Deploying NeuroMem-powered intelligence

Trainable, Responsible, Explainable

Deploying a NeuroMem engine

Load

Collect your data, visualize and curate, annotate with your labels.

Learn

Train the neurons with features extracted from your annotations.

Validate

Verify the recognition of new datasets, identify cases of uncertainty.

Deploy

Export the knowledge of the neurons to deployment platforms.

NeuroMem hardware platform

A typical NeuroMem-powered platform is composed of a processing unit (MCU, FPGA or both) interfaced to a bank of NeuroMem chips through the NeuroMem parallel bus or an intermediary serial communication controller. Depending on the targeted application, GPIOs and communication ports are connected to the MCU or FPGA. A Flash memory or SD card are convenient to store locally the MCU/FPGA configuration files and the NeuroMem pre-trained knowledge files if any.

The NeuroMem parallel architecture enables a seamless scalability of the network by cascading chips.

NeuroMem software tools

NeuroMem Knowledge Builder are a simple framework teach the neurons and validate their knowledge on your training and testing datasets. Detailed reports guide you to achieve the best compromise between the accuracy and throughput for your application. They pinpoint categories which are not properly modeled or categories subject to confusion. They guide you through the use of consolidated rules within a single network or across multiple networks trained for the same purpose on different features.

Learning criteria

  • Data reduction
  • Rate of convergence
  • Categories subject to confusion

Classification criteria

  • RBF versus K-NN classifications
  • Consolidation rules in case of uncertainties
  • Qualify accuracy per category and vector

NeuroMem I/Os

Harwdare driver to access the neurons through their parallel bus or other communication bus such as I2C and SPI.

CogniPat library

Library to manipulate vector data and broadcast them to the neurons for learning or recognition. Data can derive from any data types including  text, packets, measurements, images and waveforms.

CogniSight library

Library to manipulate pixel data, extract feature vectors from regions and broadcast them to the neurons for learning or recognition. Pixel data can come from image files or image frames from a live or stored video streams.

Integrated Software Toolchain

Training services

  • Reduce development time and costs through faster learning and targeted case studies
  • Get full insight on how to deploy NeuroMem neural networks for your applications from concept to implementation

Standard Course

  • Academic presentations, demonstrations
  • Hands-on exercises
  • 1 to 2 days off-site or on-site
  • Download information

Technology Assessment Program

  • Standard training
  • Site license for all software used during session
  • 20 hours of consulting to 2 points of contact
  • Download information