Towards 6G and Beyond
CSIT Seminar · Western Sydney University · 27 October 2026
Thanh Tung Vu
Lecturer in Computer Networking · ARC DECRA Fellow (2026-2029)
Script
I'm Tung. I work on wireless networks, and today I want to show you why they matter for the AI you build.
Optional, pointing at the map: Keep an eye on this little world. We'll come back to it.
Background
The opening introduces the speaker and the question of the talk: why wireless networks matter for the AI that people build. The map is the small isometric world used throughout the talk, with a city, a crossroads, a park, a flooded town, a factory, a farm, the sea and the sky. Its wireless links are not drawn yet; they light up on slide 3.
The future is already being built
Five examples of many, each already in trials or early use
An assistant that knows you, without your data leaving your phone
Cars that see around corners by sharing what they see
Drones that find people after a flood
A connection anywhere : the outback, the ocean, the sky
A factory line that reconfigures itself
AI is moving out of the data centre and into the physical world
Google · Gboard · US DOT, Univ. of Sydney · road trials · SA SES · search drones · Starlink · Royal Caribbean, Qatar Airways · Mercedes-Benz · Factory 56
Script
Many of us in this room build models, and we treat the network as a pipe that is simply there. Let me show you five places where it isn't. Here's a future that's already being built.
[click, five times, one line per scene, as written under each tile] An assistant that knows you, without your data leaving your phone. Cars that see around corners by sharing what they see. Drones that find people after a flood. A connection anywhere: the outback, the ocean, the sky. A factory line that reconfigures itself.
[click] None of these is science fiction. Each one is already in trials or early use. AI is moving out of the data centre and into the physical world.
Background
Five examples of AI moving out of the data centre, each described by what it does for people rather than by its technology. They are at different stages, which is why the subtitle says trials or early use . Google's Gboard trains its language models with federated learning on phones. Starlink provides connectivity on Royal Caribbean ships and Qatar Airways flights. South Australia's SES flies thermal-camera drones with AI people detection for search and rescue in general, not only after floods. Mercedes-Benz Factory 56 has run a private 5G network with automated guided vehicles since 2020; the source says production can switch between models in days. A line that reconfigures itself, without people replanning it, is where this is heading; slide 24 returns to it as an open problem. Cars that share what their sensors see are still in road trials, for example by the US Department of Transportation and the University of Sydney.
Sources: Google Research · US DOT / FHWA · University of Sydney (Gizmodo) · SA Government · Satellite Today · Business Traveller · RCR Wireless
Every one of these runs AI over a wireless link
They move, or they are far from any cable
Wireless is what holds this future together
Script
[while the map zooms out] Look at what they all share: the same five places, on one map.
Every one of them runs AI over a wireless link, because they move, or they are far from any cable. That's what 6G and beyond are expected to do: carry AI as a core service.
[click, once all the links have lit] Wireless is what holds this future together.
Background
The five scenes of slide 2 sit on one map, and every one of them depends on a wireless link: phones, cars, drones and factory carts move, and the farm, the ship and the plane are far from any cable. On the farm, the AI is the robot tractor that senses the crop as it drives. For the ship and the plane, satellites provide the link, as they already do for some airlines.
Source: Business Traveller
The wireless link is the bottleneck
A model only helps if its data and answers get through, in time
Bandwidth
Drone video and phone updates share one channel
Delay
The warning must arrive before the car reaches the corner
Reliability
A truck, a hand or a passing satellite can cut the link
Energy
Every bit sent drains a battery
Networks for AI : design the network around the AI task
Script
And every one of these systems hits the same limits, and they are set by radio resources.
[click] Bandwidth : Every drone's video and every phone's update shares one channel.
[click] Delay : The warning has to arrive before the car reaches the corner.
[click, then wait for the truck to cross, about five seconds] Reliability : A truck drives between the small cell and the phone, and the link is cut until it passes. A hand, or a satellite moving out of view, does the same.
[click] Energy , pointing at the phones' batteries, then the drone's: And every bit sent drains a battery. For a drone, a slow link also means more time in the air.
[click] So: networks for AI. Design the network around the AI task.
Background
The link that holds this future together is also its weak point. The four limits are properties of the radio channel: how many bits it can carry (bandwidth), how long they take (delay), whether they get through at all (reliability), and how much battery it costs to send them (energy). Each card reuses a scene from the map.
For a drone, most of the battery goes to flying, not to the radio; the radio matters through time, because a drone that must hover until an upload finishes pays for every extra second a slow link costs. A satellite cuts a link by moving out of view, not by blocking it the way the truck does. Privacy matters to AI but is a rule about where data may go rather than a property of the channel, so it is not on this slide; it returns with federated learning. The closing line names the first of the talk's two directions, networks for AI: design the network around the AI task.
And AI can help run the network itself
The network must decide every millisecond, faster than classic optimisation algorithms can
Networks for AI
The network serves the AI task
AI for networks
A trained model decides who transmits, and when
Two directions, one question: who gets which resource, when
3GPP TS 38.211 · 5G NR frame structure and slot timing
Script
And it works the other way round too: AI can help run the network itself.
Pointing left: On the left, what we just saw: the network serving the AI task.
[click] On the right, the network's own decisions. This access point serves two robot arms, two carts and two humanoid robots, one at a time. A trained model decides who transmits, and when, every millisecond, faster than a classic optimisation algorithm could.
[click] Two directions, one question: who gets which resource, when.
Background
The second direction, AI for networks. Here AI helps run parts of the network, such as scheduling users and choosing beams; it does not run the whole network. It is needed for speed: a 5G network schedules users slot by slot, each slot a millisecond or less, and classic optimisation algorithms are too slow for that on a large network. The strip of six slots above the access point shows whose turn it is. Both directions come down to the same question, who gets which resource, when, which the rest of the talk follows. Slide 12 shows a trained model making a slower decision of the same kind: which access points serve whom, and with how much power, as people move. Today's schedulers meet the slot deadline with simple rules such as proportional fair; what classic optimisation cannot do in time is solve the problem well on a large network.
Source: 3GPP TS 38.211
One formula sets how many bits get through
Each limit we just met, and each resource we share, is a term in it
C = B log 2 ( 1 +
P g I + N 0 B )
B
The channel the drones share
P
Paid from the battery
g
What the truck cuts
I
Signals from someone else, sent at the same time
bits ÷ C
The time a message takes to send
Who gets which resource decides how many bits get through
Shannon · A Mathematical Theory of Communication · 1948 · Tse & Viswanath · Fundamentals of Wireless Communication · 2005
Script
If you remember one formula from today, it's this. It sets how many bits can get through a wireless link.
[click] B is bandwidth: the channel the drones were sharing.
[click] P is power, and every bit of it comes out of a battery.
[click] g is the channel gain: what the truck cut.
