skip to main content

Cheap parallel computing has greatly extended the reach of ranking & selection (R&S) for simulation optimization. In this paper we present an evaluation of bi-PASS, a R&S procedure created specifically for parallel implementation and very large numbers of system designs. We compare bi-PASS to the state-ofthe- art Good Selection Procedure and an easy-to-implement subset selection procedure. This is one of the few papers to consider both computational and statistical comparison of parallel R&S procedures.
; ;
Bae, K-H; Feng, B; Kim, S; Lazarova-Molnar, S; Zheng, Z; Roeder, T; Thiesing, R
Award ID(s):
Publication Date:
Journal Name:
Proceedings of the Winter Simulation Conference
Page Range or eLocation-ID:
Sponsoring Org:
National Science Foundation
More Like this
  1. ABSTRACT We calculate the fundamental stellar parameters effective temperature, surface gravity, and iron abundance – Teff, log g, [Fe/H] – for the final release of the Mapping Nearby Galaxies at APO (MaNGA) Stellar Library (MaStar), containing 59 266 per-visit-spectra for 24 290 unique stars at intermediate resolution (R ∼ 1800) and high S/N (median = 96). We fit theoretical spectra from model atmospheres by both MARCS and BOSZ-ATLAS9 to the observed MaStar spectra, using the full spectral fitting code pPXF. We further employ a Bayesian approach, using a Markov Chain Monte Carlo (MCMC) technique to map the parameter space and obtain uncertainties.more »Originally in this paper, we cross match MaStar observations with Gaia photometry, which enable us to set reliable priors and identify outliers according to stellar evolution. In parallel to the parameter determination, we calculate corresponding stellar population models to test the reliability of the parameters for each stellar evolutionary phase. We further assess our procedure by determining parameters for standard stars such as the Sun and Vega and by comparing our parameters with those determined in the literature from high-resolution spectroscopy (APOGEE and SEGUE) and from lower resolution matching template (LAMOST). The comparisons, considering the different methodologies and S/N of the literature surveys, are favourable in all cases. Our final parameter catalogue for MaStar cover the following ranges: 2592 ≤ Teff ≤ 32 983 K; −0.7 ≤ log g ≤ 5.4 dex; −2.9 ≤ [Fe/H] ≤ 1.0 dex and will be available with the last SDSS-IV Data Release, in 2021 December.« less
  2. DNN training is extremely time-consuming, necessitating efficient multi-accelerator parallelization. Current approaches to parallelizing training primarily use intra-batch parallelization, where a single iteration of training is split over the available workers, but suffer from diminishing returns at higher worker counts. We present PipeDream, a system that adds inter-batch pipelining to intra-batch parallelism to further improve parallel training throughput, helping to better overlap computation with communication and reduce the amount of communication when possible. Unlike traditional pipelining, DNN training is bi-directional, where a forward pass through the computation graph is followed by a backward pass that uses state and intermediate data computedmore »during the forward pass. Naïve pipelining can thus result in mismatches in state versions used in the forward and backward passes, or excessive pipeline flushes and lower hardware efficiency. To address these challenges, PipeDream versions model parameters for numerically correct gradient computations, and schedules forward and backward passes of different minibatches concurrently on different workers with minimal pipeline stalls. PipeDream also automatically partitions DNN layers among workers to balance work and minimize communication. Extensive experimentation with a range of DNN tasks, models, and hardware configurations shows that PipeDream trains models to high accuracy up to 5.3X faster than commonly used intra-batch parallelism techniques.« less
  3. Introduction:Current brain-computer interfaces (BCIs) primarily rely on visual feedback. However, visual feedback may not be sufficient for applications such as movement restoration, where somatosensory feedback plays a crucial role. For electrocorticography (ECoG)-based BCIs, somatosensory feedback can be elicited by cortical surface electro-stimulation [1]. However, simultaneous cortical stimulation and recording is challenging due to stimulation artifacts. Depending on the orientation of stimulating electrodes, their distance to the recording site, and the stimulation intensity, these artifacts may overwhelm the neural signals of interest and saturate the recording bioamplifiers, making it impossible to recover the underlying information [2]. To understand how these factorsmore »affect artifact propagation, we performed a preliminary characterization of ECoG signals during cortical stimulation.Materials/Methods/ResultsECoG electrodes were implanted in a 39-year old epilepsy patient as shown in Fig. 1. Pairs of adjacent electrodes were stimulated as a part of language cortical mapping. For each stimulating pair, a charge-balanced biphasic square pulse train of current at 50 Hz was delivered for five seconds at 2, 4, 6, 8 and 10 mA. ECoG signals were recorded at 512 Hz. The signals were then high-pass filtered (≥1.5 Hz, zero phase), and the 5-second stimulation epochs were segmented. Within each epoch, artifact-induced peaks were detected for each electrode, except the stimulating pair, where signals were clipped due to amplifier saturation. These peaks were phase-locked across electrodes and were 20 ms apart, thus matching the pulse train frequency. The response was characterized by calculating the median peak within the 5-second epochs. Fig. 1 