Subscribe to the gold package and get unlimited access to Shamra Academy
Register a new userNo Multiplication? No Floating Point? No Problem! Training Networks for Efficient Inference
94
0
0.0
(
0
)
Ask ChatGPT about the research
No Arabic abstract
For successful deployment of deep neural networks on highly--resource-constrained devices (hearing aids, earbuds, wearables), we must simplify the types of operations and the memory/power resources used during inference. Completely avoiding inference-time floating-point operations is one of the simplest ways to design networks for these highly-constrained environments. By discretizing both our in-network non-linearities and our network weights, we can move to simple, compact networks without floating point operations, without multiplications, and avoid all non-linear function computations. Our approach allows us to explore the spectrum of possible networks, ranging from fully continuo
rate research
Read More
Can normal science-in the Kuhnian sense-add something substantial to the discussion about the measurement problem? Does an extended Wigners-friend Gedankenexperiment illustrate new issues? Or a new quality of known issues? Are we led to new interpretations, new perspectives, or do we iterate the previously known? The recent debate does, as we argue, neither constitute a turning point in the discussion about the measurement problem nor fundamentally challenge the legitimacy of quantum mechanics. Instead, the measurement problem asks for a reflection on fundamental paradigms of doing physics.
We consider multi-objective optimization (MOO) of an unknown vector-valued function in the non-parametric Bayesian optimization (BO) setting, with the aim being to learn points on the Pareto front of the objectives. Most existing BO algorithms do not model the fact that the multiple objectives, or equivalently, tasks can share similarities, and even the few that do lack rigorous, finite-time regret guarantees that capture explicitly inter-task structure. In this work, we address this problem by modelling inter-task dependencies using a multi-task kernel and develop two novel BO algorithms based on random scalarizations of the objectives. Our algorithms employ vector-valued kernel regression as a stepping stone and belong to the upper confidence bound class of algorithms. Under a smoothness assumption that the unknown vector-valued function is an element of the reproducing kernel Hilbert space associated with the multi-task kernel, we derive worst-case regret bounds for our algorithms that explicitly capture the similarities between tasks. We numerically benchmark our algorithms on both synthetic and real-life MOO problems, and show the advantages offered by learning with multi-task kernels.
We examine the dark matter content of satellite galaxies in Lambda-CDM cosmological hydrodynamical simulations of the Local Group from the APOSTLE project. We find excellent agreement between simulation results and estimates for the 9 brightest Galactic dwarf spheroidals (dSphs) derived from their stellar velocity dispersions and half-light radii. Tidal stripping plays an important role by gradually removing dark matter from the outside in, affecting in particular fainter satellites and systems of larger-than-average size for their luminosity. Our models suggest that tides have significantly reduced the dark matter content of Can Ven I, Sextans, Carina, and Fornax, a prediction that may be tested by comparing them with field galaxies of matching luminosity and size. Uncertainties in observational estimates of the dark matter content of individual dwarfs have been underestimated in the past, at times substantially. We use our improved estimates to revisit the `too-big-to-fail problem highlighted in earlier N-body work. We reinforce and extend our previous conclusion that the APOSTLE simulations show no sign of this problem. The resolution does not require `cores in the dark mass profiles, but, rather, relies on revising assumptions and uncertainties in the interpretation of observational data and accounting for `baryon effects in the theoretical modelling.
Since 2013 IceCube cascade showers sudden overabundance have shown a fast flavor change above 30-60 TeV up to PeV energy. This flavor change from dominant muon tracks at TeVs to shower events at higher energies, has been indebted to a new injection of a neutrino astronomy. However the recent published 54 neutrino HESE, high energy starting events, as well as the 38 external muon tracks made by trough going muon formed around the IceCube, none of them are pointing to any expected X-gamma or radio sources: no one in connection to GRB, no toward active BL Lac, neither to AGN source in Fermi catalog. No clear correlation with nearby mass distribution (Local Group), nor to galactic plane. Moreover there have not been any record (among a dozen of 200 TeV energetic events) of the expected double bang due to the tau neutrino birth and decay. An amazing and surprising unfair distribution in flavor versus expected democratic one. Finally there is not a complete consistence of the internal HESE event spectra and the external crossing muon track ones. Moreover the apparent sudden astrophysical neutrino flux rise at 60 TeV might be probably also suddenly cut at a few PeV in order to hide the (unobserved , yet) Glashow resonance peak at 6.3 PeV. A more mondane prompt charmed atmospheric neutrino component may explain most of the IceCube puzzles. If this near future, 2017-2018, it does not shine tau neutrino signals somewhere (by tau airshowers in AUGER, TA, ASHRA or double bang in IceCube) there are a list of consequences to face. These missing correlations and in particular the tau signature absence force us to claim : No Tau? No neutrino Astronomy.
Given a finite point set $P$ in the plane, a subset $S subseteq P$ is called an island in $P$ if $conv(S) cap P = S$. We say that $Ssubset P$ is a visible island if the points in $S$ are pairwise visible and $S$ is an island in $P$. The famous Big-line Big-clique Conjecture states that for any $k geq 3$ and $ell geq 4$, there is an integer $n = n(k,ell)$, such that every finite set of at least $n$ points in the plane contains $ell$ collinear points or $k$ pairwise visible points. In this paper, we show that this conjecture is false for visible islands, by constructing arbitrarily large finite point sets in the plane with no 4 collinear members and no visible island of size $2^{42}$.
suggested questions
Log in to be able to interact and post comments
comments
Fetching comments
Sign in to be able to follow your search criteria