[click] Pointing at the two small cells: I is interference. These two access points send at the same moment, and each one also lands on the other's phone.
[click] And the time to send a message is its size divided by C: that is the delay.
[click] So who gets which resource decides how many bits get through.
Background
The formula is the rate of a wireless link when interference is treated as noise. Shannon's 1948 result has only signal over noise; adding the interference I to the denominator is the standard textbook form. C is an upper limit on how fast bits can get through reliably, not what a real system achieves. B is the bandwidth, P the transmit power, g the channel gain, I the interference and N 0 the noise power spectral density; the time to send a message is its size divided by C . Each term is tied to a limit from slide 4, shown by the same small scenes.
Every term except the noise is shaped by the network's decisions: bandwidth and power are allocated directly, while beams and scheduling raise g for one user and lower I for the others. Taking turns, as in slide 5's factory, avoids interference but spends time; sending at once, as in the interference card, uses the time but adds interference. The formula returns in each line of research: far from a tower g is small (slide 9); millimetre waves bring a large B but a fragile g (slide 14); and in federated learning every round sends a model update both ways, so training time depends on C (slide 20).
Sources: Shannon, 1948 · Tse & Viswanath, 2005
Many antennas steer energy to each user
Higher g for each user, lower I for everyone else
One access point can serve many users at once, if we share power and beams well
Marzetta · Noncooperative cellular wireless with unlimited numbers of base station antennas · 2010 · Larsson et al. · Massive MIMO for next generation wireless systems · 2014
Script
With one antenna, an access point sends its energy everywhere. Most of it misses the phone it is meant for, and lands on everyone else.
[click] With two antennas it can form two beams at once, but they are wide. One serves this phone; the other has to cover these two, so they interfere.
[click] With 64 antennas, as on today's 5G base stations, it forms narrow beams: one for each phone, all at the same time. Higher g for each user, lower I for everyone else. Many antennas serving many users at once is called massive MIMO.
[click] So one access point can serve many users at once, if we share power and beams well.
Background
The same park as slide 6, with one access point and three phones. With one antenna, the access point sends its energy in all directions. With two antennas it can form two beams at once, but both are wide, so with three phones two of them share a beam and interfere. With 64 antennas, an 8 × 8 panel as on today's 5G base stations, it forms a narrow beam for each phone, all at the same time. The hardware on the pole changes with each step.
Each beam is drawn as the real main lobe of the array: the power pattern of antennas half a wavelength apart, steered at its phone, times the pattern of a single panel element, which sends most straight ahead and nothing along the panel's face. It is drawn from one first null to the other; the side lobes beyond are left out. Every beam takes a share of the total power, hence the closing's if we share power and beams well . A beam from one tower still weakens with distance (slide 9), and a narrow beam is easy to block (slide 14).
Sources: Marzetta, 2010 · Larsson et al., 2014
What makes a network that AI can rely on ?
Where my work takes the question of who gets which resource, when
What I've done
Cell-free massive MIMO
Which access points serve whom, so no one is left at the edge
What I'm doing now
My DECRA: cell-free mm‑wave
Which access points keep serving when a beam is blocked
Where I'm heading
Federated learning across space, air and ground
Which devices train, and when, as satellites come and go
Where we meet
Working together
Models and networks, designed together
Sharing well is what makes a network that AI can rely on
Script
Sharing power and beams well, deciding who gets which resource, when: that is where my work lives. I build wireless networks AI can rely on.
I did my PhD at Newcastle, then postdocs in Belfast, at Linköping, where I worked on two Ericsson patent applications, and at Macquarie; now I'm a DECRA fellow here.
Three lines of research, each taking that question somewhere new. What I've done: cell-free massive MIMO, deciding which access points serve whom, so no one is left at the edge. What I'm doing now, my DECRA: which access points keep serving when a beam is blocked. Where I'm heading: federated learning across space, air and ground, deciding which devices train, and when, as satellites come and go. And the part I care most about: designing models and networks together, with you.
[click] Sharing well is what makes a network that AI can rely on.
Background
The bridge from the vision into the speaker's research, after the earlier slides have shown why the network matters. Each of the first three cards is one line of research in the same world. Cell-free massive MIMO: two groups of access points, each serving its own pair of phones, the who-serves-whom decision of slide 11 (real groups can overlap; they are kept apart here for clarity). The DECRA project, a three-year ARC Discovery Early Career Researcher Award: when a truck blocks one small cell, the others keep serving. Federated learning across space, air and ground: satellites linking the farm, the ship and the plane. Working together: the whole connected map, with models and networks designed together.
Federated learning over ground networks is earlier work (slide 22); taking it across space, air and ground is a direction and an open question (slide 23). The speaker's background is in the notes of slide 25.
At the cell edge , service drops sharply
Far from your tower, its signal is weak and the next tower interferes
The edge exists because one tower serves each user
Script
You've all had one bar of signal at the edge of a cell. That's not bad luck, it's architecture.
As the walker leaves the tower: Today one tower serves you. Walk away from it, and its signal gets weaker. Meanwhile the next tower is busy serving its own user, just across the edge, and its signal to them reaches you too, as interference. Watch the speed bar.
And if that phone is training a model with others, each round can end up waiting for it.
[click] The edge exists because one tower serves each user.
Background
The first line of research, cell-free massive MIMO, starts at the cell edge, where sharing fails first. Two macro towers each serve their own patch of ground; the dashed line between them is the cell edge. A person walks from tower A toward the edge while tower B serves its own user just across it. Both terms of slide 6's formula go wrong together: the walker's signal g weakens with distance, and tower B's signal to its user arrives as interference I . Service drops sharply but does not stop.
The speed bar is computed from slide 6's formula, with signal falling as distance to the power 3.5; it shows the trend, not a measured rate. Placing tower B's user near the edge is the worst case: had that user been on the far side, a beam (slide 7) would point away from the walker and interfere much less. Real cells serve many users each; one per cell keeps the picture readable. In synchronous federated learning, the basic form used on slide 20, every training round waits for the slowest device, so a phone at the edge slows training for everyone.
With cell-free, many access points serve each user together
What used to interfere now also carries your data
No one is left at the edge, if we choose which access points serve whom
Ngo et al. · Cell-Free Massive MIMO Versus Small Cells · IEEE Transactions on Wireless Communications · 2017
Script
Same street, same two people, same walk. But now take the antennas off the two towers and spread them out: small access points on lamp posts, all wired to one central processor. That wire is called the fronthaul.
As the walk plays: The access points nearest to you serve you together, and they change as you walk, with no boundary to cross. The one that used to interfere now also carries your data. Watch the speed bar: it stays high.
This is cell-free massive MIMO: the many antennas of one base station, spread along the street.
And it's not just more hardware. Ngo and colleagues compared the same access points used both ways. Cooperation alone, with no extra hardware, gave the worst-served users about five times the throughput.
