TY - JOUR
AB - When a mutation with selective advantage s spreads through a panmictic population, it may cause two lineages at a linked locus to coalesce; the probability of coalescence is exp(−2rT), where T∼log(2Ns)/s is the time to fixation, N is the number of haploid individuals, and r is the recombination rate. Population structure delays fixation, and so weakens the effect of a selective sweep. However, favourable alleles spread through a spatially continuous population behind a narrow wavefront; ancestral lineages are confined at the tip of this front, and so coalesce rapidly. In extremely dense populations, coalescence is dominated by rare fluctuations ahead of the front. However, we show that for moderate densities, a simple quasi-deterministic approximation applies: the rate of coalescence within the front is λ∼2g(η)/(ρℓ), where ρ is the population density and is the characteristic scale of the wavefront; g(η) depends only on the strength of random drift, . The net effect of a sweep on coalescence also depends crucially on whether two lineages are ever both within the wavefront at the same time: even in the extreme case when coalescence within the front is instantaneous, the net rate of coalescence may be lower than in a single panmictic population. Sweeps can also have a substantial impact on the rate of gene flow. A single lineage will jump to a new location when it is hit by a sweep, with mean square displacement ; this can be substantial if the species’ range, L, is large, even if the species-wide rate of sweeps per map length, Λ/R, is small. This effect is half as strong in two dimensions. In contrast, the rate of coalescence between lineages, at random locations in space and on the genetic map, is proportional to (c/L)(Λ/R), where c is the wavespeed: thus, on average, one-dimensional structure is likely to reduce coalescence due to sweeps, relative to panmixis. In two dimensions, genes must move along the front before they can coalesce; this process is rapid, being dominated by rare fluctuations. This leads to a dramatically higher rate of coalescence within the wavefront than if lineages simply diffused along the front. Nevertheless, the net rate of coalescence due to a sweep through a two-dimensional population is likely to be lower than it would be with panmixis.
AU - Barton, Nicholas H
AU - Etheridge, Alison
AU - Kelleher, Jerome
AU - Véber, Amandine
ID - 2473
IS - 8
JF - Theoretical Population Biology
TI - Genetic hitch-hiking in spatially extended populations
VL - 87
ER -
TY - JOUR
AB - We study the problem of object recognition for categories for which we have no training examples, a task also called zero-data or zero-shot learning. This situation has hardly been studied in computer vision research, even though it occurs frequently: the world contains tens of thousands of different object classes and for only few of them image collections have been formed and suitably annotated. To tackle the problem we introduce attribute-based classification: objects are identified based on a high-level description that is phrased in terms of semantic attributes, such as the object's color or shape. Because the identification of each such property transcends the specific learning task at hand, the attribute classifiers can be pre-learned independently, e.g. from existing image datasets unrelated to the current task. Afterwards, new classes can be detected based on their attribute representation, without the need for a new training phase. In this paper we also introduce a new dataset, Animals with Attributes, of over 30,000 images of 50 animal classes, annotated with 85 semantic attributes. Extensive experiments on this and two more datasets show that attribute-based classification indeed is able to categorize images without access to any training images of the target classes.
AU - Lampert, Christoph
AU - Nickisch, Hannes
AU - Harmeling, Stefan
ID - 2516
IS - 3
JF - IEEE Transactions on Pattern Analysis and Machine Intelligence
TI - Attribute-based classification for zero-shot learning of object categories
VL - 36
ER -
TY - CONF
AB - Traditional formal methods are based on a Boolean satisfaction notion: a reactive system satisfies, or not, a given specification. We generalize formal methods to also address the quality of systems. As an adequate specification formalism we introduce the linear temporal logic LTL[F]. The satisfaction value of an LTL[F] formula is a number between 0 and 1, describing the quality of the satisfaction. The logic generalizes traditional LTL by augmenting it with a (parameterized) set F of arbitrary functions over the interval [0,1]. For example, F may contain the maximum or minimum between the satisfaction values of subformulas, their product, and their average. The classical decision problems in formal methods, such as satisfiability, model checking, and synthesis, are generalized to search and optimization problems in the quantitative setting. For example, model checking asks for the quality in which a specification is satisfied, and synthesis returns a system satisfying the specification with the highest quality. Reasoning about quality gives rise to other natural questions, like the distance between specifications. We formalize these basic questions and study them for LTL[F]. By extending the automata-theoretic approach for LTL to a setting that takes quality into an account, we are able to solve the above problems and show that reasoning about LTL[F] has roughly the same complexity as reasoning about traditional LTL.
