arXiv Science⌕ Search

arXiv subjects

Brandon Carter

Publications and source records attributed to Brandon Carter.

At least 19 recordsLinked to original sources
arXiv2021

Lost in Pruning: The Effects of Pruning Neural Networks beyond Test Accuracy

Neural network pruning is a popular technique used to reduce the inference costs of modern, potentially overparameterized, networks. Starting from a pre-trained network, the process is as follows: remove redundant parameters, retrain, and repeat while maintaining the same test accuracy. The result is a model that is a fraction of the size of the original with comparable predictive performance (test accuracy). Here, we reassess and evaluate whether the use of test accuracy alone in the terminating condition is sufficient to ensure that the resulting model performs well across a wide spectrum of "harder" metrics such as generalization to out-of-distribution data and resilience to noise. Across evaluations on varying architectures and data sets, we find that pruned networks effectively approximate the unpruned model, however, the prune ratio at which pruned networks achieve commensurate performance varies significantly across tasks. These results call into question the extent of \emph{genuine} overparameterization in deep learning and raise concerns about the practicability of deploying pruned networks, specifically in the context of safety-critical systems, unless they are widely evaluated beyond test accuracy to reliably predict their performance. Our code is available at https://github.com/lucaslie/torchprune.

cs.LG↗
arXiv2021

Maximum n-times Coverage for Vaccine Design

We introduce the maximum $n$-times coverage problem that selects $k$ overlays to maximize the summed coverage of weighted elements, where each element must be covered at least $n$ times. We also define the min-cost $n$-times coverage problem where the objective is to select the minimum set of overlays such that the sum of the weights of elements that are covered at least $n$ times is at least $\tau$. Maximum $n$-times coverage is a generalization of the multi-set multi-cover problem, is NP-complete, and is not submodular. We introduce two new practical solutions for $n$-times coverage based on integer linear programming and sequential greedy optimization. We show that maximum $n$-times coverage is a natural way to frame peptide vaccine design, and find that it produces a pan-strain COVID-19 vaccine design that is superior to 29 other published designs in predicted population coverage and the expected number of peptides displayed by each individual's HLA molecules.

q-bio.QM↗
arXiv2020

Overinterpretation reveals image classification model pathologies

Image classifiers are typically scored on their test set accuracy, but high accuracy can mask a subtle type of model failure. We find that high scoring convolutional neural networks (CNNs) on popular benchmarks exhibit troubling pathologies that allow them to display high accuracy even in the absence of semantically salient features. When a model provides a high-confidence decision without salient supporting input features, we say the classifier has overinterpreted its input, finding too much class-evidence in patterns that appear nonsensical to humans. Here, we demonstrate that neural networks trained on CIFAR-10 and ImageNet suffer from overinterpretation, and we find models on CIFAR-10 make confident predictions even when 95% of input images are masked and humans cannot discern salient features in the remaining pixel-subsets. We introduce Batched Gradient SIS, a new method for discovering sufficient input subsets for complex datasets, and use this method to show the sufficiency of border pixels in ImageNet for training and testing. Although these patterns portend potential model fragility in real-world deployment, they are in fact valid statistical patterns of the benchmark that alone suffice to attain high test accuracy. Unlike adversarial examples, overinterpretation relies upon unmodified image pixels. We find ensembling and input dropout can each help mitigate overinterpretation.

cs.LG↗
arXiv2019

Embedding Comparator: Visualizing Differences in Global Structure and Local Neighborhoods via Small Multiples

Embeddings mapping high-dimensional discrete input to lower-dimensional continuous vector spaces have been widely adopted in machine learning applications as a way to capture domain semantics. Interviewing 13 embedding users across disciplines, we find comparing embeddings is a key task for deployment or downstream analysis but unfolds in a tedious fashion that poorly supports systematic exploration. In response, we present the Embedding Comparator, an interactive system that presents a global comparison of embedding spaces alongside fine-grained inspection of local neighborhoods. It systematically surfaces points of comparison by computing the similarity of the $k$-nearest neighbors of every embedded object between a pair of spaces. Through case studies across multiple modalities, we demonstrate our system rapidly reveals insights, such as semantic changes following fine-tuning, language changes over time, and differences between seemingly similar models. In evaluations with 15 participants, we find our system accelerates comparisons by shifting from laborious manual specification to browsing and manipulating visualizations.