[click] No one is left at the edge, if we choose which access points serve whom.
Background
The same street, the same two people and the same walk as slide 9; only the architecture changes. The towers give way to eight small access points on lamp posts, wired by fronthaul to a central processor, and each user is served by its four nearest access points together, so there is no boundary to cross. The grey tick on the speed bar marks where slide 9's bar stood at the same spot.
The speed bar uses slide 6's formula with the same total power as the two towers, shared over the eight access points, each splitting its power among the users it serves. The walker's access points add up coherently and only the share they send to the other person still interferes, so the bar stays high rather than full. The picture also moves the antennas closer; the fair comparison is the paper's, which keeps the access points fixed and compares small cells, each user served by one access point, with cell-free: the 95%-likely per-user throughput, the rate the worst-served users get, was nearly five times higher with uncorrelated shadow fading and about ten times with correlated shadow fading.
The first cell-free papers had all access points serve all users; practical, user-centric cell-free gives each user a nearby group, as here. The central processor and the fronthaul matter later: slide 11 has a fronthaul limit, and the central processor is compute at the edge of the network (slide 13).
Source: Ngo et al., 2017
The hard part is deciding who serves whom
Like scheduling jobs on shared servers, where every choice changes everyone's speed
The optimisation problem
max 𝒂 , 𝜽 ∑ k ∈ 𝒦 SE k ( 𝜽 ) total speed s.t. a m k ∈ { 0 , 1 } , ∀ m , k who serves whom, yes or no θ m k 2 = 0 if a m k = 0 , ∀ m , k no link, no power ∑ k ∈ 𝒦 θ m k 2 ≤ 1 , ∀ m power limit ∑ k ∈ 𝒦 a m k SE k ( 𝜽 ) ≤ C max , ∀ m fronthaul limit ∑ k ∈ 𝒦 a m k ≤ K ˆ m , ∀ m user limit SE k ( 𝜽 ) ≥ SE QoS , ∀ k minimum speed ∑ m ∈ ℳ a m k ≥ 1 , ∀ k everyone is served
where SE k ( 𝜽 ) = τ c − τ p τ c log 2 ( 1 + SINR k ) speed of user k SINR k = ( ∑ m ∈ ℳ ρ d ( N − | 𝒮 m | ) σ m k 2 θ m k ) 2 ∑ l ∈ 𝒦 ∑ m ∈ ℳ ρ d θ m l 2 ( β m k − δ m k σ m k 2 ) + 1 all users' power,
8 × 5 yes-or-no links: 240 ≈ a trillion choices
Solving it well is too slow for users on the move
Hao, Vu et al. · Joint User Association and Power Control for Cell-Free Massive MIMO · IEEE Internet of Things Journal · 2024
Script
Now put more people on the street. Each access point has to share its power and its fronthaul among the people it serves, so the network has to decide who serves whom.
As the links flip: Watch one link switch. It's not just one speed bar that moves: the neighbours' bars move too, because power and interference are shared. It's like scheduling jobs on shared servers, where every choice changes everyone's speed.
[click] Here is the problem as we formulated it. We choose a , who serves whom, yes or no, and theta, how much power each link gets, to maximise the total speed.
[click] Each access point has a power limit, a fronthaul limit and a cap on users; each user needs a minimum speed and at least one access point.
[click] And here is why it's hard. Each user's speed depends on a ratio: their own signal on top, and underneath, the power sent to every user, from every access point. So no choice is local: change one person's power and you change everyone's speed. Point at the double sum in the denominator.
[click] On this one street, that's two to the forty, about a trillion ways to choose who serves whom, before we even choose the powers.
[click] Solving it well is too slow for users on the move.
Background
With more people on the street, each access point shares its power and fronthaul among the people it serves, so the network has to decide who serves whom. Five people are each served by their three nearest access points; every second one link switches on (thick red) or off (grey, with a cross), and the speed bars of two or three people change, not one. The bars follow slide 6's formula and show the trend, not measured rates.
The card is problem (21) of Hao, Vu et al. (2024), with constraints (8) and (15) to (20), and on the second step the speed itself, Eqs. (13) and (14), in the paper's symbols: a θ θ ≥ 0 is left off to fit. Beamforming is fixed in this paper (local partial zero-forcing), not optimised. The objective is total speed; the minimum-speed constraint protects the user at the edge.
In the SINR, τc and τp are the coherence block and pilot lengths, ρd each access point's maximum power, N the antennas per access point, 𝒮m m serves with zero-forcing, σ²mk mk mk ml l appears in user k 's SINR, which is why every choice changes everyone's speed.
The paper calls the problem mixed-integer nonconvex, with a tight coupling between the binary a and the continuous θ , visible in the fronthaul constraint where a multiplies SE(θ ). The 240 counts every yes-or-no pattern on this street, an upper bound since the constraints rule some out; the paper's smallest network, 25 access points and 7 users, already has 2175 . The paper's optimisation method, successive convex approximation (SCA), took about 26 seconds for 25 access points and 7 users, and two and a half to four minutes for 150 to 300 access points and 40 users, on one CPU core (Tables III and IV). The inputs are large-scale fading, which changes as people move.
Source: Hao, Vu et al., 2024
A neural network decides 1000 times faster
Trained on our slow algorithm's solutions, it keeps about 95% of the performance
Teacher
Our SCA algorithm
26 s for 25 access points and 7 users
→
Training
28,000 solved problems
A CNN learns to give SCA's solution
→
Student
A trained network
2 ms for the same problem
A slow algorithm can teach a fast network
Hao, Vu et al. · Joint User Association and Power Control for Cell-Free Massive MIMO · IEEE Internet of Things Journal · 2024
Script
So how do we decide fast enough? We let our slow algorithm teach. Point at the left picture: We designed an algorithm for this problem, called SCA. It works, but it's still running on this street while people walk away: for 25 access points and 7 users it takes about 26 seconds.
[click] So we ran it offline, 28,000 times, and used its solutions to train a neural network: a CNN that takes the grid of channel strengths and gives two answers, how much power, and who serves whom.
[click] Given a new street, the trained network answers in 2 milliseconds, and it keeps re-deciding as people walk. That decision follows how people move, so it does not have to be remade every millisecond. In our simulations, twenty-six seconds down to two milliseconds: more than a thousand times faster, at about 95% of the total speed.
[click] A slow algorithm can teach a fast network.
Background
The answer to slide 11's closing, and slide 5's AI-for-networks direction in practice: a trained model decides who serves whom, and with how much power. The three pictures run left to right: the teacher, the paper's SCA algorithm, still working while people walk; the training, a CNN that takes channel strengths and gives two answers; and the student, the trained network re-deciding as people walk. The grids are illustrative, computed from the street, not taken from the paper.
