CORTEZ1

The CORTEZ design implements a simple Neural Network trained to recognize 3 vowels (o, u and i) represented as a 3x3 matrix. The CORTEZ1 chip is based on the caravel SoC and it targets GF180MCU/D technology.

Network architecture

The CORTEZ Neural Network has the following features:

• Bipolar input values, inspired by the Madaline network;
• Training is based on the back-propagation algorithm;
• The activation function is implemented by means of a piecewise linear approximation of the tanh(x) function;
• Fixed-point primitives rather than floating-point, to save area and reduce costs;
• All values (inputs, intermediate values, weights and biases) are 8-bit fixed point, with 5 bits for the fractional part;
• The input consists of 9 values;
• The hidden layer consists of 6 neurons. Inputs are fully connected to the input layer;
• The output layer is made of 3 neurons. Outputs are fully connected to the hidden layer.