Under review at ICLR 2027 · September 2026
Eric Kryski
An untrained network of 1,024 coupled oscillators, read by a linear readout, set against its own input, the same network uncoupled, leaky-integrator banks and five trained networks of the same size on noisy spoken digits from held-out speakers. Across eight experiments and 6,353 runs, most of the accuracy comes from the input and the readout: the dynamics add memory, a free oscillator amplitude is the one design choice that moves accuracy by more than a point, and neither the lattice geometry, a cochlea included, nor the phase coupling function does.
SSRN Working Paper, September 2026
Eric Kryski
A survey of coupled-oscillator networks as a machine-learning substrate, organizing eighteen published oscillatory neural network systems around whether gradients reach the oscillator dynamics and how a model is trained around them. Argues that a substrate whose native operations are resonance and entrainment resembles how neurons evolved to sense physical signals, and closes with open directions for shared data, model evaluation, and controls.
INTERACT 2011 · LNCS vol. 6948 · Springer
Eric Kryski, Ehud Sharlin
Design of interactive personal vehicles that express behavioral, personality-like traits through motion to make commuting more satisfying. Presents the design goals, the evolution of the vehicle prototypes, and preliminary findings from a design critique evaluation.