cs.HC↗
arXiv2018

What made you do this? Understanding black-box decisions with sufficient input subsets

Local explanation frameworks aim to rationalize particular decisions made by a black-box prediction model. Existing techniques are often restricted to a specific type of predictor or based on input saliency, which may be undesirably sensitive to factors unrelated to the model's decision making process. We instead propose sufficient input subsets that identify minimal subsets of features whose observed values alone suffice for the same decision to be reached, even if all other input feature values are missing. General principles that globally govern a model's decision-making can also be revealed by searching for clusters of such input patterns across many data points. Our approach is conceptually straightforward, entirely model-agnostic, simply implemented using instance-wise backward selection, and able to produce more concise rationales than existing techniques. We demonstrate the utility of our interpretation method on various neural network models trained on text, image, and genomic data.

cs.LG↗
arXiv2012

The regular conducting fluid model for relativistic thermodynamics

The "regular" model presented here can be considered to be the most natural solution to the problem of constructing the simplest possible relativistic analogue of the category of classical Fourier--Euler thermally conducting fluid models as characterised by a pair of equations of state for just two dependent variables (an equilibrium density and a conducting scalar). The historically established but causally unsatisfactory solution to this problem due to Eckart is shown to be based on an ansatz that is interpretable as postulating a most unnatural relation between the (particle and entropy) velocities and their associated momenta, which accounts for the well known bad behaviour of that model which has recently been shown to have very pathological mixed-elliptic-hyperbolic comportments. The newer (and more elegant) solution of Landau and Lifshitz has a more mathematically respectable parabolic-hyperbolic comportment, but is still compatible with a well posed initial value problem only in such a restricted limit-case such as that of linearised perturbations of a static background. For mathematically acceptable behaviour undermore general circumstances, and a fortiori for the physically motivated requirement of subluminal signal propagation, only strictly hyperbolic behaviour is acceptable. Attention is drawn here to the availability of a more modern "regular" solution which, is fully satisfactory as far as all these requirements are concerned. This "regular" category of relativistic conducting fluid models arises naturally within a recently developed variational approach, in which the traditionally important stress--momentum-energy density tensor is relegated to a secondary role, while the relevant covariant 4-momentum co-vectors are instead brought to the fore.

gr-qc↗
arXiv2012

Classical Anthropic Everett model: indeterminacy in a preordained multiverse

Although ultimately motivated by quantum theoretical considerations, Everett's many-world idea remains valid, as an approximation, in the classical limit. However to be applicable it must in any case be applied in conjunction with an appropriate anthropic principle, whose precise formulation involves an anthropic quotient that can be normalised to unity for adult humans but that would be lower for infants and other animals. The outcome is a deterministic multiverse in which the only function of chance is the specification of one's particular identity.

gr-qc↗
arXiv2011

Classical dynamics of strings and branes, with application to vortons

These notes offer an introductory overview of the essentials of classical brane dynamics in a space-time background of arbitrary dimension, using a systematic geometric treatment emphasising the role of the second fundamental tensor and its trace, the curvature vector $K^\mu$. This approach is applied to the problem of stability of vorton equilibrium states of cosmic string loops in an ordinary 4-dimensional background.

hep-th↗
arXiv2010

Hominid evolution: genetics versus memetics

The last few million years on planet Earth have witnessed two remarkable phases of hominid development, starting with a phase of biological evolution characterised by rather rapid increase of the size of the brain. This has been followed by a phase of even more rapid technological evolution and concomitant expansion of the size of the population, that began when our own particular `sapiens' species emerged, just a few hundred thousand years ago. The present investigation exploits the analogy between the neo-Darwinian genetic evolution mechanism governing the first phase, and the memetic evolution mechanism governing the second phase. From the outset of the latter until very recently -- about the year 2000 -- the growth of the global population N was roughly governed by an equation of the form dN/Ndt= N/T*, in which T* is a coefficient introduced (in 1960) by von Foerster, who evaluated it empirically as about two hundred thousand million years. It is shown here how the value of this hitherto mysterious timescale governing the memetic phase is explicable in terms of what happenned in the preceding genetic phase. The outcome is that the order of magnitude of the Foerster timescale can be accounted for as the product of the relevant (human) generation timescale, about 20 years, with the number of bits of information in the genome, of the order of ten thousand million. Whereas the origin of our `homo' genus may well have involved an evolutionary hard step, it transpires that the emergence of our particular `sapiens' species was rather an automatic process.

