-
Human-interpretable model explainability on high-dimensional data
Authors:
Damien de Mijolla,
Christopher Frye,
Markus Kunesch,
John Mansir,
Ilya Feige
Abstract:
The importance of explainability in machine learning continues to grow, as both neural-network architectures and the data they model become increasingly complex. Unique challenges arise when a model's input features become high dimensional: on one hand, principled model-agnostic approaches to explainability become too computationally expensive; on the other, more efficient explainability algorithm…
▽ More
The importance of explainability in machine learning continues to grow, as both neural-network architectures and the data they model become increasingly complex. Unique challenges arise when a model's input features become high dimensional: on one hand, principled model-agnostic approaches to explainability become too computationally expensive; on the other, more efficient explainability algorithms lack natural interpretations for general users. In this work, we introduce a framework for human-interpretable explainability on high-dimensional data, consisting of two modules. First, we apply a semantically meaningful latent representation, both to reduce the raw dimensionality of the data, and to ensure its human interpretability. These latent features can be learnt, e.g. explicitly as disentangled representations or implicitly through image-to-image translation, or they can be based on any computable quantities the user chooses. Second, we adapt the Shapley paradigm for model-agnostic explainability to operate on these latent features. This leads to interpretable model explanations that are both theoretically controlled and computationally tractable. We benchmark our approach on synthetic data and demonstrate its effectiveness on several image-classification tasks.
△ Less
Submitted 20 December, 2021; v1 submitted 14 October, 2020;
originally announced October 2020.
-
Linear hyperfine tuning of donor spins in silicon using hydrostatic strain
Authors:
John Mansir,
Pierandrea Conti,
Zai** Zeng,
Jarryd J. Pla,
Patrice Bertet,
Michael W. Swift,
Chris G. Van de Walle,
Mike L. W. Thewalt,
Benoit Sklenard,
Yann-Michel Niquet,
John J. L. Morton
Abstract:
We experimentally study the coupling of Group V donor spins in silicon to mechanical strain, and measure strain-induced frequency shifts which are linear in strain, in contrast to the quadratic dependence predicted by the valley repopulation model (VRM), and therefore orders of magnitude greater than that predicted by the VRM for small strains $|\varepsilon| < 10^{-5}$. Through both tight-binding…
▽ More
We experimentally study the coupling of Group V donor spins in silicon to mechanical strain, and measure strain-induced frequency shifts which are linear in strain, in contrast to the quadratic dependence predicted by the valley repopulation model (VRM), and therefore orders of magnitude greater than that predicted by the VRM for small strains $|\varepsilon| < 10^{-5}$. Through both tight-binding and first principles calculations we find that these shifts arise from a linear tuning of the donor hyperfine interaction term by the hydrostatic component of strain and achieve semi-quantitative agreement with the experimental values. Our results provide a framework for making quantitative predictions of donor spins in silicon nanostructures, such as those being used to develop silicon-based quantum processors and memories. The strong spin-strain coupling we measure (up to 150~GHz per strain, for Bi-donors in Si), offers a method for donor spin tuning --- shifting Bi donor electron spins by over a linewidth with a hydrostatic strain of order $10^{-6}$ --- as well as opportunities for coupling to mechanical resonators.
△ Less
Submitted 23 March, 2018; v1 submitted 2 October, 2017;
originally announced October 2017.
-
Strain-induced spin resonance shifts in silicon devices
Authors:
J. J. Pla,
A. Bienfait,
G. Pica,
J. Mansir,
F. A. Mohiyaddin,
Z. Zeng,
Y. M. Niquet,
A. Morello,
T. Schenkel,
J. J. L. Morton,
P. Bertet
Abstract:
In spin-based quantum information processing devices, the presence of control and detection circuitry can change the local environment of a spin by introducing strain and electric fields, altering its resonant frequencies. These resonance shifts can be large compared to intrinsic spin line-widths and it is therefore important to study, understand and model such effects in order to better predict d…
▽ More
In spin-based quantum information processing devices, the presence of control and detection circuitry can change the local environment of a spin by introducing strain and electric fields, altering its resonant frequencies. These resonance shifts can be large compared to intrinsic spin line-widths and it is therefore important to study, understand and model such effects in order to better predict device performance. Here we investigate a sample of bismuth donor spins implanted in a silicon chip, on top of which a superconducting aluminium micro-resonator has been fabricated. The on-chip resonator provides two functions: first, it produces local strain in the silicon due to the larger thermal contraction of the aluminium, and second, it enables sensitive electron spin resonance spectroscopy of donors close to the surface that experience this strain. Through finite-element strain simulations we are able to reconstruct key features of our experiments, including the electron spin resonance spectra. Our results are consistent with a recently discovered mechanism for producing shifts of the hyperfine interaction for donors in silicon, which is linear with the hydrostatic component of an applied strain.
