Home
neat-dnfs is a neuroevolutionary framework that extends NEAT to evolve Dynamic Neural Field (DNF) architectures—optimizing both intrinsic field dynamics (e.g., kernel parameters, time constants) and inter-field topology (fields + couplings).
Our evaluation follows a hierarchy of tasks where behaviour must emerge from continuous-time and space DNF dynamics. Each task page includes:
- A short task description + trial structure
- The highest-fitness solution (with minimal topology)
- Experiment-level statistics across runs (success rates, generations-to-threshold, architecture complexity, performance, notable runs)
These tasks target foundational DFT behaviours: detecting input, maintaining working memory, and selecting among competitors.
-
Detection Instability A transient stimulus should create a peak that propagates to the output, then both fields should return to baseline after stimulus removal.
-
Memory Instability Like detection, but the output must maintain a self-sustained peak after stimulus offset (working memory), while the input returns to baseline.
-
Selection Instability Two stimuli are presented; the input represents both, while the output must settle into winner-take-all selection of a single location, then decay after offset.
These require structural innovation—typically introducing an intermediate field that stores an internal trace and biases later decisions.
-
DMTS (Delayed Match-to-Sample) Sample → delay → test (match vs non-match). The system must internally retain the sample information during the delay and then select the matching item in the test phase.
-
IOR (Inhibition of Return) Cue → delay → test (old + new). The model should detect the novel cue first, via a persisting inhibitory trace of the cued location.
Simple boolean tasks included for comparison/coverage:
- AND
- XOR