SCA (successive convex approximation, Algorithm 1 of the paper) finds stationary solutions at a high computational cost, so its outputs are good solutions rather than guaranteed optima. The network, JointCFNet, takes the large-scale fading matrix (access points by users) through convolutions and three ResNet blocks into two heads with a sigmoid each, one for the powers θ and one for the association a . It is trained by supervised learning, mean squared error against SCA's solutions with both heads weighted equally, using Adam, batches of 512 and up to 400 epochs; generating the 28,000 examples took SCA about 24 hours.
Both methods were timed on one core of the same CPU (Table III): 25.7 s for SCA against 0.002 s for the network with 25 access points and 7 users, a ratio of 12,850; 19.2 s against 0.002 s with 36 and 5, a ratio of about 9,600. The paper's introduction describes this as three orders of magnitude. The network reaches about 95% of SCA's median sum SE (30 against 31.5 bit/s/Hz) with 25 and 7, and up to 97% with 36 and 5.
Two limits. The network's raw output can violate a constraint, so a correction step rescales powers and speeds; even then the decisions are feasible in up to 95% of cases, and the paper scores the rest as zero speed, so the figure of about 95% of the total speed already counts them as zero. And the network was tested only on small networks; for 150 to 300 access points the paper uses another of its algorithms, accelerated projected gradient (APG), about ten times faster than SCA. All results are from simulation.
Why this architecture, if asked (Sec. III-B of the paper). The input is the large-scale fading matrix, access points by users, which is easier to collect than instantaneous channel state and changes only as people move. Convolutions with 3 × 3 kernels and shared weights keep the parameter count small, where a fully connected network grows quickly with the network size. Stride 1 and zero padding keep every layer at access points by users, so each output cell is one access point and user pair. Each of the three ResNet blocks runs a convolution, batch normalisation and ReLU, then a second convolution and batch normalisation, adds the block's input back through the skip connection, and ends with a ReLU. The skip connection carries the earlier features forward, in the paper's words “to fully make use of all the features from the original LSF coefficients”, and batch normalisation eases training. One shared trunk feeds two heads because power and association are coupled. Earlier networks for cell-free power control learned one variable, and this design can take one head per variable. The sigmoid keeps both outputs between 0 and 1: the powers θ are normalised coefficients, and the association head a is trained toward SCA's yes-or-no labels. The paper describes no separate rounding step. The size was found by search: the paper varied the number of ResNet blocks, the kernel size, the number of filters (64) and the activations until the training error met the target. It needed 28,000 examples, against 2 million for a model the paper cites.
If asked why convolutions suit a grid whose row and column order is arbitrary: the paper chose a CNN to keep the model small, and it matched SCA on fixed small networks. A CNN is not invariant to reordering the access points or users, and it is trained for one network size, so do not call the design optimal or say it works for other sizes. Graph neural networks, which treat access points and users as nodes, are a natural next step, and the paper cites one for power control.
Source: Hao, Vu et al., 2024
Open question: who answers each LLM request ?
Each request needs a radio link and an edge server, and both are shared
The open question
For each request, choose the radio and the compute together
Choose
which access points carry it which edge server runs the model Shared
radio: power, fronthaul compute: memory, queue Goal
every answer arrives quickly
Radio and compute have to be decided together
Script
Now let's look ahead. The same street, but now these phones are asking an AI model something. Point at the speech bubbles. Each request rides the radio to the nearest access points, then the fronthaul to an edge server that runs the model. And those servers have queues.
[click] So for every request, someone has to choose the access points and the server, and both are shared. Earlier I said this was like scheduling jobs on shared servers. Here the jobs really are requests, and the servers really are GPUs.
[click] Radio and compute have to be decided together. If you work on serving models, I'd love to work on this with you. Next: what happens when the radio link itself is fragile.
Background
The open question of the cell-free research: running large language models at the network edge. It is a question, not a result. The same five people now send prompts to an AI model (the speech bubbles); each request travels over the radio to the nearest access points, then over the fronthaul to one of two edge servers, whose queues fill and empty.
Slide 11's comparison with scheduling jobs on shared servers becomes literal here: the jobs are requests and the servers are GPUs. Choosing who serves whom now has a compute half, which server and when, and the radio half changes how long the upload and the answer take. Deciding them separately can leave a fast link waiting on a busy server, or a free server waiting on a slow link, hence decided together . Two things make it harder than slide 11: answers are generated piece by piece, so how long a server is busy depends on an answer length nobody knows in advance; and each server can hold only so many requests in memory at once.
The cell-free research thus ends with both directions from slide 5: slide 12 is AI for networks, and this slide is networks for AI. Whether slide 12's approach, a slow algorithm teaching a fast network, also helps here is untested.
Millimetre waves bring far more bandwidth, but are easily blocked
So today they stay in small hotspots, not across a wide area
Far more bandwidth
One mm-wave band is 6.5 times as wide as today's 5G mid-band
Easily blocked
Vehicles, people and buildings cut the link
The bandwidth is there but reliability is not
3GPP TS 38.101-1, 38.101-2 · bands n78, n258 · Bai et al. · IEEE Transactions on Wireless Communications · 2014 · Jain et al. · IEEE JSAC · 2019
Script
The second line of research is about millimetre waves, the high frequencies in 5G and 6G. Look at the scale. Point at the sliver on the left. This is today's 5G mid-band: 500 megahertz. And this is one millimetre-wave band: 3,250 megahertz, six and a half times as wide.
[click] The catch. Remember the truck? At millimetre waves, that happens easily. The wavelength is about a centimetre, so the signal doesn't bend around the truck or go through it, and the narrow beams from earlier are easy to cut. Vehicles, people, buildings: any of them can cut the link. So today, operators use millimetre waves only in small hotspots, not across a wide area.
[click] The bandwidth is there but reliability is not.
Background
The second line of research is about millimetre waves, the high frequencies used in 5G and planned for 6G. The bands on the left are 3GPP definitions, drawn to scale: n78, a common 5G mid-band, runs from 3.3 to 3.8 GHz (500 MHz); n258, often called the 26 GHz band, runs from 24.25 to 27.5 GHz (3,250 MHz). The factor of 6.5 compares one band with another; it is not total spectrum and not a speed. Per carrier, the widest 5G channel is 100 MHz below 6 GHz and 400 MHz at millimetre waves.
The catch is blockage. Millimetre-wave signals suffer high attenuation and penetrate obstacles poorly, so moving vehicles, people and buildings easily block them. To travel far they need narrow beams, which must be aligned precisely and are easy to cut, as on slide 7. Deployments so far have mostly been limited to small hotspots rather than wide areas. The right picture is slide 4's road, closer.
