Sampling from exponential distributions in the time domain with superparamagnetic tunnel junctions

Temitayo N. Adeyeye, Sidra Gibeault, Daniel P. Lathrop, Matthew W. Daniels, Mark D. Stiles, Jabez J. McClelland, William A. Borders, Jason T. Ryan, Philippe Talatchian, Ursula Ebels, Advait Madhavan · 2024

Read Sampling from exponential distributions in the time domain with superparamagnetic tunnel junctions by Temitayo N. Adeyeye, Sidra Gibeault, Daniel P. Lathrop, Matthew W. Daniels, Mark D. Stiles, Jabez J. McClelland, William A. Borders, Jason T. Ryan, Philippe Talatchian, Ursula Ebels, Advait Madhavan (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-10317.txt.

Introduction

Temitayo N Adeyeye and ten co-authors build computer parts that draw random numbers from the physics of tiny magnets. Their 2024 paper uses magnetic tunnel junctions, small devices that flip between two resistance states at random. The waiting time before each flip follows an exponential pattern, a shape many computing tasks need.

The team made a circuit cell that starts a current step and times the first flip. Repeated tests confirmed the waiting times match the wanted pattern. They then show how rows of these cells can power two heavy tasks: a Metropolis-Hastings stepper and a weighted random sampler. Using natural randomness could make such sampling far cheaper in energy.

Catalog record

CREATORTemitayo N. Adeyeye, Sidra Gibeault, Daniel P. Lathrop, Matthew W. Daniels, Mark D. Stiles, Jabez J. McClelland, William A. Borders, Jason T. Ryan, Philippe Talatchian, Ursula Ebels, Advait Madhavan
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-10317

