Abstract:
A Binary Neural Network (BNN) organizes neurons using logic functions and bitsets for
their activation and connections. Typical BNN research focuses on the discretization
of
oating-point gradients with xed models. This work followed a di erent approach:
the framework Binary Spectrum-diverse Uni ed Neuroevolution Architecture (BiSUNA)
employs population metaheuristics to nd dynamic neural network models composed
of binary neurons and discrete weights. These adaptive arrangements provide multiple
solutions to Reinforcement Learning (RL) problems.
In other words, this research produced a state-of-the-art technique to train evolutionary
Binary Neural Networks (eBNN). It enables discrete models to resolve discrete
RL problems, a perspective previously not explored in the literature.
Several results show how BiSUNA nds solutions to standardized OpenAI Gym and
Atari 8-bit video games. It also dispatches competitive rewards compared to other Deep
RL implementations. It solves Copy, DuplicatedInput or the SpaceInvaders video game
with only 1,000 generations. BiSUNA is an addition to the Deep Learning tools that
allow researchers to solve problems where discrete values are essential.
Employ Look Up Tables (LUT) instead of Digital Signal Processor (DSP) decreases
circuitry complexity for de ned applications. Because of this, the BiSUNA framework
shrinks power consumption when it executes on recon gurable hardware compared to
mainstream computational components. This work can train binary agents by taking
advantage of FPGA high-level synthesis tools using only logic functions. The change to
binary sets allows eBNN populations to use fewer resources and execute faster than similar
continuous neurons, up to 60% with comparable
oating-point agents on recon gurable
hardware. This project can deploy workloads on reprogrammable hardware to train
neural networks as the rst step in a long road to diversify the CPU/GPU duopoly.
Another contribution is the development of an unhackneyed adversarial environment
with three di erent populations. BiSUNA agents create a communication system with
confusion and di usion properties comparable to the human-made Advanced Encryption
Standard (AES). By using the Normalized Cross-Correlation, eBNN score up to 86%
better than AES.