Sources: 3GPP TS 38.101-1 · 3GPP TS 38.101-2 · Bai, Vaze and Heath, IEEE Transactions on Wireless Communications, 2014 · Jain et al., IEEE Journal on Selected Areas in Communications, 2019
If one beam is blocked, the other access points keep serving
A truck is unlikely to block every access point at once
Many access points turn blockage from failure into a scheduling decision
Script
So how do we make millimetre waves reliable? Don't tie the phone to one access point.
Point at the three beams: Here three access points serve the phone at the same time: two across the road, one on its side. Watch the speed bar. Here comes the truck. It cuts one beam, and the speed dips, but the link survives: the other two keep serving. Then it cuts the second one, and again the others carry on.
A moment ago, one access point and one truck meant no link at all. A truck is unlikely to block every access point at once. That's the cell-free idea from earlier, at millimetre waves.
[click] Many access points turn blockage from failure into a scheduling decision.
Background
The answer to slide 14, on the same road. Switching to another cell after a blockage is a handover, which cellular networks already do. In a cell-free network several access points serve the phone at the same moment, so a blocked beam costs a share of the speed, not the link. Two access points stand across the road, where their beams cross the truck's lane; the third stands on the phone's side, where the truck never crosses its beam. The beams are drawn as on slide 7.
The speed bar adds the beams' signals together and drops a beam while the truck cuts it; it shows the dip, not a measured rate. The second access point is placed so that one truck never cuts both of the first two beams at once. With more trucks or fewer access points every beam could be blocked, so this is a matter of likelihood, not a guarantee. The research questions are in the details: every beam has to stay aligned with a moving phone, and when one is blocked the network has to re-share power and beams among the access points still in view. That is a scheduling decision, the who-serves-whom decision of slides 11 and 12, with beams and blockage added.
Project summary: ARC DECRA DE260100816, Cell-Free Wide-Area mm-Wave Communications
Some access points receive while others send
So the network sends and receives at once, with ordinary half-duplex hardware
Half-duplex, today
All send, then all receive: time is split in two
red : to users · dark : from users
Full-duplex
Every access point does both, and must cancel its own signal
Network-assisted full-duplex
Each access point does one thing; the network does both
The network does both at once, if it chooses who sends and who receives
Mohammadi, Vu, Ngo, Matthaiou · Network-Assisted Full-Duplex Cell-Free Massive MIMO: Spectral and Energy Efficiencies · IEEE JSAC · 2023
Script
The second resource we share is time. Today, a network takes turns. Point at the left panel and its slot strip. Every access point sends to its users, then every one receives from them. Red is to the users, dark is from them. Time is split in two.
[click] Full-duplex would do both at once at every access point. But then each access point hears its own signal, far stronger than the faint one it's trying to receive, and needs special circuits to cancel it.
[click] Network-assisted full-duplex does something simpler. These two access points send while those two receive, at the same time, on the same frequencies. No access point does two things at once. The network does. The price is new interference, shown dashed: the sending access points reach the receiving ones, and the sending phones reach the other phones.
[click] The network does both at once, if it chooses who sends and who receives.
Background
Time is shared too. In half-duplex networks today, an access point sends and then receives, so time is split in two. Full-duplex lets every access point do both at once, but each must cancel its own transmission, which is far stronger than the signal it is trying to receive. Network-assisted full-duplex (NAFD) keeps every access point half-duplex: some send while others receive, at the same time and on the same frequencies, so the network as a whole is full-duplex.
The three panels use slide 8's park scene: four access points and four phones in two groups. Red is downlink (to the phones) and dark is uplink (from the phones); the arrow above each access point marks send or receive. The price of NAFD is cross-link interference, drawn dashed: from sending access points to receiving ones (red), and from sending phones to other phones (dark). The paper jointly optimises each access point's mode with power control, a mixed-integer nonconvex problem solved with SCA, the same kind of problem as slide 11. Because of the interference, the network does not get twice the capacity; slide 17 gives the simulated gains.
Source: Mohammadi, Vu, Ngo, Matthaiou, 2023
Choosing who sends and who receives saves time and energy
Each access point still only sends or receives, so it needs no full-duplex circuits
about 30%more speed
than half-duplex
up to 200%more energy efficiency
than half-duplex and full-duplex
My DECRA · a three-year ARC fellowship · three aims, at millimetre waves
Aim 1 · beam alignment
Point every beam precisely, as people move
Aim 2 · dynamic link management
Decide beams, and who sends and who receives, together
Aim 3 · machine learning, at scale
Scale Aims 1 and 2 up to a whole area
My DECRA brings these ideas to millimetre waves, to make them reliable across a wide area
Mohammadi, Vu, Ngo, Matthaiou · Network-Assisted Full-Duplex Cell-Free Massive MIMO: Spectral and Energy Efficiencies · IEEE JSAC · 2023 · ARC DECRA DE260100816
Script
So does choosing who sends and who receives help? In our paper's simulations, yes: about 30% more speed than half-duplex, and up to 200% more energy efficiency than both half- and full-duplex.
Then the three pictures: My DECRA, a three-year ARC fellowship, takes these ideas to millimetre waves, in three aims. First, point every beam precisely, even as people move. Second, decide the beams and who sends and who receives together: network-assisted full-duplex, now with beams. Third, machine learning to make all of this work at the scale of a whole area, not one street.
[click] My DECRA brings these ideas to millimetre waves, to make them reliable across a wide area.
Background
The numbers are from Mohammadi, Vu, Ngo and Matthaiou (IEEE JSAC, 2023), in simulation. The spectral-efficiency gain is against half-duplex and saturates at around 30% as access points are added, because of interference between access points (Fig. 2a). The energy-efficiency gain is up to 200%, against both half- and full-duplex (Fig. 3). Full-duplex does better when there are more than six users in each direction (Fig. 2b), and the comparison is generous to full-duplex, assuming its self-interference cancellation costs no power. The paper's conclusion names machine-learning-based resource allocation as a next step.
The DECRA project, a three-year ARC Discovery Early Career Researcher Award, Cell-Free Wide-Area mm-Wave Communications , takes these ideas to millimetre waves. As its public summary describes, it proposes a cell-free network architecture for outdoor, wide-area millimetre-wave service, with dynamic link management algorithms to overcome signal propagation and blockage, and planning tools for network operators. The three pictures illustrate this direction; they are drawn as what is new, not as earlier results.
Sources: Mohammadi, Vu, Ngo, Matthaiou, 2023 · ARC DECRA DE260100816
Open question: can the network see the blocker coming ?
Then it could move its beams before the link is cut
The open question
See the blocker, predict it, move first
Sense
the network's own radio signals, as they reflect off objects Predict
where the blocker will be Act
move beams, and who serves whom, before the cut, across many access points
If the network can see ahead, a blockage becomes a plan, not a surprise
Script
The same architecture opens up the next problem I'd like to work on with colleagues. Point at the teal echoes. What if the network could see the truck coming? Its own radio signals bounce off the truck, like radar. If it knows where the truck will be, it can move the beams before the link is cut. Watch the beam fade out before the truck arrives: no cut.
