-
Random feedback weights support learning in deep neural networks
Authors:
Timothy P. Lillicrap,
Daniel Cownden,
Douglas B. Tweed,
Colin J. Akerman
Abstract:
The brain processes information through many layers of neurons. This deep architecture is representationally powerful, but it complicates learning by making it hard to identify the responsible neurons when a mistake is made. In machine learning, the backpropagation algorithm assigns blame to a neuron by computing exactly how it contributed to an error. To do this, it multiplies error signals by ma…
▽ More
The brain processes information through many layers of neurons. This deep architecture is representationally powerful, but it complicates learning by making it hard to identify the responsible neurons when a mistake is made. In machine learning, the backpropagation algorithm assigns blame to a neuron by computing exactly how it contributed to an error. To do this, it multiplies error signals by matrices consisting of all the synaptic weights on the neuron's axon and farther downstream. This operation requires a precisely choreographed transport of synaptic weight information, which is thought to be impossible in the brain. Here we present a surprisingly simple algorithm for deep learning, which assigns blame by multiplying error signals by random synaptic weights. We show that a network can learn to extract useful information from signals sent through these random feedback connections. In essence, the network learns to learn. We demonstrate that this new mechanism performs as quickly and accurately as backpropagation on a variety of problems and describe the principles which underlie its function. Our demonstration provides a plausible basis for how a neuron can be adapted using error signals generated at distal locations in the brain, and thus dispels long-held assumptions about the algorithmic constraints on learning in neural circuits.
△ Less
Submitted 2 November, 2014;
originally announced November 2014.
-
The C/O ratio at low metallicity: constraints on early chemical evolution from observations of Galactic halo stars
Authors:
D. Fabbian,
P. E. Nissen,
M. Asplund,
M. Pettini,
C. Akerman
Abstract:
We present new measurements of the abundances of carbon and oxygen derived from high-excitation C I and O I absorption lines in metal-poor halo stars, with the aim of clarifying the main sources of these two elements in the early stages of the chemical enrichment of the Galaxy. We target 15 new stars compared to our previous study, with an emphasis on additional C/O determinations in the crucial…
▽ More
We present new measurements of the abundances of carbon and oxygen derived from high-excitation C I and O I absorption lines in metal-poor halo stars, with the aim of clarifying the main sources of these two elements in the early stages of the chemical enrichment of the Galaxy. We target 15 new stars compared to our previous study, with an emphasis on additional C/O determinations in the crucial metallicity range -3<[Fe/H]<-2. Departures from local thermodynamic equilibrium were accounted for in the line formation for both carbon and oxygen. The non-LTE effects are very strong at the lowest metallicities but, contrary to what has sometimes been assumed in the past due to a simplified assessment, of different degrees for the two elements. In addition, for the 28 stars with [Fe/H]<-1 previously analysed, stellar parameters were re-derived and non-LTE corrections applied in the same fashion as for the rest of our sample, giving consistent abundances for 43 halo stars in total. The new observations and non-LTE calculations strengthen previous suggestions of an upturn in C/O towards lower metallicity (particularly for [O/H]<-2). Adopting the H collisional cross-sections estimated from the classical Drawin formula leads to [C/O]~0 at [O/H]~-3. To remove the upturn in C/O, near-LTE formation for O I lines would be required, which could only happen if the H collisional efficiency with the Drawin recipe is underestimated by factors of up to several tens of times, which we consider unlikely. The high C/O values derived at the lowest metallicities may be revealing the fingerprints of Population III stars or may signal rotationally-aided nucleosynthesis in more normal Population II stars.
△ Less
Submitted 1 October, 2008;
originally announced October 2008.
-
Sulphur and zinc abundances in Galactic halo stars revisited
Authors:
P. E. Nissen,
C. Akerman,
M. Asplund,
D. Fabbian,
F. Kerber,
H. U. Kaufl,
M. Pettini
Abstract:
High resolution UVES spectra of 40 main-sequence stars with -3.3 < [Fe/H] < -1.0 are used to derive S, Fe and Zn abundances from lines in the 400 - 950 nm region. For one star we also present novel observations of the SI triplet at 1.046 micron carried out with the ESO VLT CRIRES spectrograph. Comparison of sulphur abundances from the weak and strong SI lines provides important constraints on no…
▽ More
High resolution UVES spectra of 40 main-sequence stars with -3.3 < [Fe/H] < -1.0 are used to derive S, Fe and Zn abundances from lines in the 400 - 950 nm region. For one star we also present novel observations of the SI triplet at 1.046 micron carried out with the ESO VLT CRIRES spectrograph. Comparison of sulphur abundances from the weak and strong SI lines provides important constraints on non-LTE effects. The high sulphur abundances reported by others for some metal-poor stars are not confirmed; instead, when taking non-LTE effects into account, the Galactic halo stars distribute around a plateau at [S/Fe] = +0.2 dex with a scatter of 0.07 dex only. This indicates that sulphur in Galactic halo stars has been made by alpha-capture processes in massive SNe. The observed scatter in S/Fe is, however, much smaller than predicted from current stochastic models of the chemical evolution of the early Galaxy, suggesting that either the models or the calculated yields of massive SNe should be revised. [Zn/Fe] is close to zero for metallicities in the range -2.0 < [Fe/H] < -1.0 but increases to a level of [Zn/Fe] = +0.15 dex in the range -2.7 < [Fe/H] < -2.0. At still lower metallicities [Zn/Fe] rises steeply to a value around [Zn/Fe] = +0.5 dex at [Fe/H] = -3.2. We also examine the behaviour of S/Zn and find that departures from the solar ratio are significantly reduced at all metallicities if non-LTE corrections to the abundances of these two elements are adopted. This effect, if confirmed, would reduce the usefulness of the S/Zn ratio as a diagnostic of past star-formation activity, but would bring closer together the values measured in damped Lyman-alpha systems and in Galactic stars.
