Qualcomm and Nokia Bell Labs showed how multiple-vendor AI fashions can work collectively in an interoperable method in wi-fi networks.
Carl Nuzman, Bell Labs Fellow at Nokia Bell Labs and Rachel Wang, principal engineer at Qualcomm, stated in a weblog put up that they demonstrated the pliability of sequential studying, which may facilitate community
decoder-first or machine encoder-first coaching.
They stated Qualcomm and Nokia Bell Labs are persevering with to work collectively to display the worth of interoperable, multi-vendor AI in wi-fi networks. At Cellular World Congress 2024, they first demonstrated over-the-air interoperability of AI-enhanced channel state suggestions encoder and decoder fashions.
These have been operating in reference cellular gadgets with Qualcomm’s 5G modem-RF system and a Nokia prototype base station, respectively. These interoperable fashions have been developed by the 2 corporations utilizing a brand new approach known as sequential studying. Now they’re again at MWC 2025 with extra.
Channel state suggestions helps the community work out one of the simplest ways to ship knowledge to your machine. As wi-fi situations change, so does the optimum route a transmission takes from the community to the machine. Qualcomm and Nokia have been capable of make the community smarter and extra environment friendly by producing exact beams with AI.
With sequential studying, a number of corporations can co-design interoperable AI fashions without having to share proprietary particulars of their implementations. As an alternative, a coaching dataset of mannequin enter/output pairs is shared from one firm to the opposite.
Constructing on this proof-of-concept, the businesses have since continued working collectively to display the worth, flexibility and scalability of interoperable AI for channel state suggestions.
Wi-fi AI robustness in several bodily environments
As AI applied sciences are deployed in real-world networks, you will need to make sure that fashions work robustly in numerous environments. Coaching datasets needs to be sufficiently numerous for AI fashions to be taught successfully; nonetheless, it’s unrealistic for them to cowl all doable situations.
Thus, it’s crucial for AI fashions to generalize their coaching to deal with new conditions. Within the collaboration, the corporations studied three very completely different cell websites: an outside suburban location and two completely different indoor environments.
Within the first situation, they in contrast the efficiency of a typical AI mannequin educated with numerous datasets with hyper-local fashions which can be educated at particular places. They discovered the widespread AI mannequin can work in several environments with comparable efficiency as hyper native fashions.
The businesses later tailored the widespread mannequin to incorporate knowledge from Indoor Website 2 (the Tailored Widespread mannequin). Then they measured the consumer knowledge throughput at 4 completely different places inside Indoor Website 2. The widespread mannequin got here inside 1% of the efficiency of the Tailored Widespread mannequin in all circumstances, exhibiting the robustness of the overall widespread mannequin to new situations.
AI-enhanced channel state suggestions permits the community to transmit in a extra exact beam sample, bettering the acquired sign energy, lowering interference, and in the end offering larger knowledge throughput. We measured this enchancment by logging knowledge throughputs skilled with AI-based suggestions and grid-of-beam-based suggestions (3GPP Kind I) because the cellular consumer moved between varied places within the cell.
Use of the AI suggestions yielded larger throughput, with per-location throughput positive factors starting from 15% to 95%. The throughput positive factors that will likely be noticed in industrial techniques below AI-enhanced CSF will rely on many elements. Nonetheless, the outcomes of this proof-of-concept, along with quite a few simulation research, counsel that the throughput with AI enhancements will likely be constantly larger than the that achieved with legacy approaches.
Sequential studying may be carried out in two methods, both machine encoder-first or community decoder first, which has completely different implications for deployment and standardization. To assist 3GPP’s growing curiosity within the decoder-first method, this yr we changed our unique encoder-first demonstrations with decoder-first mannequin coaching.
With the encoder-first method demonstrated in MWC 2024, Qualcomm designed an encoder mannequin, generated a coaching dataset of enter/output pairs, after which shared the dataset with Nokia, which subsequently designed an interoperable decoder.
This yr, with the decoder-first method, Nokia designed a decoder mannequin and generated and shared a coaching dataset of decoder enter/output pairs for Qualcomm Applied sciences to make use of in designing an interoperable encoder. We discovered that fashions designed by each modalities carried out equally nicely, inside just a few share factors.
Backside line
The prototype that Qualcomm Applied sciences and Nokia Bell Labs have collectively demonstrated represents a key step in shifting AI-enhanced communication from idea to actuality. The outcomes present that the consumer expertise may be considerably improved, in a strong method, through a number of studying modalities. As we be taught to design interoperable, multi-vendor AI techniques, we will begin to notice enhanced capability, improved reliability, and diminished vitality consumption.