[click] Sense, predict, move first. For a single base station, people have used cameras and radar to see blockers coming. Across a cell-free network, with the network's own signals, deciding who serves whom before the cut: that's an open problem.
[click] If the network can see ahead, a blockage becomes a plan, not a surprise. If you build perception models, this is where I'd love your help. Next: training AI models over the network, without collecting the data.
Background
The open question of the millimetre-wave research takes slide 15 one step further: instead of reacting after a beam is cut, the network sees the blocker coming and moves first, so the picture never shows a cut. The two access points across the road send faint teal echoes that find the approaching truck (teal marks what devices sense, as on slide 2); the dashed teal outline shows where the truck will be in two seconds, and a beam fades out as soon as that outline reaches it, while the other access points keep the phone served.
The open problem is the combination: the network's own radio signals sense a blocker, the network predicts it, and it re-plans beams and who serves whom across a cell-free millimetre-wave network before the cut. Related work so far includes blockage prediction at a single base station with cameras (Charan, Alrabeiah and Alkhateeb, 2021) and with radar (Demirhan and Alkhateeb, 2022), and, in cell-free millimetre-wave networks, choosing for each access point whether to communicate or sense (Yan et al.) and using sensing to choose access points ahead of time for beamforming. This is the AI-for-networks direction of slide 5: learning to predict what the environment will do.
Related work: Charan et al., 2021 · Demirhan & Alkhateeb, 2022 · Yan et al. · arXiv 2510.06936
Your data stays on your device
Each device trains on its own data; only model updates travel
Collect the data
Every phone sends its data to a server, which trains the model
Federated learning
The model comes to the phones; only updates go back
Federated learning shares the model, not the data
Script
The third line of research is about training AI models across many devices. Here's the usual way. Point at the left panel. Every phone sends its data to a server, and the server trains the model. Your photos, your messages, all leave the phone.
[click] Federated learning turns that around. The model comes to the phones. Each phone learns on its own data, which never leaves it, and sends back only an update. The server combines the updates into a better model, and the next round begins.
Remember the assistant that knows you, without your data leaving your phone? This is how it can work: Google's Gboard keyboard is already trained this way.
[click] Federated learning shares the model, not the data.
Background
The third line of research starts from federated learning on ground networks, the foundation of the speaker's earlier work (slide 22), and heads for networks across space, air and ground (slide 23). The two panels use the park from slides 2, 6 and 7: three phones, the park's access point, and a server beside it, wired to the access point. On the left, each phone's data (the document icons) streams to the server. On the right, the data stays with each phone (the dotted circle, as on slide 2); the model comes down in red, each phone learns locally (teal, the deck's colour for what a device does on its own), and only the updates go back to be combined.
The raw data stays on the device, but an update can still reveal something about the data it came from, so practical systems add protections such as secure aggregation or differential privacy. The difference shown is what travels, data or model, not how much: for a large model an update can be bigger than the data. Google's Gboard trains its language models with federated learning on phones, the first example of slide 2. Privacy is the rule that slide 4 left aside because it is not a property of the radio channel; here it shapes how the network is used.
The closing leads to slide 20: the model is shared in every training round, so it crosses the air both ways, round after round.
Source: Google Research
Every training round crosses the air twice
And every step waits for the slowest device
T train = N rounds · ( max k t k DL download + max k t k comp learning + max k t k UL upload )
Training time depends on the network, not just the model
Vu, Ngo, Tran, Ngo, Dao, Middleton · Cell-Free Massive MIMO for Wireless Federated Learning · IEEE Transactions on Wireless Communications · 2020
Script
So the model is shared. But look at what one training round costs. As the playhead moves: First everyone downloads the model, over the air. Then everyone learns on their phone. Then everyone uploads their update, over the air again. Twice across the air, every round.
In the basic, synchronous form, each step waits for the slowest phone. This one has the weakest link, like the walker at the cell edge earlier: its download and upload take longest, and the other phones wait.
[click] So the training time is the number of rounds times the length of a round, and a round is the slowest download, plus the slowest learning, plus the slowest upload. And the number of rounds isn't fixed either: it depends on how much each phone learns per round. Hold that thought.
[click] Training time depends on the network, not just the model.
Background
One round of federated learning, as modelled in the TWC 2020 paper on cell-free massive MIMO for federated learning (Algorithm 1): the central processor sends the global model down to every device, each device computes its local update, each device sends its update up, and the updates are combined. Each step must be completed for all devices before the next one starts, so every step lasts as long as its slowest device, and the faster devices wait. The chart shows three devices, near the access point, a little further, and at the cell edge, with illustrative durations.
The formula simplifies the paper's Eqs. (27) and (28). Each max is taken over the devices k , so each term is the slowest device's time for that step. Each transfer time is the size of the update divided by the device's rate, the bits-divided-by-C of slide 6, so the network's choices of power and rate enter the training time directly. The number of rounds depends on how accurately each device solves its local problem, for a fixed target accuracy of the global model: a more accurate local step means fewer rounds but longer learning per round. The paper writes the number of rounds as G (θ ), with θ the local accuracy; the slide writes N rounds instead, because on slide 11 θ denotes power, in a different paper's notation. The paper also includes the time over the backhaul between access points and the central processor, left out here for space, and it treats the time to combine the updates as negligible.
This is why the device at the cell edge matters for AI (slide 9), and why cell-free networks suit federated learning: the paper notes they make the training less prone to users with unfavourable links (slide 10). Slide 21 turns the formula into a design problem.
Source: Vu, Ngo, Tran, Ngo, Dao, Middleton, 2020
We tune the learning and the network together
Learning more per round means fewer rounds, but longer ones
The optimisation problem
min ε , 𝐩 , 𝐑 , 𝐟 𝔼 { T train } average training time s.t. E k send ( p k , R k UL ) + E k learn ( ε , f k ) ≤ E max , ∀ k energy per phone R k DL ≤ r k DL ( 𝐩 ) , ∀ k download rate the channel allows R k UL ≤ r k UL ( 𝐩 ) , ∀ k upload rate the channel allows ∑ k p m k ≤ P max , ∀ m access-point power f min ≤ f k ≤ f max , ∀ k CPU speed
over ε : local accuracy · 𝐩: powers · 𝐑: data rates · 𝐟: CPU speeds
where r k UL ( 𝐩 ) = τ c − τ p τ c B log 2 ( 1 + SINR k UL ) upload rate of phone k SINR k UL = p k ( ∑ m σ m k 2 ) 2 ∑ l ≠ k p l ( ∑ m σ m k 2 β m l / β m k ) 2 | 𝝋 k H 𝝋 l | 2 shared pilots + ∑ l p l ∑ m σ m k 2 β m l every phone’s power + ∑ m σ m k 2 noise
Co-design beats tuning each side alone
Vu, Ngo, Tran, Ngo, Dao, Middleton · Cell-Free Massive MIMO for Wireless Federated Learning · IEEE Transactions on Wireless Communications · 2020
Script
Here's the thought I asked you to hold. Each phone can learn a little each round, or learn more. As the two rows play: Learning a little means many short rounds, and every round crosses the air twice. Learning more means fewer rounds, but each one takes longer and drains more battery. Both reach the same target accuracy. Which finishes first depends on the network.