q-bio.PE↗
arXiv2010

Fields in nonaffine bundles. IV. Harmonious non-Abelian currents in string defects

This article continues the study of the category of harmonious field models that was recently introduced as a kinetically non-linear generalisation of the well known harmonic category of multiscalar fields over a supporting brane wordsheet in a target space with a curved Riemannian metric. Like the perfectly harmonious case of which a familiar example is provided by ordinary barotropic perfect fluids, another important subcategory is the simply harmonious case, for which it is shown that as well as ``wiggle'' modes of the underlying brane world sheet, and sound type longitudinal modes, there will also be transverse shake modes that propagate at the speed of light. Models of this type are shown to arise from a non-Abelian generalisation of the Witten mechanism for conducting string formation by ordinary scalar fields with a suitable quartic self coupling term in the action.

hep-th↗
arXiv2009

Fields in nonaffine bundles. III. Effective symmetries and conserved currents in strings and higher branes

The principles of a previously developed formalism for the covariant treatment of multi-scalar fields for which (as in a nonlinear sigma model) the relevant target space is not of affine type -- but curved -- are recapitulated. Their application is extended from ordinary harmonic models to a more general category of "harmonious" field models, with emphasis on cases in which the field is confined to a string or higher brane worldsheet, and for which the relevant internal symmetry group is non Abelian, so that the conditions for conservation of the corresponding charge currents become rather delicate, particularly when the symmetry is gauged. Attention is also given to the conditions for conservation of currents of a different kind -- representing surface fluxes of generalised momentum or energy -- associated with symmetries not of the internal target space but of the underlying spacetime background structure, including the metric and any relevant gauge field. For the corresponding current to be conserved the latter need not be manifestly invariant: preservation modulo a gauge adjustment will suffice. The simplest case is that of "strong" symmetry, meaning invariance under the action of an effective Lie derivative (an appropriately gauge adjusted modification of an ordinary Lie derivative). When the effective symmetry is of the more general "weak" kind, the kinetic part of the current is not conserved by itself but only after being supplemented by a suitable contribution from the background.

hep-th↗
arXiv2009

Stability of winding cosmic wall lattices with X type junctions

This work confirms the stability of a class of domain wall lattice models that can produce accelerated cosmological expansion, with pressure to density ratio $w=-1/3$ at early times, and with $w=-2/3$ at late times when the lattice scale becomes large compared to the wall thickness. For walls of tension $T_{I}$, the relevant X type junctions could be unstable (for a sufficiently acute intersection angle $α$) against separation into a pair of Y type junctions joined by a compound wall, only if the tension $T_{II}$ of the latter were less than $2T_{I}$ (and for an approximately right-angled intersection if it were less that $\sqrt{2} T_{I}$) which can not occur in the class considered here. In an extensive category of multicomponent scalar field models of forced harmonic (linear or non-linear) type it is shown how the relevant tension -- which is the same as the surface energy density $U$ of the wall -- can be calculated as the minimum (geodesic) distance between the relevant vacuum states as measured on the space of field values $Φ^i$ using a positive definite (Riemannian) energy metric $dU^2=\tilde G_{ij} dΦ^i dΦ^j$ that is obtained from the usual kinetic metric (which is flat for a model with ordinary linear kinetic part) by application of a conformal factor proportional to the relevant potential function $V$. For suitably periodic potential functions there will be corresponding periodic configurations -- with parallel walls characterised by incrementation of a winding number -- in which the condition for stability of large scale bunching modes is shown to be satisfied automatically. It is suggested that such a configuration -- with a lattice lengthscale comparable to intergalactic separation distances -- might have been produced by a late stage of cosmological inflation.

hep-ph↗
arXiv2009

Fields In Nonaffine Bundles II. Gauge coupled generalization of harmonic mappings and their Bunting identities

The general purpose bitensorially gauge-covariant differentiation procedure set up in the preceding article is specialised to the particular case of bundles with nonlinear fibres that are endowed with a torsion free Riemannian or pseudo-Riemannian structure. This formalism is used to generalize the class of harmonic mappings between Riemannian or pseudo-Riemannian spaces to a natural gauge coupled extension in the form of a class of field sections of a bundle having the original image space as fibre, with a nonintegrable gauge connection $\Amr$ belonging to the algebra of the isometry group of the fibre space. The Bunting identity that can be used for establishing uniqueness in the strictly positive Riemannian case with negative image space curvature is shown to be generalizable to this gauge coupled extension.

