Machine Learning (ML), a subset of AI technology, helps use historic data and generate intelligent response. To achieve this, a DNN is designed and trained using historic data. This network generates intelligence response for subsequent fresh input sequence. Earlier this was done in cloud based AI software running on powerful servers. With availability of multiple cpu cores within chip, a dedicated hardware logic implementing DNNs and yet consume low power , it is now possible to implement ML on embedded devices. Advantage of this is now AI-ML algorithm can be implemented right in the control system, edge device interfacing IoT sensor.
Applications of embedded ML are numerous. User can be provided improved personal assistance on smartphones, wearables. It can be used to provide improved drive experience in automobile, Improved production efficiency & predictive maintenance in Industrial production setup.
Blueberry has capability to model DNN, deploy on embedded device having multiple CPU cores or dedicated compute hardware. Blueberry has done projects in designing compiler tool chain for application specific hardware for DNN. The flexible compiler developed, compiles DNN models developed in any of the popular DNN modelling frameworks. The framework, methodology/technology used are ONNX, MLIR, LLVM.