[click] So we choose them together: how much each phone learns per round, the transmit powers, the data rates, and each phone's CPU speed, to minimise the training time.
[click] And each phone has one energy budget, for sending and for learning.
[click] And the rate the channel allows is the one formula I asked you to remember, with everyone in it. Every phone's upload power sits under every other phone's rate, and phones that send the same reference signal, called a pilot, leak into each other. That one rate sets both how long a phone takes to send, and how much battery it spends. Point at the denominator.
[click] Co-design beats tuning each side alone.
Background
The trade-off behind slide 20's number of rounds. In the TWC 2020 paper, how accurately each device solves its local problem in a round, its local accuracy, sets how many rounds are needed to reach a fixed target accuracy for the global model: more local work means fewer rounds, but each round's learning takes longer and costs more energy. The two rows show one phone taking the two routes on one clock, with illustrative numbers; here the slow link makes many rounds expensive, so learning more per round finishes first, but with a faster link the balance can tip the other way. The battery shows the energy spent.
The card simplifies the paper's problem (29). It chooses the local accuracy, the download and upload powers, the data rates and each device's CPU frequency together, to minimise the average training time over changing channel conditions. Each device's energy for sending and for learning shares one budget. Sending energy depends on the device's upload power and upload rate, and learning energy on the local accuracy and its CPU frequency (the paper's Eqs. (23) and (24)), so the network's variables and the learning's variables meet in one constraint. That is why the two sides cannot be tuned separately (the energy limit of slide 4); the rates are bounded by what the channel allows for the chosen powers, each access point's power is limited, and each CPU runs within its range. The paper normalises the powers, and the card shows them as plain limits. The paper writes the local accuracy as θ ; the card uses ε , since θ denotes power on slide 11, in a different paper's notation. It calls the problem mixed-timescale, stochastic and nonconvex, and solves it online with successive convex approximation, splitting it into fast decisions per channel condition and a slow decision on the local accuracy. Beamforming and user association are not variables in this paper.
The second click shows the upload rate, the paper's Eq. (21), which is slide 6's formula with the share of each coherence block spent on pilots removed. The card writes each device's upload power as p (the paper writes ρ u ζ k τ p , as on slide 11 (the paper writes τ t ). B is the bandwidth, τ c the coherence block length, σ²mk m 's channel estimate for device k , βmk φ k k 's pilot sequence. The denominator has three parts: pilot contamination, where devices whose pilots overlap corrupt each other's channel estimates; the interference from every device's upload power; and the noise. The download rate, the paper's Eq. (13), takes the same form with the access points' powers. The upload rate sets both the upload time, S u /R k UL , and the sending energy, so every device's power reaches every device's time and battery, and the problem cannot be split device by device.
The closing is tested in the paper: with 50 access points and 8 devices, jointly optimising everything reduces the training time by up to 55% compared with fixed settings and by up to 49% compared with tuning only the local accuracy, and it is also below tuning only the powers at every point of the paper's Figs. 5 and 6; slide 22 shows these results. The card is the "models and networks, designed together" of slide 8, in symbols.
Source: Vu, Ngo, Tran, Ngo, Dao, Middleton, 2020
Co-design cuts training time by up to 55%
Each side alone gets only part of the way
Designing the network around the learning makes training faster
Vu, Ngo, Tran, Ngo, Dao, Middleton · Cell-Free Massive MIMO for Wireless Federated Learning · IEEE Transactions on Wireless Communications · 2020 · Fig. 6
Script
Here's what that buys. In simulation: fifty access points, two to eight phones, and every scheme is set to reach the same target accuracy. The grey lines tune one side, or neither: fixed settings at the top, then tuning only the learning, and tuning only the network. Point at where the two lower grey lines cross: With few phones, tuning the learning helps more; with many, tuning the network does. Which side matters depends on the network.
[click] Tune both, and it's the fastest for every number of phones: up to 55% less training time than fixed settings, and up to 49% less than tuning the learning alone.
[click] That's the idea from the start of this talk, now with numbers: designing the network around the learning makes training faster.
Background
The chart redraws Fig. 6 of the TWC 2020 paper: the effective training time, averaged over large-scale fading realisations, with 50 access points and 2 to 8 devices, every scheme reaching the same target global accuracy, 10−2 in the paper. The values are the four plotted points of each line, taken from the paper's vector figure; the lines join them, with the paper's dash styles. Fixed powers and accuracy is the paper's baseline 1: equal download power for every device, maximum upload power, and a fixed local accuracy, with the data rates and processing frequencies still optimised. Tune the learning only is baseline 2, which also optimises the local accuracy; tune the network only is baseline 3, which also optimises the download and upload powers; tune both is the paper's Algorithm 4.
At 8 devices, tuning both reduces the training time by 55% compared with baseline 1 and by 49% compared with baseline 2, the reductions the paper reports. Training takes longer as devices are added, which the paper attributes to stronger interference and pilot contamination; tuning only the powers keeps that rise small, so the two single-side lines cross near 4 devices. That is the point of slide 21's closing: which side matters more depends on the network, and tuning both is lowest throughout.
These are simulations of a small network, computed from the paper's model of how many rounds training needs and how long each takes; the paper does not train on a dataset, and refers to the original paper of the learning framework it uses for experiments on real data. The closing returns to the idea of slide 5, designing the network around the AI task, with this result as the evidence. It holds for the setting simulated here: synchronous rounds, in which every step must be completed by all devices before the next begins (slide 20), with each round finishing within one large-scale coherence time, a period over which each link's average strength stays the same. Slide 23 turns to networks where links come and go.
The paper also compares network architectures under the same design: cell-free massive MIMO needs up to 57% less training time than collocated massive MIMO (30 access points, a 1 km × 1 km area), and up to 94% less than cell-free TDMA, where devices take turns to send (8 devices). This links back to the cell-free research: the cell-free network of slide 10 also suits federated learning.
Source: Vu, Ngo, Tran, Ngo, Dao, Middleton, 2020
Open question: how do we train across space, air and ground ?
