-
PikeLPN: Mitigating Overlooked Inefficiencies of Low-Precision Neural Networks
Authors:
Marina Neseem,
Conor McCullough,
Randy Hsin,
Chas Leichner,
Shan Li,
In Suk Chong,
Andrew G. Howard,
Lukasz Lew,
Sherief Reda,
Ville-Mikko Rautio,
Daniele Moro
Abstract:
Low-precision quantization is recognized for its efficacy in neural network optimization. Our analysis reveals that non-quantized elementwise operations which are prevalent in layers such as parameterized activation functions, batch normalization, and quantization scaling dominate the inference cost of low-precision models. These non-quantized elementwise operations are commonly overlooked in SOTA…
▽ More
Low-precision quantization is recognized for its efficacy in neural network optimization. Our analysis reveals that non-quantized elementwise operations which are prevalent in layers such as parameterized activation functions, batch normalization, and quantization scaling dominate the inference cost of low-precision models. These non-quantized elementwise operations are commonly overlooked in SOTA efficiency metrics such as Arithmetic Computation Effort (ACE). In this paper, we propose ACEv2 - an extended version of ACE which offers a better alignment with the inference cost of quantized models and their energy consumption on ML hardware. Moreover, we introduce PikeLPN, a model that addresses these efficiency issues by applying quantization to both elementwise operations and multiply-accumulate operations. In particular, we present a novel quantization technique for batch normalization layers named QuantNorm which allows for quantizing the batch normalization parameters without compromising the model performance. Additionally, we propose applying Double Quantization where the quantization scaling parameters are quantized. Furthermore, we recognize and resolve the issue of distribution mismatch in Separable Convolution layers by introducing Distribution-Heterogeneous Quantization which enables quantizing them to low-precision. PikeLPN achieves Pareto-optimality in efficiency-accuracy trade-off with up to 3X efficiency improvement compared to SOTA low-precision models.
△ Less
Submitted 29 March, 2024;
originally announced April 2024.
-
Constraining the top-light initial mass function in the extended ultraviolet disk of M83
Authors:
R. P. V. Rautio,
A. E. Watkins,
H. Salo,
A. Venhola,
J. H. Knapen,
S. Comerón
Abstract:
The universality or non-universality of the initial mass function (IMF) has significant implications for determining star formation rates and star formation histories from photometric properties of stellar populations. We reexamine whether the IMF is deficient in high-mass stars (top-light) in the low-density environment of the outer disk of M83 and constrain the shape of the IMF therein. Using ar…
▽ More
The universality or non-universality of the initial mass function (IMF) has significant implications for determining star formation rates and star formation histories from photometric properties of stellar populations. We reexamine whether the IMF is deficient in high-mass stars (top-light) in the low-density environment of the outer disk of M83 and constrain the shape of the IMF therein. Using archival Galaxy Evolution Explorer (GALEX) far ultraviolet (FUV) and near ultraviolet (NUV) data and new deep OmegaCAM narrowband H$α$ imaging, we constructed a catalog of FUV-selected objects in the outer disk of M83. We counted H$α$-bright clusters and clusters that are blue in FUV$-$NUV in the catalog, measured the maximum flux ratio $F_{\mathrm{H}α}/f_{λ\mathrm{FUV}}$ among the clusters, and measured the total flux ratio $ΣF_{\mathrm{H}α}/Σf_{λ\mathrm{FUV}}$ over the catalog. We then compared these measurements to predictions from stellar population synthesis models made with a standard Salpeter IMF, truncated IMFs, and steep IMFs. We also investigated the effect of varying the assumed internal extinction on our results. We are not able to reproduce our observations with models using the standard Salpeter IMF or the truncated IMFs. It is only when assuming an average internal extinction of $0.10 < A_{\mathrm{V}} < 0.15$ in the outer disk stellar clusters that models with steep IMFs ($α> 3.1$) simultaneously reproduce the observed cluster counts, the maximum observed $F_{\mathrm{H}α}/f_{λ\mathrm{FUV}}$, and the observed $ΣF_{\mathrm{H}α}/Σf_{λ\mathrm{FUV}}$. Our results support a non-universal IMF that is deficient in high-mass stars in low-density environments.
△ Less
Submitted 6 November, 2023;
originally announced November 2023.
-
The multifarious ionization sources and disturbed kinematics of extraplanar gas in five low-mass galaxies
Authors:
R. P. V. Rautio,
A. E. Watkins,
S. Comerón,
H. Salo,
S. Díaz-García,
J. Janz
Abstract:
We investigate the origin of the extraplanar diffuse ionized gas (eDIG) and its predominant ionization mechanisms in five nearby (17-46 Mpc) low-mass ($10^9\text{-}10^{10}$ $M_{\odot}$) edge-on disk galaxies: ESO 157-49, ESO 469-15, ESO 544-27, IC 217, and IC 1553. We acquired Multi Unit Spectroscopic Explorer (MUSE) integral field spectroscopy and deep narrowband H$α$ imaging of our sample galaxi…
▽ More
We investigate the origin of the extraplanar diffuse ionized gas (eDIG) and its predominant ionization mechanisms in five nearby (17-46 Mpc) low-mass ($10^9\text{-}10^{10}$ $M_{\odot}$) edge-on disk galaxies: ESO 157-49, ESO 469-15, ESO 544-27, IC 217, and IC 1553. We acquired Multi Unit Spectroscopic Explorer (MUSE) integral field spectroscopy and deep narrowband H$α$ imaging of our sample galaxies. To investigate the connection between in-plane star formation and eDIG, we perform a photometric analysis of our narrowband H$α$ imaging. We measure eDIG scale heights of $h_{z\text{eDIG}} = 0.59 \text{-} 1.39$ kpc and find a positive correlation between them and specific star formation rates. In all galaxies, we also find a strong correlation between extraplanar and midplane radial H$α$ profiles. Using our MUSE data, we investigate the origin of eDIG via kinematics. We find ionized gas rotation velocity lags above the midplane with values between 10 and 27 km s$^{-1}$ kpc$^{-1}$. While we do find hints of an accretion origin for the ionized gas in ESO 157-49, IC 217, and IC 1553, overall the ionized gas kinematics of our galaxies do not match a steady galaxy model or any simplistic model of accretion or internal origin for the gas. We also construct standard diagnostic diagrams and emission-line maps (EW(H$α$), [NII]/H$α$, [SII]//H$α$, [OIII]/H$β$) and find regions consistent with mixed OB star and hot low-mass evolved stars (HOLMES) ionization, and mixed OB-shock ionization. Our results suggest that OB stars are the primary driver of eDIG ionization, while both HOLMES and shocks may locally contribute to the ionization of eDIG to a significant degree. Despite our galaxies' similar structures and masses, we find a surprisingly composite image of ionization mechanisms and a multifarious origin for the eDIG.
△ Less
Submitted 3 January, 2022;
originally announced January 2022.