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Sampling from exponential distributions in the time domain with superparamagnetic tunnel junctions Temitayo N. Adeyeye, Sidra Gibeault, Daniel P. Lathrop, Matthew W. Daniels, Mark D. Stiles, Jabez J. McClelland, William A. Borders, Jason T. Ryan, Philippe Talatchian, Ursula Ebels, Advait Madhavan Abstract In the superparamagnetic regime, magnetic tunnel junctions switch between two resistance states due to random thermal fluctuations. The dwell time distribution in each state is exponential. We sample this distribution using a temporal encoding scheme, in which information is encoded in the time at which the device switches between its resistance states. We then develop a circuit element known as a probabilistic delay cell that applies an electrical current step to a superparamagnetic tunnel junction and a temporal measurement circuit that measures the timing of the first switching event. Repeated experiments confirm that these times are exponentially distributed. Temporal processing methods then allow us to digitally compute with these exponentially distributed probabilistic delay cells. We describe how to use these circuits in a Metropolis-Hastings stepper and in a weighted random sampler, both of which are computationally intensive applications that benefit from the efficient generation of exponentially distributed random numbers. Preprint: APS/123-QED I Introduction Though exponential distributions are ubiquitous in statistical physics and related computational models, directly sampling them from device behavior is rarely done. The magnetic tunnel junction (MTJ) operating in the superparamagnetic regime, a key device in probabilistic computing, is known to naturally exhibit exponentially distributed temporal switching dynamics [ 1 , 2 , 3 ] . Contemporary approaches focus on sampling the instantaneous state of the device [ 4 , 5 ] and treat each switching event as a Bernoulli trial resulting in a discrete sample. To sample an exponential distribution with an MTJ, we need to measure it in the time domain, which is challenging with traditional techniques. A time-domain-based sampling technique would result in exponentially distributed analog time values, which can be further processed based on application requirements. The importance of the exponential distribution cannot be overstated. Characterized by a rate parameter, it is central to any Markovian description of evolution through state space and to any energy-based objective function description since the Boltzmann distribution that describes the occupancy of the states at equilibrium is an exponential function of their energies [ 6 ] . This property is exploited in the Metropolis-Hastings algorithm for Markov chain Monte Carlo sampling, by accepting transitions to higher energy states with a probability proportional to an exponential function of the difference in energies [ 7 ] . Exponential distributions can also be used to generate weighted random samples by selecting the argument of the smallest drawn sample from a set of exponential variates, a technique known as the “exponential clocks method” [ 8 ] . This can be used to generate arbitrarily weighted random samples, akin to rolling an arbitrarily biased n -sided die. While sampling techniques based on measuring the instantaneous states of MTJ dynamics are extremely useful for probabilistic computations, using the resulting Bernoulli distributions to sample from exponential distributions would require significant memory and computational overhead. The exponential distributions reside in the temporal characteristics, not state distributions, of SMTJ switching events. The ability to efficiently sample exponential distributions has the potential to enhance the efficiency of different types of probabilistic computing. In this work, we design a circuit aimed at drawing samples from exponential distributions by measuring the temporal statistics of the first switching events of MTJs when exposed to a current step. As shown in Fig. 1 (a,b), the MTJ is first biased to be in the parallel state just outside the superparamagnetic regime (on experimentally relevant time scales). Subsequently, a current step is applied, putting the MTJ in a superparamagnetic regime where there is eventually a transition to the antiparallel state after a probabilistic delay that depends on the magnitude of the current step. The time elapsed before this switching event occurs (shown in Fig. 1 (c)) is measured with embedded electronics and is then used to demonstrate the exponential switching characteristics of this probabilistic delay cell (PDC). By creating switching events that are temporally encoded and drawn from an exponential distribution, we can utilize this delay cell alongside other temporal computing primitives to make Boolean decisions based on the arrival times of both deterministic and stochastic temporal signals. We extend this approach to the “exponential clocks method” and propose a circuit to draw weighted random samples from arbitrary distributions. Figure 1: Extracting temporal stochasticity with a current step: Panel (a) shows the current flowing through the device as a function of time. Initially, a low current value keeps the device in the parallel state. At a clock edge, a current step is applied across the device, putting it in the superparamagnetic state, causing a switching event after some delay as is shown in panel (b). The time taken for this switching event t is the quantity of interest, and is measured using on-board electronics. Panel (c) shows the experimentally measured exponential distribution of the times of the first arriving switch, as measured with the circuit shown in Fig. 2 (a). The slope of the straight line fit to the data determines the empirical mean \tau=\langle t\rangle . The error bars are calculated as the square root of the number of entries in each bin. Approximately 10,000 switching events were collected for a current step with a constant amplitude and width. While the statistical uncertainties are smaller at longer times due to the reduced number of counts, they appear larger on the logarithmic scale. The reduced chi-squared statistic for the fit is \chi^{2} = 1.92, indicating that the statistical variations provide the primary source of uncertainty in the data. The rest of the paper is organized as follows. Section II gives background information on probabilistic computing, magnetic tunnel junctions, and come existing methods of producing random samples. Section III describes the circuit and the experimental setup used to interrogate the device and collect data. Section IV describes the statistical properties of the first switching event and its tunability with current. Section V describes how these circuits can be used in applications such as a Metropolis-Hastings stepper and a weighted random sampler, followed by a discussion in Sec. VI . II Background Probabilistic computing [ 9 , 10 ] is a modern computing paradigm inspired by the statistical physics of interacting bodies, such as gas molecules in a box [ 11 ] , metal atoms during an annealing process [ 7 ] , or magnetic moments in a ferromagnetic material [ 12 ] . Various combinatorial optimization problems with real-world utility can be solved by formulating the constraints of the problem into an objective function, which is analogous to the energy landscape of a physical system [ 7 ] . The objective function mathematically describes the interactions between participating bodies in configuration space. The configuration that minimizes this objective function corresponds to the solution [ 7 ] . Broadly speaking, there are two approaches to probabilistic computing. One approach [ 13 , 14 ] encodes the configuration space of such problems in probabilistic bits, also known as p -bits, each of which can take one of two binary values, 0 or 1, but do

Read the complete text