A low-orbit satellite is in view for only a few minutes
Passes over one ground station · illustrative
Co-design the learning and the network around the passes
Matthiesen, Razmi, Leyva-Mayorga, Dekorsy, Popovski · Federated Learning in Satellite Constellations · IEEE Network · 2024
Script
This is where I'm heading, and it's still open. Up here, the links go through satellites, and a low satellite is in view for only a few minutes before the next one takes over.
[click] Say the satellites themselves learn, from the images they take, and a ground station combines their updates. Each one gets the model on one pass, learns while it's out of view, and can only send its update on a later pass. The passes ahead are outlined, because we already know them, from the orbits.
[click] So the idea is to co-design the learning and the network around the passes: which devices train, and when, and how they send. Unlike the truck on the road, a pass is known in advance. If you work on learning across satellites and drones, I'd love to work on this with you.
Background
A research direction, not a result: the speaker has not yet published on federated learning across satellite, aerial and ground networks, often called space-air-ground integrated networks (SAGIN). The subtitle comes from the survey by Matthiesen and colleagues (IEEE Network, 2024): contacts between a low-Earth-orbit satellite and a ground station typically last a few minutes, and because they follow from orbital mechanics, when each satellite can take part is predictable rather than random.
The timeline is illustrative and follows the survey's main scenario, in which satellites train on data they collect and a ground station acts as the server. A satellite receives the model during one pass, computes its update while out of view, and returns it on a later pass; the outlined passes ahead are known from the orbits. The survey shows how slow a synchronous algorithm such as FedAvg can be in this setting (in its example, the first round takes over nine hours), and reviews two responses: choosing which satellites take part in each round, and asynchronous algorithms such as FedSat, with scheduling that predicts when each satellite will be in view.
In the scene, the ship and the plane, like other users of large constellations, are served continuously as one satellite hands over to the next; each satellite is still in view for only minutes. The closing names the direction: deciding which devices train, when, and with which radio resources, around passes known in advance, carrying the co-design of slides 21 and 22 into the sky. Unlike the truck of slide 18, which the network has to see coming, a satellite's pass is known from its orbit.
Source: Matthiesen, Razmi, Leyva-Mayorga, Dekorsy, Popovski, IEEE Network, 2024
Our future still needs open problems solved
Problems I'd like to explore
An assistant that knows you, without your data leaving your phone
Networks for AI Running LLMs at the edge, and fine-tuning them on data that stays on the phone
Cars that see around corners by sharing what they see
AI for networks AI that lets networks sense their surroundings while they communicate
Drones that find people after a flood
AI for networks Semantic communication: send the meaning , not the video stream
A connection anywhere : the outback, the ocean, the sky
Networks for AI Federated learning across space, air and ground
A factory line that reconfigures itself
AI for networks AI agents that let the network reconfigure itself
Each needs wireless and AI researchers working together
Gündüz et al. · Beyond Transmitting Bits · IEEE JSAC · 2023 · IETF RFC 9315 · 3GPP TS 28.312
Script
Let's go back to where we started: five futures that are already being built. Each one still has open problems. Here's one for each, the ones I'd like to explore with colleagues here.
[click] The assistant: language models running at the edge, as in the open question on LLM requests, and fine-tuned on data that stays on the phone, which is federated learning again.
[click] Cars see around corners when the network itself can sense the road while it communicates, like seeing the blocker coming.
[click] Drones could send the meaning, "person on this roof", not the video stream. That's semantic communication.
[click] Once there's a connection anywhere, we want to learn anywhere: federated learning across space, air and ground, as with the satellite passes.
[click] And a factory line that reconfigures itself needs a network that reconfigures itself too: someone says "switch the line to the new product", and AI agents work out which machine gets which connection, and when.
[click] Each needs wireless and AI researchers working together.
Background
The talk closes by returning to the five futures of slide 2, each already in trials or early use, and naming one open problem behind each. They are problems the speaker would like to explore with colleagues, not results, and each future has many more. Each tile carries one of the two directions of slide 5: networks for AI, where the network serves an AI task, and AI for networks, where AI helps run the network. Three of the problems closed the three lines of research (slides 13, 18 and 23); two are new.
The assistant: answering requests with large language models at the network edge is slide 13's open question, and fine-tuning them on personal data without moving it is federated learning (slide 19). On its click, the tile's park gains an edge server beside its access point, where slide 19 placed its server: the phones send requests and the server answers, while each phone fine-tunes on data that stays with it. Cars that see around corners: a network that senses its surroundings while it communicates could detect what lies beyond a corner and share it, the direction of slide 18. Drones: semantic, or task-oriented, communication sends what the receiver needs rather than every bit; on its click, the tile's drone sends only that a person is on this roof, where on slide 2 it streamed its video, and help is still sent; Gündüz and colleagues survey the field (IEEE JSAC, 2023), and the idea goes back to Weaver's 1949 distinction between the technical and the semantic levels of communication. A connection anywhere: federated learning across satellite, aerial and ground networks (slide 23). The factory: a line that reconfigures itself needs a network that reconfigures itself too. Intent-based networking lets operators state what they want rather than how to configure it (IETF RFC 9315, and 3GPP TS 28.312 for mobile networks); AI agents that turn such goals into network actions, and keep adjusting them, are an open research direction.
The closing says why these are problems to work on with colleagues: each has a wireless side and an AI side, so each needs both kinds of expertise.
Sources: Gündüz et al., IEEE JSAC, 2023 · IETF RFC 9315 · 3GPP TS 28.312
Let's find the questions we share
Wireless networks that AI can rely on.
Cell-free massive MIMO (IoTJ 2024) · Network-assisted full-duplex (JSAC 2023) · ARC DECRA DE260100816 · Wireless federated learning (TWC 2020)
Script
If your model or your data will ever travel over a wireless link, that link is already deciding your training time, your latency and your battery. So one of those five problems is probably yours too. Tell me which one, now or in one line by email, and we'll write the question down together.
Questions, or better, problems: I'd love to hear them.
Background
The last slide stays on screen during the questions. The line under the title is the talk in one sentence, the network that slide 8 asks about and answers. The grey line lists the work shown, in the order of the three lines of research: cell-free massive MIMO (slides 9 to 12), network-assisted full-duplex and the DECRA project (slides 14 to 17), and federated learning over cell-free massive MIMO (slides 19 to 22). The code links to the speaker's homepage.
Thanh Tung Vu is a Lecturer in Computer Networking at Western Sydney University and holds an ARC Discovery Early Career Researcher Award (DE260100816, 2026-2029), Cell-Free Wide-Area mm-Wave Communications . After a PhD in electrical engineering at the University of Newcastle, Australia (2021), the speaker was a postdoctoral research fellow at Queen's University Belfast (2021-2022), Linköping University (2022-2023) and Macquarie University (2023-2025). Work at Linköping made the speaker a co-inventor on two patent applications filed by Ericsson AB.
Sources: thanhtungvu.com · ARC grant summary, DE260100816