shows a representative response of the right temporal grid (RTG), with the stimulation channel at RTG electrodes 14 and 15. It also shows a hypothetical amplifier saturation contour of an implantable, bi-directional, ECoG-based BCI prototype [2], assuming the supply voltage of 2.2 V and a gain of 66 dB. Finally, we quantify the worstcase scenario by calculating the largest distance between the saturation contour and the midpoint of each stimulating channel.Discussion:Our results indicate that artifact propagation follows a dipole potential distribution with the extent of the saturation region (the interior of the white contour) proportional to the stimulation amplitude. In general, the artifacts propagated farthest when a 10 mA current was applied with the saturation regions extending from 17 to 32 mm away from the midpoint of the dipole. Consistent with the electric dipole model, this maximum spread happened along the direction of the dipole moment. An exception occurred at stimulation channel RTG11-16, for which an additional saturation contour emerged away from the dipole contour (not shown), extending the saturation region to 41 mm. Also, the worst-case scenario was observed at 6 mA stimulation amplitude. This departure could be a sign of a nonlinear, switch-like behavior, wherein additional conduction pathways could become engaged in response to sufficiently high stimulation.Significance:While ECoG stimulation is routinely performed in the clinical setting, quantitative studies of the resulting signals are lacking. Our preliminary study demonstrates that stimulation artifacts largely obey dipole distributions, suggesting that the dipole model could be used to predict artifact propagation. Further studies are necessary to ascertain whether these results hold across other subjects and combinations of stimulation/recording grids. Once completed, these studies will reveal practical design constraints for future implantable bi-directional ECoG-based BCIs. These include parameters such as the distances between and relative orientations of the stimulating and recording electrodes, the choice of the stimulating electrodes, the optimal placement of the reference electrode, and the maximum stimulation amplitude. These findings would also have important implications for the design of custom, low-power bioamplifiers for implantable bi-directional ECoG-based BCIs.References:[1] Hiremath, S. V., et al. "Human perception of electrical stimulation on the surface of somatosensory cortex." PloS one 12.5 (2017): e0176020.[2] Rouse, A. G., et al. "A chronic generalized bi-directional brain-machine interface." Journal of Neural Engineering 8.3 (2011): 036018« less
  4. We build the plethora of counterexamples to bi-parameter two weight embedding theorems. Two weight one parameter embedding results (which is the same as results of boundedness of two weight classical paraproducts, or two weight Carleson embedding theorems) are well known since the works of Sawyer in the 80’s. Bi-parameter case was considered by S. Y. A. Chang and R. Fefferman but only when underlying measure is Lebesgue measure. The embedding of holomorphic functions on bi-disc requires general input measure. In [9] we classified such embeddings if the output measure has tensor structure. In this note we give examples that withoutmore »tensor structure requirement all results break down.« less
  5. A gr e at d e al of i nt er e st s urr o u n d s t h e u s e of tr a n s cr a ni al dir e ct c urr e nt sti m ul ati o n (t D C S) t o a u g m e nt c o g niti v e tr ai ni n g. H o w e v er, eff e ct s ar e i n c o n si st e nt a cr o s s st u dimore »e s, a n d m et aa n al yti c e vi d e n c e i s mi x e d, e s p e ci all y f o r h e alt h y, y o u n g a d ult s. O n e m aj or s o ur c e of t hi s i n c o n si st e n c y i s i n di vi d u al diff er e n c e s a m o n g t h e p arti ci p a nt s, b ut t h e s e diff er e n c e s ar e r ar el y e x a mi n e d i n t h e c o nt e xt of c o m bi n e d tr ai ni n g/ sti m ul ati o n st u di e s. I n a d diti o n, it i s u n cl e ar h o w l o n g t h e eff e ct s of sti m ul ati o n l a st, e v e n i n s u c c e s sf ul i nt er v e nti o n s. S o m e st u di e s m a k e u s e of f oll o w- u p a s s e s s m e nt s, b ut v er y f e w h a v e m e a s ur e d p erf or m a n c e m or e t h a n a f e w m o nt hs aft er a n i nt er v e nti o n. H er e, w e utili z e d d at a fr o m a pr e vi o u s st u d y of t D C S a n d c o g niti v e tr ai ni n g [ A u, J., K at z, B., B u s c h k u e hl, M., B u n arj o, K., S e n g er, T., Z a b el, C., et al. E n h a n ci n g w or ki n g m e m or y tr ai ni n g wit h tr a n scr a ni al dir e ct c urr e nt sti m ul ati o n. J o u r n al of C o g niti v e N e u r os ci e n c e, 2 8, 1 4 1 9 – 1 4 3 2, 2 0 1 6] i n w hi c h p arti ci p a nts tr ai n e d o n a w or ki n g m e m or y t as k o v er 7 d a y s w hil e r e c ei vi n g a cti v e or s h a m t D C S. A n e w, l o n g er-t er m f oll o w- u p t o a ss es s l at er p erf or m a n c e w a s c o n d u ct e d, a n d a d diti o n al p arti ci p a nt s w er e a d d e d s o t h at t h e s h a m c o n diti o n w a s b ett er p o w er e d. W e a s s e s s e d b a s eli n e c o g niti v e a bilit y, g e n d er, tr ai ni n g sit e, a n d m oti v ati o n l e v el a n d f o u n d si g nifi c a nt i nt er a cti o ns b et w e e n b ot h b as eli n e a bilit y a n d m oti v ati o n wit h c o n diti o n ( a cti v e or s h a m) i n m o d els pr e di cti n g tr ai ni n g g ai n. I n a d diti o n, t h e i m pr o v e m e nt s i n t h e a cti v e c o nditi o n v er s u s s h a m c o n diti o n a p p e ar t o b e st a bl e e v e n a s l o n g a s a y e ar aft er t h e ori gi n al i nt er v e nti o n. ■« less