Abstract
Quantum extreme learning machines (QELM) is a framework for quantum machine learning which aims to leverage the high dimensionality
of the Hilbert space of a quantum reservoir to construct a rich feature map of classical data inputs. There are promising results mostly
obtained on ideal simulators, but it is known that the inherent noise in quantum devices makes the state quickly converge to a specific mixed state
independent of the input, making the QELM input-agnostic and thus useless[1]. As a step towards practical application of QELM before fully
fault-tolerant hardware is realized, we propose an analysis of the performance of random reservoir unitaries with a controlled T count, which
can be easily constructed fault-tolerantly for CSS codes, compared to the widely used chaotic Ising model reservoir in terms of Fourier richness of
the outputs, and study their impact on the state convergence phenomenon when implemented fault-tolerantly for a small system.