Conferences and supporting programme
Effective Embedded AI with NeuroMem Technology: From IP to Applications
Embedded AI, especially hardware, has a high potential to become pervasive. NeuroMem technology is a parallel architecture with dedicated memory that is able to learn patterns such as image, sound or sensor data, and recognize news patterns in few microseconds regardless of the number of patterns. Its cognitive engine is agonistic to data type and can recognize by association, either fuzzy or exact, with real-time and providing a recognition quality factor. Hence, supporting Radial Basis Function (RBF), K-Nearest Neighbor (KNN) classification methods. The learning is adaptive, incremental and can be done during all the life of the product. Moreover, the knowledge builder is on-the-go, portable and encompasses traceability. Some of key outcome features are ubiquitous technology, high-performance and low-power hardware thanks to its efficient architecture, since memory and processing logic are in a same cell, natively parallel architecture, with low and deterministic latency, and easy to integrate into designs through standard interfaces. Indeed, similarly to neurons, NeuroMem is combining data, learnt data, learning and recognizing within the same hardware architecture avoiding processors bottlenecks from data, memory and processing. However, the most effective usage of this technology is in synergy with hardware architecture such as Harvard and GPU, for pattern recognition allowing lower hardware complexity, thus, lower power consumption and lower system cost. The comparison, role and possible synergies with other architecture will be presented. This technology has been implemented by other companies in standalone chips (500 and 1000 neurons such as NM500) and system-on-chip. This is allowing to develop new applications with minimum software and shorter time. Some key performance of these implementations will be presented. Furthermore, development of applications is becoming easier since built-in learning engine, life-long learning and transferable learning. This incremental and adaptive development are also closer to innovative and user driven design methods where Agile or Rational Unified Process (RUP) are necessary but difficult to use for applications using hardware. After explaining NeuroMem technology, showing some implementation and application examples, we’ll explore some of new fields that can be addressed efficiently such as medical monitoring, gate and activity analyzing, industrial process and machines, and vision.
--- Date: 27.02.2018 Time: 10:30 AM - 11:00 AM Location: Conference Counter NCC Ost