Hardware Implications

B5L-B: Hardware Implications

Session Type: Lecture
Session Code: B5L-B
Location: Room 2
Date & Time: Thursday March 23, 2023 (15:20-16:20)
Chair: Philippe Pouliquen
Track: 1

Paper ID	Paper Title	Authors	Abstract
3047	Impedance Estimation at Multi-Antenna Receivers	Shaohan Wu{1}, Brian Hughes{2}	This paper considers antenna impedance matrix estimation using training sequences at multi-antenna receivers. We formulate this problem in a classical estimation framework, treating the unknown channel matrix as a nuisance parameter. The maximum-likelihood (ML) impedance estimator is derived as a function of the top block eigenvector of the sample covariance matrix in i.i.d. Rayleigh fading channels. Fundamental lower bounds, e.g., Cramer-Rao bounds (CRB) and Miller-Chang bounds (MCB), on these estimators are derived, the latter of which better predicts the ML estimator\'s behavior using numerical simulations. Another finding is that transmit diversity may lead to diminishing returns if the total transmitted power is the constraint.
3082	Application of Machine Learning to Quantum Cascade Laser Design	Andres Correa Hernandez, Claire Gmachl	A framework to innovate quantum cascade laser design was developed using machine learning. An 8.2 μm laser with an operating field of 51 kV/cm and 131.7 eV ps Å^2 for the figure of merit was chosen as the starting design. A dataset of 13200 different designs was generated from the original, each design consisting of 22 well/barrier thicknesses with random well/barrier changes in the [-5, +5] Å range and an applied electric field from 20-70 kV/cm, for a total of 23 inputs. A second completely random dataset with 22 well/barrier thicknesses in the [9, 57] Å range was also developed, with 13200 designs and the same electric field sweep. The lasing transition figure of merit, gain coefficient, dipole matrix element, scattering times, and electronic state-pair energy difference were identified for each of these designs and were the outputs to be predicted. Single-output regression and multi-output regression were used to predict the figure of merit. Single-output regression was able to give root-mean-square error of 16 to 24 for the figure of merit in a matter of seconds. Multi-output regression predicts the root-mean-square error from 17 to 19 when trained for 10 to 30 minutes, sweeping through neural network parameters of interest. Both types of algorithms predicted data that was similar to training data very well but did not perform as well when new data was introduced.
3171	Linearization of Non-Uniform Quantizers via Adaptive Non-Subtractive Dithering	Morriel Kasher{2}, Predrag Spasojevic{2}, Michael Tinston{1}	Non-subtractive dithering is an effective method to improve quantizer performance by injecting input noise that reduces statistical correlations between input signals and quantization error. Existing non-subtractive dither theory has primarily designed dither signal distributions for linear, uniform quantizers, neglecting real-world non-idealities including non-uniformity and finite-level saturation. We develop a generalized analytical condition to guarantee independence of the quantization error moments from the input signal for an arbitrary finite-level non- linear quantizer characteristic. We use this to propose a novel asymmetric, adaptive dither technique for effective linearization of non-uniform quantizers via reduction of the first conditional quantization error moment. These adaptive dither distributions are shown to completely eliminate the first error moment in Lloyd-Max quantizers and significantly reduce it in non-linear quantizers. This allows the use of time-averaging to converge to an arbitrarily precise signal estimate in non-uniform quantizers.