Towards the Neuroevolution of Low-level Artificial General Intelligence

Sidney Pontes-Filho, Kristoffer Olsen, Anis Yazidi, Michael A. Riegler, Pål Halvorsen, Stefano Nichele

Abstract

This paper proposes a bottom-up approach to Artificial General Intelligence (AGI) by focusing on "low-level" intelligence modeled after primitive biological organisms. Rather than optimizing static weights for specific tasks, the researchers introduce the NAGI (Neuroevolution of Artificial General Intelligence) framework to evolve Spiking Neural Networks (SNNs) that possess innate learning machinery. By evolving local plasticity rules and network topologies, agents can adapt to mutable environments in real-time starting from randomized initial weights. Experimental results across foraging, logic emulation, and control tasks demonstrate the robustness of evolved plasticity in achieving adaptive behavior.

Introduction

Modern AI excels at narrow tasks through global optimization on static datasets but often fails to replicate the general adaptability of simple biological systems, such as the C. elegans worm. Traditional deep learning relies on fixed weights determined during training, making "on-the-fly" adaptation to environmental shifts difficult. This research argues for a paradigm shift: building AGI from the bottom up by evolving systems that can reorganize themselves based on sensory feedback like pleasure or pain.

The NAGI Framework

The core of this approach is the NAGI framework, which utilizes an extension of the NEAT (NeuroEvolution of Augmenting Topologies) algorithm. Unlike standard NEAT, which evolves connection weights, NAGI evolves the parameters of learning. This includes the network's topology, neurotransmitter types (excitatory vs. inhibitory), and specific local learning rules.

The NAGI framework evolves genotypes consisting of node and connection genes, which are then expressed as a functional SNN phenotype.

Spiking Neural Networks and Plasticity

The system employs Spiking Neural Networks (SNNs), where neurons communicate via discrete temporal spikes. Adaptation occurs during an agent's lifetime through Spike-Timing-Dependent Plasticity (STDP). The evolution process selects from various STDP curves—such as Hebbian or Anti-Hebbian, symmetric or asymmetric—to determine how synaptic strengths change based on the timing of pre- and post-synaptic spikes.

Four types of evolved STDP learning rules: (a) Asymmetric Hebbian, (b) Asymmetric Anti-Hebbian, (c) Symmetric Hebbian, and (d) Symmetric Anti-Hebbian.

To maintain stability, the framework incorporates a homeostasis mechanism. This prevents runaway excitation or total silence, ensuring the network remains in a state capable of processing new information.

Experimental Methodology

The framework was evaluated in three "mutable" environments where the rules of the task change during the agent's lifetime:

  1. Food Foraging: A binary task where agents must eat or avoid items based on color. The edibility of colors is periodically swapped, requiring the agent to relearn associations.
  2. Logic Gates: The agent must emulate different logic operations (e.g., AND, XOR). The required operation switches mid-simulation.
  3. Cart-pole Balancing: A control task where the pole length is modified during the run, forcing the controller to adapt its dynamics.

In these tasks, agents receive "reward" and "penalty" spikes that act as biological feedback, triggering the evolved STDP rules to rewire the network in real-time.

Key Results and Insights

The NAGI framework demonstrated high levels of adaptivity across all benchmarks:

  • Foraging: Achieved an average accuracy of 88.4%, with end-of-sample accuracy reaching 95.9%.
  • Logic Gates: Reached an average accuracy of 85.7% and an end-of-sample accuracy of 94.2%.
  • Cart-pole: Achieved near-perfect fitness (0.999), successfully adapting to pole lengths ranging from 0.3 to 0.7.

A significant finding was that complex adaptive behavior could emerge from surprisingly simple topologies, sometimes requiring no hidden layers. This suggests that well-tuned local learning rules are more critical for general adaptability than sheer network depth.

Discussion and Future Work

While the results are promising, the authors noted instability in performance across generations due to the randomness of weight initialization at "birth." Future work will focus on stabilizing these evolutionary trajectories and expanding the framework to multi-class classification and more complex control tasks, such as quadruped walking. Additionally, the researchers aim to relax the "one rule per neuron" constraint to allow for more sophisticated dendritic processing.

Summary updated

BibTeX
@article{pontesfilho2022towards,
  author = {Sidney Pontes-Filho and Kristoffer Olsen and Anis Yazidi and Michael A. Riegler and Pål Halvorsen and Stefano Nichele},
  title = {{Towards the Neuroevolution of Low-level Artificial General Intelligence}},
  journal = {Frontiers Robotics AI},
  year = {2022},
  doi = {10.3389/frobt.2022.1007547},
}
Figure