△ Less
Submitted 9 January, 2018; v1 submitted 25 August, 2016;
originally announced August 2016.
-
Charge dynamics and spin blockade in a hybrid double quantum dot in silicon
Authors:
M. Urdampilleta,
A. Chatterjee,
C. C. Lo,
T. Kobayashi,
J. Mansir,
S. Barraud,
A. C. Betz,
S. Rogge,
M. F. Gonzalez-Zalba,
J. J. L. Morton
Abstract:
Electron spin qubits in silicon, whether in quantum dots or in donor atoms, have long been considered attractive qubits for the implementation of a quantum computer due to the semiconductor vacuum character of silicon and its compatibility with the microelectronics industry. While donor electron spins in silicon provide extremely long coherence times and access to the nuclear spin via the hyperfin…
▽ More
Electron spin qubits in silicon, whether in quantum dots or in donor atoms, have long been considered attractive qubits for the implementation of a quantum computer due to the semiconductor vacuum character of silicon and its compatibility with the microelectronics industry. While donor electron spins in silicon provide extremely long coherence times and access to the nuclear spin via the hyperfine interaction, quantum dots have the complementary advantages of fast electrical operations, tunability and scalability. Here we present an approach to a novel hybrid double quantum dot by coupling a donor to a lithographically patterned artificial atom. Using gate-based rf reflectometry, we probe the charge stability of this double quantum dot system and the variation of quantum capacitance at the interdot charge transition. Using microwave spectroscopy, we find a tunnel coupling of 2.7 GHz and characterise the charge dynamics, which reveals a charge T2* of 200 ps and a relaxation time T1 of 100 ns. Additionally, we demonstrate spin blockade at the inderdot transition, opening up the possibility to operate this coupled system as a singlet-triplet qubit or to transfer a coherent spin state between the quantum dot and the donor electron and nucleus.
△ Less
Submitted 19 March, 2015; v1 submitted 3 March, 2015;
originally announced March 2015.
-
Hybrid optical-electrical detection of donor electron spins with bound excitons in silicon
Authors:
C. C. Lo,
M. Urdampilleta,
P. Ross,
M. F. Gonzalez-Zalba,
J. Mansir,
S. A. Lyon,
M. L. W. Thewalt,
J. J. L. Morton
Abstract:
Electrical detection of spins is an essential tool in understanding the dynamics of spins in semiconductor devices, providing valuable insights for applications ranging from optoelectronics and spintronics to quantum information processing. For electron spins bound to shallow donors in silicon, bulk electrically-detected magnetic resonance has relied on coupling to spin readout partners such as pa…
▽ More
Electrical detection of spins is an essential tool in understanding the dynamics of spins in semiconductor devices, providing valuable insights for applications ranging from optoelectronics and spintronics to quantum information processing. For electron spins bound to shallow donors in silicon, bulk electrically-detected magnetic resonance has relied on coupling to spin readout partners such as paramagnetic defects or conduction electrons which fundamentally limits spin coherence times. Here we demonstrate electrical detection of phosphorus donor electron spin resonance by transport through a silicon device, using optically-driven donor-bound exciton transitions. We use this method to measure electron spin Rabi oscillations, and, by avoiding use of an ancillary spin for readout, we are able to obtain long intrinsic electron spin coherence times, limited only by the donor concentration. We go on to experimentally address critical issues for adopting this scheme for single spin measurement in silicon nanodevices, including the effects of strain, electric fields, and non-resonant excitation. This lays the foundations for realising a versatile readout method for single spin readout with relaxed magnetic field and temperature requirements compared with spin-dependent tunneling.
△ Less
Submitted 5 November, 2014;
originally announced November 2014.
