• The first gyroscope-on-rod-only design fails as a classifier: each gyroscope has no way to influence the others, so input information never reaches the output gyroscopes, and the system cannot be trained until springs connect them.
• Connecting 100 gyroscopes in a 10x10 hexagonal grid with 261 springs, the trained system reached 83.4% accuracy on the PenDigits stroke-classification task, versus 56.2% for a similarly complex linear classifier and 89.6% for an LSTM.
• Training uses ordinary backpropagation through a differentiable ODE model of the gyroscope-and-spring dynamics, the same machinery used for Neural ODEs, so gyroscope mass, rod length, and spring/gyroscopic coefficients are learned parameters.
• After training on GPU, inference can run on the physical gyroscope-and-spring system itself, or be emulated on a different physical system such as an electrical circuit, where 1 second of gyroscope time can map to 1 microsecond of circuit time.
• The authors note the emulated dynamics don't need to be physically realistic: a hardware surrogate could emulate 'negative mass' to boost the system's expressivity even though that value is not physically meaningful.

• inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.

• articleIntroducing Un-0: Generating Images with Coupled OscillatorsUnconventional AI
• articleSpeaking the Language of Physics: Introducing the First Instruction Set Architecture (ISA) for Dynamical System HardwareAng Li
• articleSystem Modeling: The Substrate for Neural Co-EvolutionCarlos H. Morales

Checking sign-in…

Loading comments…