△ Less
Submitted 25 April, 2007; v1 submitted 26 February, 2007;
originally announced February 2007.
-
Zn and Cr abundances in damped Lyman alpha systems from the CORALS survey
Authors:
Chris J. Akerman,
Sara L. Ellison,
Max Pettini,
Charles C. Steidel
Abstract:
We present metal abundances in 15 damped Lyman alpha systems (DLAs) from the Complete Optical and Radio Absorption Line System (CORALS) survey, designed to be free from any biasing effects due to extinction of QSOs by dust in intervening absorbers. It has long been suggested that such biasing may explain differences in metallicity between DLAs and coeval luminous galaxies, and between model pred…
▽ More
We present metal abundances in 15 damped Lyman alpha systems (DLAs) from the Complete Optical and Radio Absorption Line System (CORALS) survey, designed to be free from any biasing effects due to extinction of QSOs by dust in intervening absorbers. It has long been suggested that such biasing may explain differences in metallicity between DLAs and coeval luminous galaxies, and between model predictions and observations. We use our measured zinc and chromium abundances (combined with those for five more CORALS DLAs from the literature, giving us a very nearly complete sample) to test whether the metallicity and degree of dust depletion in CORALS DLAs are significantly different from those of existing, larger, samples of DLAs drawn from magnitude limited, optical surveys. We find that the column density weighted metallicity of CORALS DLAs, [<(Zn/H)_DLA>] = -0.88 +/- 0.21 in the redshift interval 1.86 < z(abs) < 3.45, is only marginally higher than that of a control sample from the recent compilation by Kulkarni et al., [<(Zn/H)_DLA>] = -1.09 +/- 0.10. With the present limited statistics this difference is not highly significant. Furthermore, we find no evidence for increased dust depletions in CORALS DLAs--their [Cr/Zn] ratios conform to the known trend of increasing depletion (decreasing [Cr/Zn]) with increasing metallicity, and we have encountered no cases where Cr is as depleted as in local cold interstellar clouds. These results, when combined with the earlier findings of the CORALS survey reported by Ellison et al. in 2001, make it difficult to invoke a dust-induced bias to explain the generally low level of chemical evolution exhibited by most DLAs. Rather, they indicate that large scale optical QSO surveys give a fair census of the population of high redshift absorbers.
△ Less
Submitted 8 June, 2005;
originally announced June 2005.
-
The evolution of the C/O ratio in metal-poor halo stars
Authors:
Chris J. Akerman,
Leticia Carigi,
Poul E. Nissen,
Max Pettini,
Martin Asplund
Abstract:
We report new measurements of carbon and oxygen abundances in 34 F and G dwarf and subgiant stars belonging to the halo population and spanning a range of metallicity from [Fe/H] = -0.7 to -3.2 . The survey is based on observations of four permitted lines of C I near 9100 A, and the O I 7774 A triplet, all recorded at high signal-to-noise ratios with the UVES echelle spectrograph on the ESO VLT.…
▽ More
We report new measurements of carbon and oxygen abundances in 34 F and G dwarf and subgiant stars belonging to the halo population and spanning a range of metallicity from [Fe/H] = -0.7 to -3.2 . The survey is based on observations of four permitted lines of C I near 9100 A, and the O I 7774 A triplet, all recorded at high signal-to-noise ratios with the UVES echelle spectrograph on the ESO VLT. The line equivalent widths were analysed with the 1D, LTE, MARCS model atmosphere code to deduce C and O abundances; corrections due to non-LTE and 3D effects are discussed. When combined with similar published data for disk stars, our results confirm the metallicity dependence of the C/O ratio known from previous stellar and interstellar studies: C/O drops by a factor of 3-4 as O/H decreases from solar to 1/10 solar. Analysed within the context of standard models for the chemical evolution of the solar vicinity, this drop results from the metallicity dependence of the C yields from massive stars with mass loss, augmented by the delayed release of C from stars of low and intermediate mass. The former is, however, always the dominant factor. Our survey has also uncovered tentative evidence to suggest that, as the oxygen abundance decreases below [O/H] = -1, [C/O] may not remain constant at [C/O] = -0.5, as previously thought, but increase again, possibly approaching near-solar values at the lowest metallicities ([O/H] approximately equal to -3). With the current dataset this is no more than a 3 sigma effect and it may be due to metallicity-dependent non-LTE corrections to the [C/O] ratio which have not been taken into account. However, its potential importance as a window on the nucleosynthesis by Population III stars is a strong incentive for future work, both observational and theoretical, to verify its reality.
△ Less
Submitted 16 October, 2003;
originally announced October 2003.