hep-th↗
arXiv2009

Fields in Nonaffine Bundles. I. The general bitensorially covariant differentiation procedure

The standard covariant differentiation procedure for fields in vector bundles is generalised so as to be applicable to fields in general nonaffine bundles in which the fibres may have an arbitrary nonlinear structure. In addition to the usual requirement that the base space should be flat or endowed with its own linear connection, and that there should be an ordinary gauge connection on the bundle, it is necessary to require also that there should be an intrinsic, bundle-group invariant connection on the fibre space. The procedure is based on the use of an appropriate primary-field (i.e. section) independent connector that is constructed in terms of the natural fibre-tangent-vector realisation of the gauge connection. The application to gauged harmonic mappings will be described in a following article.

hep-th↗
arXiv2008

The logarithmic equation of state for superconducting cosmic strings

This investigation follows up the suggestion that the equation of state for superconducting cosmic strings provided by Witten's prototype biscalar field model can be well represented by an effective Lagrangian of simple logarithmic form depending on only 3 independent parameters. The numerical work described here confirms the validity of this approximation, and initiates the evaluation of the 3 required parameters, as functions of the masses and other parameters specifying the underlying U(1) x U(1) scalar field model in the limit for which the relevant gauge coupling constants are small.In this limit, subject to calibration of the relevant length and mass scales, the scalar field model is characterised by just 3 dimensionless ratios which (in order to provide conducting strings) must be subject to three inequalities (of which two have obvious analytic expressions). It is found here that when all three of these inequalities are satisfied by a reasonably large margin, there is a simple empirical formula that can be used to provide a fairly accurate prescription for the algebraic dependence on these 3 dimensionless ratios of the 3 parameters required for the logarithmic equation of state.

hep-th↗
arXiv2007

Five or six step scenario for evolution?

The prediction that (due to the limited amount of hydrogen available as fuel in the Sun) the future duration of our favourable terrestrial environment will be short (compared with the present age of the Earth) has been interpreted as evidence for a hard step scenario. This means that some of the essential steps (such as the development of eukaryotes) in the evolution process leading to the ultimate emergence of intelligent life would have been hard, in the sense of being against the odds in the available time, so that they are unlikely to have been achieved in most of the earth-like planets that may one day be discovered in nearby extra-solar systems. It was originally estimated that only one or two of the essential evolutionary steps had to have been hard in this sense, but it has become apparent that this figure may need upward revision, because recent studies of climatic instability suggest that the possible future duration of our biologically favourable environment may be shorter than had been supposed, only about one Giga year rather than five. On the basis of the statistical requirement of roughly equal spacing between hard steps, it is argued that the best fit with the fossil record is now obtainable by postulating the number of hard steps to be five, if our evolution was exclusively terrestrial, or six if, as now seems very plausible, the first step occurred on Mars.

astro-ph↗
arXiv2007

Objective and subjective time in anthropic reasoning

The original formulation of the (weak) anthropic principle was prompted by a question about objective time at a macroscopic level, namely the age of the universe when ``anthropic'' observers such as ourselves would be most likely to emerge. Theoretical interpretation of what one observes requires the theory to indicate what is expected, which will commonly depend on where, and particularly when, the observation can be expected to occur. In response to the question of where and when, the original version of the anthropic principle proposed an {it a priori} probability weighting proportional to the number of ``anthropic'' observers present. The present discussion takes up the question of the time unit characterising the biological clock controlling our subjective internal time, using a revised alternative to a line of argument due to Press, who postulated that animal size is limited by the brittleness of bone. On the basis of a static support condition depending on the tensile strength of flesh rather than bone, it is reasoned here that our size should be subject to a limit inversely proportional to the terrestrial gravitation field g, which is itself found to be proportional (with a factor given by the 5/2 power of the fine structure constant) to the gravitational coupling constant.This provides an animal size limit that will in all cases be of the order of a thousandth of the maximum mountain height, which will itself be of the order of a thousandth of the planetary radius. The upshot, via the (strong) anthropic principle, is that the need for brains, and therefore planets, that are large in terms of baryon number may be what explains the weakness of gravity relative to electromagnetism.

hep-th↗