AU - Almagor, Shaull
AU - Boker, Udi
AU - Kupferman, Orna
ID - 2517
IS - Part 2
TI - Formalizing and reasoning about quality
VL - 7966
ER -
TY - CONF
AB - A class of valued constraint satisfaction problems (VCSPs) is characterised by a valued constraint language, a fixed set of cost functions on a finite domain. An instance of the problem is specified by a sum of cost functions from the language with the goal to minimise the sum. We study which classes of finite-valued languages can be solved exactly by the basic linear programming relaxation (BLP). Thapper and Živný showed [20] that if BLP solves the language then the language admits a binary commutative fractional polymorphism. We prove that the converse is also true. This leads to a necessary and a sufficient condition which can be checked in polynomial time for a given language. In contrast, the previous necessary and sufficient condition due to [20] involved infinitely many inequalities. More recently, Thapper and Živný [21] showed (using, in particular, a technique introduced in this paper) that core languages that do not satisfy our condition are NP-hard. Taken together, these results imply that a finite-valued language can either be solved using Linear Programming or is NP-hard.
AU - Kolmogorov, Vladimir
ID - 2518
IS - 1
TI - The power of linear programming for finite-valued CSPs: A constructive characterization
VL - 7965
ER -
TY - CONF
AB - We propose a probabilistic model to infer supervised latent variables in
the Hamming space from observed data. Our model allows simultaneous
inference of the number of binary latent variables, and their values. The
latent variables preserve neighbourhood structure of the data in a sense
that objects in the same semantic concept have similar latent values, and
objects in different concepts have dissimilar latent values. We formulate
the supervised infinite latent variable problem based on an intuitive
principle of pulling objects together if they are of the same type, and
pushing them apart if they are not. We then combine this principle with a
flexible Indian Buffet Process prior on the latent variables. We show that
the inferred supervised latent variables can be directly used to perform a
nearest neighbour search for the purpose of retrieval. We introduce a new
application of dynamically extending hash codes, and show how to
effectively couple the structure of the hash codes with continuously
growing structure of the neighbourhood preserving infinite latent feature
space.
AU - Quadrianto, Novi
AU - Sharmanska, Viktoriia
AU - Knowles, David
AU - Ghahramani, Zoubin
ID - 2520
SN - 9780974903996
T2 - Proceedings of the 29th conference uncertainty in Artificial Intelligence
TI - The supervised IBP: Neighbourhood preserving infinite latent feature models
ER -
TY - JOUR
AB - We consider non-interacting particles subject to a fixed external potential V and a self-generated magnetic field B. The total energy includes the field energy β∫B2 and we minimize over all particle states and magnetic fields. In the case of spin-1/2 particles this minimization leads to the coupled Maxwell-Pauli system. The parameter β tunes the coupling strength between the field and the particles and it effectively determines the strength of the field. We investigate the stability and the semiclassical asymptotics, h→0, of the total ground state energy E(β,h,V). The relevant parameter measuring the field strength in the semiclassical limit is κ=βh. We are not able to give the exact leading order semiclassical asymptotics uniformly in κ or even for fixed κ. We do however give upper and lower bounds on E with almost matching dependence on κ. In the simultaneous limit h→0 and κ→∞ we show that the standard non-magnetic Weyl asymptotics holds. The same result also holds for the spinless case, i.e. where the Pauli operator is replaced by the Schrödinger operator.
