Artificial Intelligence is poised to unleash the next wave of digital disruption. AI is being packaged in small devices, chips, sensors and agents for intelligent decision making. There is a trend of moving the AI inference engine from server/cloud/high-performance-computing towards the edge (i.e. PCs, IoT devices, etc.). Having devices perform machine learning locally versus relying solely on the cloud is a trend driven by the need to reduce latency, to ensure persistent availability, to reduce costs, and to address privacy concerns. Both software and hardware approaches are feeding into the edge-based processing for AI.
The Eximius team, with its expertise in ASIC, FPGA, Storage, Connectivity, Software and Machine Learning has put together End-to-End solutions and provided services towards Artificial Intelligence in multiple domains.
We have built and deployed Deep Learning models with Text, Images and Speech data within AI applications. Our experience with ASIC & FPGA allows us to develop and validate energy efficient, real time inferencing engines. Combining our IoT and Connectivity skills, we help enterprises in an efficient conversion of their Concept to Models and Products.
At Eximius, we have applied Deep Learning for Natural Language Processing, Computer Vision and Speech Recognition. We have trained Fully Connected, Convolutional (CNN) and Sequential Neural Network (RNN, LSTM) models using multiple Deep Learning Frameworks (Tensorflow, Torch, CNTK, Caffe2, Theano, Keras). We have also programmed Hardware Accelerators (GPU, FPGA) with CUDA and OpenCL to optimize the performance of learning algorithms and trained models.
The Eximius team has experience in the development of an Object detection and classification solution for Advanced driver-assistance systems (ADAS) – specifically for Pedestrian Detection. Here, firstly, for Object detection, a rich set of object proposals (i.e., a set of image regions which are likely to contain an object) was generated using a fast (but possibly inaccurate) algorithm. Secondly, a convolutional neural network (CNN) classifier is applied on each of the proposals. The customer specific solution was developed and optimized on an FPGA based system.
Natural Language Processing
Eximius team has designed multiple Natural Language Processing pipelines using Deep Learning approaches. The team has experience working on Sentiment Analysis and Intent Classification. Recurrent Neural Networks (RNN) with Bi-directional Long Short-Term Memory (LSTM) units were used for Sentiment Classification. The trained model was then optimized for performance on an FPGA platform. Eximius team has also worked on Intent Classification from text. It has trained LSTM based models for joint Intent Detection and Slot Filling.
Eximius team is working on an android application that performs on-device inferencing to understand voice commands. Embedded processors and SDKs such as Qualcomm® Snapdragon™ Neural Processing Engine (NPE) software development kit (SDK) are being used for solutions that place an increasing demand for running deep neural networks efficiently on mobile and other edge devices.
"The most technologically efficient machine that man has ever invented is the book."
– Northrop Frye