Sony Patents an AI System That Requests More Data Only When It Needs To
Instead of waiting for a full dataset before making a decision, Sony's patent describes an AI that pulls in data piece by piece and only asks for more if it isn't confident enough yet.
How Sony's on-demand AI data feed works
Today, most AI systems need all their input data delivered upfront before they can start working. That means if you're on a slow or spotty connection, everything stalls until the full payload arrives.
Sony's patent describes a different approach: the AI receives data in small chunks, makes a running guess after each one, and only asks the sender for another chunk if its current answer still isn't good enough. Once it's confident, it stops asking and delivers the result. Think of it like a doctor who orders tests one at a time instead of running every test at once, stopping when the diagnosis is clear.
The practical payoff is that a lot of tasks could finish using far less data than they would today. Your device (or a server) sends as little as possible, and the AI on the other end signals when it has enough.
… a generator that generates, from the original input data, a plurality of pieces of input data with reduced data amounts; a data transmitter that sequentially transmits pieces of input data of the plurality of pieces of input data; and a request receiver that receives feedback information based on whether there is a need for an additional piece of input data …
Translation: It shrinks data into smaller chunks, sends them one by one, and listens for requests for more information.
How the feedback loop gates each data chunk
The patent describes a two-sided system: a transmission entity (the sender) and a reception entity (the receiver, which runs the AI model).
On the sending side, a generator takes the original raw data and produces multiple smaller versions of it, each with a reduced data footprint. These are queued up and sent one at a time rather than all at once.
On the receiving side, a multi-input AI/ML model (an AI designed to work with a sequence of partial inputs, not just one complete one) ingests each chunk as it arrives and immediately produces an inference value (a confidence-weighted answer). The system updates its output after every chunk. A dedicated determination section then evaluates whether the latest answer is good enough. If it is, the process stops. If not, it sends feedback information back to the sender requesting the next chunk.
Key components include:
- Generator on the sender that creates the ranked partial-data sequence
- Inference section on the receiver that scores each incoming chunk
- Determination section that decides whether confidence is sufficient
- Request transmitter that fires the feedback signal back to the sender
… an inference section that sequentially inputs the pieces of input data of the plurality of pieces of input data to a multi-input AI/ML model to sequentially acquire inference values of a plurality of inference values; an output section that sequentially updates inference results on the basis of the respective inference values …
Translation: The system feeds the data pieces into the AI model one at a time to constantly update its conclusions.
What this means for AI running on tight networks
For AI running over constrained links, such as mobile networks, satellite connections, or edge devices with limited bandwidth, sending full datasets every time is expensive and slow. A system that short-circuits the data stream the moment it has enough information could cut transmission costs and latency in a meaningful way, especially for real-time applications like video analysis or sensor processing.
Sony's steady investment in networked AI processing suggests this fits into a broader push to make AI inference work well beyond the data center. Whether the gains are large or modest depends entirely on how predictable the data requirements turn out to be in practice, and that is something only real-world deployment would reveal.
This is the tenth Sony filing we've tracked in AI training and infrastructure since June, a topic that already includes an AI gaming coach and training on private approximate data.
The core idea here is appealingly simple: stop sending data when you don't need it anymore. The concept is sound, and in bandwidth-constrained scenarios it could genuinely reduce waste.
The path to a shipping product, though, requires several pieces to be in place at once. You need a sender capable of generating the ranked partial-data sequence, a receiver running a model specifically designed to accept those partial inputs, and a low-latency feedback channel between them. That's three coordinated components, not one drop-in software update. The feedback loop itself adds a round-trip delay that could eat into the savings on very fast networks.
This reads as foundational infrastructure work, the kind of patent that defines a mechanism rather than a product. It's the right kind of thinking for constrained AI deployment, but it would take significant system-level engineering before anything like this shows up in a consumer device or a standardized protocol.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
25 drawing sheets from US 2026/0270156 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →