Developing AI to build the next generation of wireless communications requires advanced new computers capable of simulating the electromagnetic spectrum around us. DeepSig has designed and built our in-house training cluster for AI-native wireless communications to further advance the AI research and development of our software products. The training cluster, or supercomputer, is used for AutoML and part of our MLOps process to ensure our customers are getting finely tuned models that generalize across an extensive range of RF conditions and datasets.
What does this mean for our customers?
AutoML, or automated machine learning, is the process of automating adjustments to model architectures and training parameters to maximize accuracy, speed, and generalization of a machine-learning based product. DeepSig uses state-of-the-art AutoML practices to tune AI parameters and options to maximize model performance and generalization. This produces faster and more accurate models for our products, OmniSIG and OmniPHY-5G. MLOps, or machine learning operations at DeepSig ensures that our products, which are leveraging first-of-its kind AI capabilities, do not regress in performance as our engineers and data teams continue pushing the AI-native wireless communications boundaries.
These new cluster capabilities increase our compute power by 20x, enabling AutoML to search through dozens to hundreds of training models per day to find the optimal AI solutions for our customers.
By finding models that can fit within the low-latency time-sensitive requirements of 3GPP processing while improving wireless performance, the cluster accelerates our path to continue to create AI 5G-RAN improvements. Our AI advancements in the 5G-RAN also improve wireless performance in the harsh wireless conditions that cellular networks operate in, including high-multipath factories, urban canyons, and the cell-edge. Through AutoML we automatically improve our AI models and training routines by leveraging intelligence-guided searches to fine-tune models. This supercomputer capability allows DeepSig to train and replay candidates in this model search process in a simulated environment, such as pedestrians walking through urban canyons or high-speed rail up to 8 times faster than a traditional wireless channel emulator, allowing for more evaluation of channel and interference conditions. This results in a significantly increased RAN throughput when deploying to partner O-RAN networks.
The cluster advances our efforts to reduce dataset burdens for wideband spectrum awareness and recognition through a combination of faster AutoML, MLOps, and expanded research capacity. This will enable our pioneering application of self-supervised learning to wideband spectrum awareness. In addition to on-going OmniSIG R&D, DeepSig is bringing self-supervised learning to phased array data, which means reducing dataset burden for future spectrum awareness and wideband digital array signal processing. This advanced compute capacity brings the large-scale search requirements to a reality for all of our OmniSIG customers.