Over-the-air compute (OAC) is the idea that, instead of transmitting data from multiple senders to a receiver, who then computes some function over that data, the senders code their data so that the combination of the codes in the receiver produces the desired result of that function directly as a result of the radio signal/codes decoding.
One obvious advantage of this is that instead of contending for a shared radio channel (and avoiding collision via various time, space or other multiplexing/control or contention avoiding protocols), the senders send simultaneously, so the result is much quicker and the radio channel use therefore also much more efficient, with possible energy saving at senders and recxeivers (not having to stay powered on for so long, at least). This is clearly useful in IoT/Smart Environment settings with very low power devices/sensors.
The thesis linked above devises various codes for various scenarios (including dealing with noise o nthe channel) and differs from earlier OAC work which was in the analog domain by being digital. Various functions (mean, min/max, average) are shown to be feasible and efficient (thoguh not that cheap to design, but once designed, not expensive to code).
Two possible extensions to this: one would be to add feedback so that recursive functions could be applied (or even just adaptation). The second is to look at how the coding is designed and see if the process is amenable to some type of compiler technology (take a high level definition of the desired function, and see if we can build an automated way to generate the constellation/coding). One cool application would be for super-efficient Federated Learning.
Ethernet for AI includes links to the htsim platform. I'm suggesting using that to explore this idea for using a field based model for managing traffic flows where you have tools like packet spraying and ultra-ethernet potential congestion signaling mechanism to provide for latching of state in switches bi-dirctionally.
Could make use of Ryan and Josh spatial platform too...
In practice, this could look like an extension of Congestion Managers, Co-Flow Scheduling and HCSFQ
Verse is a new programming language for games
Metaphors is a proposal for a mechanism for handling variable perforance between players in multiplayer games/virtual environments.
This project would be to implement that latter in the former, and build a demo app.
Review papers are extremely useful and popular publications which serve the community by checkpointing all the work across an area in a structured way so that people can see what has been done, and also where there are gaps in current research. They are often very highly cited.
This project would be to take a corpus of review papers and build a machine learning environment that finds what features are "most useful" (in the sense that they lead to most citations around the time of publication or later). The idea would be to take apply some fairly standard NLP techniques to find properties of a review paper, and then look at how thse impact the paper's effectiveness.
Of course, an obvious goal for this project is to produce a "review of review papers" paper, for publication!
For an example of biblioemtrics, see this earlier work we did on networking papers over 50 years
This project is about using a trusted execution environment (in this case, Cheri or ARM Morello) to protect code&data running a confidential computation, e.g. learning some statistics from personal healthcare or finanical records.
The idea starts from work on secure containers and confidential map/reduce style computational frameworks (Hadoop, Spark etc), but in this instance, the task is to take a different framework, namely dataflow, built in a different language (not java/scala, but Rust).
Timely Dataflow and see also differential dataflow . I believe there's a rust compiler that can use the ARM capability mechanism now or soon.
An extra challenge might be to figure out how to do attestation e.g. using a secure co-processor, even a SIM, even something like this.
last edited, 3 Oct 2025