Luke Minks | TACOCAT

Motherboard

Sun, 19 Apr 2020 15:50:34 -0400

The TACOCAT motherboard design includes all of the network’s hardware-based artificial neurons along with the synapses connecting the hidden-layer neurons to the output-layer neurons.
The motherboard also includes a 32-bit ARM microcontroller that measures neuron voltages and drives the 100 pixel input lines that connect to each daughterboard.
The motherboard is connected to a host PC via I2C datalink.

Circuit Design Leads: Deven and German

PCB Design Lead: Luke

Sun, 19 Apr 2020 14:58:51 -0400

Each daughterboard implements the 100 hardware synapses that are required for each hidden-layer neuron.
The daughterboard performs 100 multiplication operations simultaneuosly, and the sum of the products is calculated in logarithmic time complexity using an analog adder tree.
TACOCAT uses three daughterboard PCBs.

Circuit Design Leads: Deven and German

PCB Design Lead: Luke

Sun, 19 Apr 2020 13:59:06 -0400

The TACOCAT software is written in Python 3.7. It contains modules for loading image data, training a neural network (software- or hardware-based), simulating a hardware-based neural network with SPICE, and connecting to the TACOCAT neural network’s microcontroller. Design Lead: Justin

Sun, 19 Apr 2020 11:59:10 -0400

Firmware for the STM32H743 microcontroller was written in C.
The microcontroller acts as an interface between the software-based training algorithm and the neural network hardware, mediating the exchange of weight values, input data, and neuron outputs between the two entities.

Design Lead: Justin

Sat, 18 Apr 2020 22:59:11 -0400

Fri, 17 Apr 2020 00:00:00 +0000

Tue, 14 Apr 2020 00:00:00 +0000