell you more details about that so
• 19:00 - 19:30 at the moment we are really focusing on the on the with something like 20 Master students per year on how we can use AI for our infrastructure the maintenance and energy saving with conversational AI on you know generative AI large language models um as mentioned before on the strategy and Business Development on all kind of data analytics and then also for data Ai and the interaction with security cyber
• 19:30 - 20:00 security how to do secure analytics and collaborations and you see some of the um students that just finished uh successfully finished their Master feces with us working on on such topics um one point that I would like to add your attention to is synthetic tblo data generation with large language models so as Telos as techos uh collaboration with researchers with
• 20:00 - 20:30 other companies with startup is something that is very very difficult because by regulation in most cases we cannot share data for very good reason don't misunderstand me it is very um for very good reason we cannot share data but this makes also collaboration difficult and this is where synthetic data generation can help so that one can create data that looks similar enough so that researchers scientists security researchers can look
• 20:30 - 21:00 at this data and make conclusion out of this and this is where basically we look into all kind of cons considerations of bias in this data so do we have in this data some um biases like gender or are there some privacy risks with this kind of synthetic data generation here intentionally blurred you see somebody kind of uh data distribution from an origin data set and then basically an
• 21:00 - 21:30 large language model on the right side created synthetic data set that looks very very similar has most of the features that basically this original data set has but are totally um secure in terms of privacy concerns um graph anom detection this is something that we show you we presented um at the last applied machine learning days and will present again the the goal is really to see how we can work with this kind of
• 21:30 - 22:00 huge databases of Knowledge Graph databases and find anomalies in there um I told you a little bit about the energy saving this is where we also did quite some advantage of seeing how can we do additional um data savings sorry Energy savings on our networks and how we can use all kind of um how how can we use all kind of
• 22:00 - 22:30 strategies machine learning uh strategies to reduce the energy consumption that we have there so to try to summarize a little bit um the last minute um do we have the right data sets in the right places um I cannot answer for all Telos obviously and I um as you also know there's a huge variant between different
• 22:30 - 23:00 techos Telos worldwide for swisscom and some other partners that have been talking about I can say that we are really very close to that that we have a huge opportunity to put um Ai and generative AI really in in the use of the benefit of our Core Business that uh we are exploring and operating some of this um already today and exploring the future
• 23:00 - 23:30 and my advice to everybody who feels that they don't have the right data in the right place is it is not too late but it's it's becoming close to be basically really challenging if you don't start to put now your data really available for all kind of AI algorithms accessible and explorable then it will be very tough to you know uh in the in the future to really to really keep up and and so the
• 23:30 - 24:00 time for the ground work basically is today well uh that was really interesting Daniel that's uh life right in the middle of it all um so we have a couple of questions from the audience uh the first was from massud who asks is there any search uh any research going on um to secure Telecom sites and to prevent theft I
• 24:00 - 24:30 um you I don't know yeah so I I can answer with you know let me try to answer in two two parts I think in uh in Switzerland this is not an a big issue as a country where um where we have very rare cases of of situations where this is necessary talking to some some other colleagues yes this is always a a concern and where
• 24:30 - 25:00 basically there is a lot of research going into monitoring and you know early detection of this kind of risks um not not not in connection with AI I would say this are more classical kind of data analytics or surveillance techniques or you know yeah maybe in the area of um where we could add some AI is
• 25:00 - 25:30 if you want to do some automatic analysis of video recordings of a site where AI can really help to you know go through a waste amount of sites a waste amount of of video recordings and identify uh early indicators of of of ffts or security issues before okay thank you and Alexandra um congratulates you on an excellent presentation and asks uh can you tell us
• 25:30 - 26:00 if and how the Swiss government supports fundamental AI research for toos um yes with great pleasure and thanks for this wonderful question um I I think I need to modify the question a little bit and try to answer it in two parts and hope that helps to get a better answer um the what I would try to separate out is fun FAL AI research I
• 26:00 - 26:30 think the Swiss government is heavily supporting fundament AI research for the universities and Switzerland has a huge tradition of having excellent research on AI is that directly applicable to to Telos in many cases not and we see now the results of good um fundamental AI research support from five or 10 years ago do they support the use case and the
• 26:30 - 27:00 applied research of AI in uh in Telos yes that's at well but not at the same time so these are basically two different streams of funding and support so there's one fundamental support for the universities and then there's an applied support for the Telos okay and we have a f I'm afraid we've only got you need to be quick with this I'm sorry Daniel we're out of time but toan asks why is now the time to locate data appropriately for AI what
• 27:00 - 27:30 clock is ticking I guess just Tiff I don't know I I try to answer that very very quickly is because um we are just at the beginning um it is we will see in the next years that there will be more and more and better and better models if you if you now have basically your data sorted out so that you can go through this follow this kind of evolution of better and better models you will be
• 27:30 - 28:00 able to to change uh without any issues through this kind of better and better models so it's a little bit the time of uh Teenage laugh I would say basically you know sort your sort out your uh data um date a lot of different AI models in the in the next years Fall In Love sometimes but don't marry yet because you will basically go through a lot of different AI models in the next years um so having
• 28:00 - 28:30 basically the data sorted a case being being ready to exploit opportunity yes as it comes up and as it suits your particular business and situation I guess yes I I definitely agree to it Daniel that was great thank you very much thank you to our audience for their questions um yeah I think uh we we look forward to catching up with you further
• 28:30 - 29:00 down the line and see um see how it's all going thank you so much