AU - Erdös, László
AU - Fournais, Søren
AU - Solovej, Jan
ID - 2698
IS - 6
JF - Journal of the European Mathematical Society
TI - Stability and semiclassics in self-generated fields
VL - 15
ER -
TY - CONF
AB - Even though both population and quantitative genetics, and evolutionary computation, deal with the same questions, they have developed largely independently of each other. I review key results from each field, emphasising those that apply independently of the (usually unknown) relation between genotype and phenotype. The infinitesimal model provides a simple framework for predicting the response of complex traits to selection, which in biology has proved remarkably successful. This allows one to choose the schedule of population sizes and selection intensities that will maximise the response to selection, given that the total number of individuals realised, C = ∑t Nt, is constrained. This argument shows that for an additive trait (i.e., determined by the sum of effects of the genes), the optimum population size and the maximum possible response (i.e., the total change in trait mean) are both proportional to √C.
AU - Barton, Nicholas H
AU - Paixao, Tiago
ID - 2718
T2 - Proceedings of the 15th annual conference on Genetic and evolutionary computation
TI - Can quantitative and population genetics help us understand evolutionary computation?
ER -
TY - CONF
AB - Prediction of the evolutionary process is a long standing problem both in the theory of evolutionary biology and evolutionary computation (EC). It has long been realized that heritable variation is crucial to both the response to selection and the success of genetic algorithms. However, not all variation contributes in the same way to the response. Quantitative genetics has developed a large body of work trying to estimate and understand how different components of the variance in fitness in the population contribute to the response to selection. We illustrate how to apply some concepts of quantitative genetics to the analysis of genetic algorithms. In particular, we derive estimates for the short term prediction of the response to selection and we use variance decomposition to gain insight on local aspects of the landscape. Finally, we propose a new population based genetic algorithm that uses these methods to improve its operation.
AU - Paixao, Tiago
AU - Barton, Nicholas H
ID - 2719
T2 - Proceedings of the 15th annual conference on Genetic and evolutionary computation
TI - A variance decomposition approach to the analysis of genetic algorithms
ER -
TY - JOUR
AB - Knowledge of the rate and fitness effects of mutations is essential for understanding the process of evolution. Mutations are inherently difficult to study because they are rare and are frequently eliminated by natural selection. In the ciliate Tetrahymena thermophila, mutations can accumulate in the germline genome without being exposed to selection. We have conducted a mutation accumulation (MA) experiment in this species. Assuming that all mutations are deleterious and have the same effect, we estimate that the deleterious mutation rate per haploid germline genome per generation is U = 0.0047 (95% credible interval: 0.0015, 0.0125), and that germline mutations decrease fitness by s = 11% when expressed in a homozygous state (95% CI: 4.4%, 27%). We also estimate that deleterious mutations are partially recessive on average (h = 0.26; 95% CI: –0.022, 0.62) and that the rate of lethal mutations is <10% of the deleterious mutation rate. Comparisons between the observed evolutionary responses in the germline and somatic genomes and the results from individual-based simulations of MA suggest that the two genomes have similar mutational parameters. These are the first estimates of the deleterious mutation rate and fitness effects from the eukaryotic supergroup Chromalveolata and are within the range of those of other eukaryotes.
AU - Long, Hongan
AU - Paixao, Tiago
AU - Azevedo, Ricardo
AU - Zufall, Rebecca
ID - 2720
IS - 2
JF - Genetics
TI - Accumulation of spontaneous mutations in the ciliate Tetrahymena thermophila
VL - 195
ER -
TY - JOUR
AB - We consider random n×n matrices of the form (XX*+YY*)^{-1/2}YY*(XX*+YY*)^{-1/2}, where X and Y have independent entries with zero mean and variance one. These matrices are the natural generalization of the Gaussian case, which are known as MANOVA matrices and which have joint eigenvalue density given by the third classical ensemble, the Jacobi ensemble. We show that, away from the spectral edge, the eigenvalue density converges to the limiting density of the Jacobi ensemble even on the shortest possible scales of order 1/n (up to log n factors). This result is the analogue of the local Wigner semicircle law and the local Marchenko-Pastur law for general MANOVA matrices.
AU - Erdös, László
AU - Farrell, Brendan
ID - 2782
IS - 6
JF - Journal of Statistical Physics
TI - Local eigenvalue density for general MANOVA matrices
VL - 152
ER -