We use the latest machine-learning algorithms with efficient implementation in software and hardware. Our main scientific tool are multi-layered neural networks.
These network are composed of multiple layers of processing. Each layer performs: feature extraction with filter templates, reduction of input resolution, non-linear operations
The sequence of multiple layers of this neural network allows to compress images into short code-words that can then be used for recognition or for retrieval of targets of interest. We are experienced in training deep networks with the latest supervised and unsupervised algorithms. When data is plentiful, we can provide superlative results in precision by using supervised techniques. In case data is scarce or not existent we use unsupervised training, and even one-shot training. This is commonly performed while tracking novel objects and object parts, and also for identifying completely new objects never seen before.
We use microchip technologies to implement fast neural networks in custom hardware. Our implementation is several tens of times faster than conventional high-end laptop computers, and is as fast as many mobile graphic cards. Power usage is more than 10 times lower than commercial processors.
Want to find out more? Read our recent publications: