In today’s landscape of Artificial Intelligence, Deep Learning and its numerous inference engines are monopolizing the front stage, but other technologies have essential benefits such as field training and real-time adaptivity, novelty detection, learning causality and traceability.
Among them, the NeuroMem NM500 chip is a digital neural network chip capable of intrinsic learning and recognition of patterns derived from multimedia such as images and sounds, but also instruments, text and other data types. Manufactured by nepes (Korea) under a license from General Vision, the NM500 features 576 neurons which can be trained on small datasets and cleverly tuned to deliver the best compromise between throughput and accuracy for a given application. For example, one may prefer to train a NeuroMem network acknowledging when it is uncertain or even ignorant rather than guessing or reporting a “closest” match which can still be quite far. This is made possible when the neurons are used as a Radial Basis Function classifier, and not as the commonly known K-Nearest Neighbor. It is this notion of ignorance and uncertainty which can trigger the intelligent decision to learn more or to have the wise recourse to another opinion. By combining multiple NeuroMem networks (or experts) trained differently on the same subject, accurate decisions can be made taking advantage of their complementarity or their domains of mutual exclusivity.
To experiment with NeuroMem networks, General Vision’s NeuroMem Knowledge Builder is a simple framework to train and test the neurons on your datasets while producing rich diagnostics. The company also offers simple APIs and tools for generic pattern recognition as well as image recognition. They all integrate a cycle accurate simulation of 8000 neurons and can also interface to a hardware NeuroMem network such as the Brilliant USB dongle (2304 neurons), the Arduino-compatible NeuroShield board (576 neurons) with expandable network capacity, and soon a cognitive SSD with high speed throughput and high network capacity. In addition to the NM500 chip, NeuroMem is available as an IP core for FPGAs from Xilinx, Altera and Lattice and also for licensing.
NeuroMem Knowledge Builder (NMKB) is a simple framework to experiment with the power and adaptivity of the NeuroMem RBF classifier. It lets you train and test the neurons on your datasets while producing rich diagnostics reports to help you find the best compromise between accuracy and throughput, waive uncertainties and reduce confusion. The application runs under Windows and integrates a cycle accurate simulation of 8000 neurons, but can also interface to a hardware NeuroMem network such as the Brilliant USB dongle (2304 neurons), the Arduino-compatible NeuroShield board (576 neurons) with expandable network capacity, and soon a cognitive SSD with high speed throughput and high network capacity.
Simple toolchain for non AI experts
NMKB can import labeled datasets deriving from any data types such as text, heterogeneous measurements, images, video and audio files. The neurons can build a knowledge in a few seconds and diagnostics reports indicate if the training dataset was significant and sufficient, how many models were retained by the neurons to describe each category, and more. For the classification of new datasets, you can choose to use the neurons as a Radial Basis Function or K-Nearest Neighbor classifier. Other settings include the value K and a consolidation rule to produce a single output in case of uncertainties. The throughput and accuracy of the classification are reported per categories.
Building a knowledge with traceability
The application produces a data log to easily track and compare the settings and workflow which were tested on a given dataset. The traceability of the knowledge built by the NeuroMem neurons is conveniently exploited by NMKB. For example, you can filter the vectors classified incorrectly and comprehend why by comparing their profile to the firing models. This utility may even pinpoint errors in the input datasets!
Primitive and custom knowledge bases
Finally, the knowledge built by the neurons can be saved and exported to other NeuroMem platforms which can themselves use the knowledge “as is” or possibly enrich it if they are configured with a learning logic. A typical NeuroMem platform features a Field Programmable Gate Array (FPGA) and a bank of NM500 chips interconnected together either directly or through the FPGA along with the necessary GPIOs and communication ports for the targeted application. They all have in common that latencies to learn and recognize are deterministic and independent of the complexity of your datasets. The network’s parallel architecture also enables a seamless scalability of the network by cascading chips.
In addition to the NeuroMem Knowledge Builder, General Vision offers SDKs interfacing to NeuroMem networks for generic pattern recognition and image recognition with examples in C/C++, C#, Python, MatLab and LabVIEW.
What are the key differences between Deep Learning and RBF-based solutions? Local vs Remote learning, Additive learning vs Forget-and-Learn-from-scratch learning, Predictable recognition latency, Knowledge Traceability…
Philippe Lambinet, president of Cogito Instruments, explains why the NeuroMem AI chips featuring RBF classifiers and integrated in Cogito’s products help solve industrail IOT applications.
SINGAPORE, October 17, 2018 – Longway A.I. Technologies Pte Ltd (Singapore) and General Vision Inc®, one of the top 10 Artificial Intelligence companies in the world1, have signed a research collaboration agreement to develop Cognitive SSD, a “2.5” SSD with a configurable neuromorphic search engine that enables energy efficiency and parallel high-speed sorting through big data using Artificial Intelligence (A.I.).
General Vision is cited again among the leaders of the self-Learning Neuromorphic Chip Market 2018-2023.