We developed a natural language text classification technology based on deep neural networks.
The technology supports all languages: for different languages we construct and train specific models. We use both convolutional and recurrent (LSTM and BiRNN) neural networks.
OBJECT DETECTORS (IMAGES AND VIDEOS)
We develop algorithms for object detection and classification of images and videos. Detectors work at a speed of about 100 frames per second for 640x480 frames on NVIDIA GTX 1080. Precision and recall vary from task to task and ___ depend on required speed.
It can be used for tasks of satellite images segmentation or analysis, advanced videoanalytics, autonomous cars vision and many other applications.
PUZZLELIB DEEP LEARNING FRAMEWORK
Deep Learning library PuzzleLib is used for construction, training and usage of deep neural networks. The framework supports all modern architectures of neural networks and uses.
PuzzleLib supports most popular operation systems: Windows, Linux, Mac OS, iOS, Android.
MOBILE DEEP LEARNING
Our Deep Learning framework PuzzleLib supports Apple devices (iPhone 5s and newer). Both forward computations (object recognition) and backward computations (training the neural network directly on the device) are supported. Full Android support is coming soon.
PuzzleLib Mobile can be used for AR apps and photo/video recognition apps.
DEEP LEARNING ON NVIDIA JETSON
Our framework can work on NVIDIA Jetson TX2. It provides an opportunity to use neural networks in robotics, autonomous cars and drones.
Deep neural networks built with PuzzleLib can process from 20 to 100 frames per second on Jetson.
Jetson can be used in tasks where on-board data processing is necessary: when the speed is important and there is no time to transfer data to the server.
DEEP NEURAL NETWORKS ON BAIKAL CPU
Our algorithms can work on Baikal computers. Baikal is certified for military applications and use in government agencies of Russia.