This paper learns deep local patch descriptor (for replacing SIFT) by hard negative mining using current mini-batch. It outperforms SIFT and deep competitors on Oxford5K and Paris6K retrieval datasets.

Reports of rapid growth in nature-based tourism and recreation add significant weight to the economic case for biodiversity conservation but seem to contradict widely voiced concerns that people are becoming increasingly isolated from nature. This apparent paradox has been highlighted by a recent study showing that on a per capita basis, visits to natural areas in the United States and Japan have declined over the last two decades. These results have been cited as evidence of "a fundamental and pervasive shift away from nature-based recreation" - but how widespread is this phenomenon? We address this question by looking at temporal trends in visitor numbers at 280 protected areas (PAs) from 20 countries. This more geographically representative dataset shows that while PA visitation (whether measured as total or per capita visit numbers) is indeed declining in the United States and Japan, it is generally increasing elsewhere. Total visit numbers are growing in 15 of the 20 countries for which we could get data, with the median national rate of change unrelated to the national rate of population growth but negatively associated with wealth. Reasons for this reversal of growth in the richest countries are difficult to pin down with existing data, but the pattern is mirrored by trends in international tourist arrivals as a whole and so may not necessarily be caused by disaffection with nature. Irrespective of the explanation, it is clear that despite important downturns in some countries, nature-related tourism is far from declining everywhere, and may still have considerable potential both to generate funds for conservation and to shape people's attitudes to the environment.

Nature-based tourism is frequently described as one of the fastest growing sectors of the world's largest industry, and a very important justification for conservation. However, a recent, high profile report has interpreted declining visit rates to US and Japanese national parks as evidence of a pervasive shift away from nature tourism. Here we use the largest database so far compiled on trends in visits to Protected Areas around the world to resolve this apparent paradox. We find that, while visit rates—measured in two different ways—are indeed declining in some wealthy countries, in roughly three-quarters of the nations where data are available, visits to Protected Areas are increasing. Internationally, rates of growth in the number of visits to such areas show a clear negative association with per capita income, which interestingly is matched by trends in foreign arrivals as a whole. Our results therefore suggest that, despite worrying local downturns, nature-related tourism is far from declining everywhere, and may still have considerable potential to generate funds for conservation and engage people with the environment.

Updating the parameters in a LSTM language model based on the observed sequence during testing. A slice of text is first processed and then used for a gradient descent update step. A regularisation term is also proposed which draws the parameters back towards the original model.

https://i.imgur.com/zikOowE.png

## Introduction

A variable x is said to obey a power-law if it is drawn from a probability distribution function (pdf) of the form $p(x) = Cx^{-\alpha}$ where $C$ is called the **normalization constant** and $\alpha$ is called **scaling parameter** or exponent. Often, the power-law applies only for values greater than some minimum $x$, called $x_\text{min}$. The paper describes various statistical techniques to test if a given distribution follows a power-law or not.

Power-law distributions come in both continuous and discrete flavor with the discrete case being more involved than the continuous one. So, the discrete power-law behavior is often approximated by continuous power-law behavior for the sake of convenience. One reliable approximation is to assume that discrete values of $x$ are generated from a continuous power-law and then rounded to nearest integer to get the discrete values.

Sometimes, complementary cumulative distribution function (or CDF) is also considered where $P(X) = p(x ≥ X)$

## Fitting power-laws to empirical data

Power-law distribution makes a straight line on  the log-log plot. This slope can be calculated using the method of least square linear regression. But simple line fitting does not guarantee that data follows a power-law distribution. Moreover, the assumption of independent, Gaussian noise, which is a pre-requisite for linear regression, does not hold for this case.

### Estimating scaling parameter

Assuming that we know the value of $x_\text{min}$, the value of $\alpha$ can be obtained by the *method of maximum likelihood*. Maximum likelihood estimator (MLE) for continuous case is given as:

https://github.com/shagunsodhani/powerlaw/raw/master/paper/assets/MLEContinous.png

and that for the discrete case is given as:

https://github.com/shagunsodhani/powerlaw/raw/master/paper/assets/MLEDiscrete.png

The equation for the discrete case is only an approximation as there is no exact MLE for discrete case.

MLE method outperforms several linear regression based approaches like line fitting on the log-log plot, line fitting after performing logarithmic binning (done to reduce fluctuations in the tail of the distribution), line fitting to CDF with constant size bins and line fitting to CDF without any bins. But for any finite sample size $n$ and any choice of $x_\text{min}$, there is bias present which decays as $O(1/n)$ and can be ignored for $n ≥ 50$


### Estimating $x_\text{min}$

If we choose a value of $x_\text{min}$ less than the original value, then we will get a biased value of $\alpha$ as we will be fitting power-law to non power-law region as well. If we choose a value larger than the original value, we will be losing legitimate data points (leading to statistical errors). But it is more acceptable to make a higher estimate of $x_\text{min}$ than the original value.

One approach is to plot the PDF or CDF on the log-log plot and mark the point beyond which the distribution becomes roughly straight or to plot $\alpha$ as a function of $x_\text{min}$ and mark the point beyond which the value appears relatively stable. But these approaches are not objective as roughly straight and relatively stable are not quantified.

The approach proposed by Clauset et al. [Clauset, A., M. Young, and K. S. Gleditsch, 2007, Journal of Conflict Resolution 51, 58] is as follows:

Choose a value of $x_\text{min}$ such that the probability distribution of the measured data and best-fit power-law model are as similar as possible. Similarity between distributions can be measured using Kolmogorov-Smirnov (KS) statistic which is defined as:

https://github.com/shagunsodhani/powerlaw/raw/master/paper/assets/KSStatistics.png

where $S(x)$ is CDF of given data with values greater than or equal to $x_\text{min}$ and *P(x)* is the CDF of power-law model that best fits the data in the region $x ≥ x_\text{min}$. This method works well for both continuous and discrete data and is recommended for the general case. Other measures, like weighted KS or Kuiper statistics, can also be used in place of KS statistic.

## Testing the power-law hypothesis

MLE and other approaches do not tell us whether power-law is a possible fit to the given data - all they do is find the best fit values of $x_\text{min}$ and $\alpha$ assuming the data comes from a power-law distribution. A basic approach would be to calculate the value of $x_\text{min}$ and $\alpha$ and use them to hypothesize a power-law distribution from which the data is drawn. We then check the validity of this hypothesis using the goodness-of-fit tests.

### Goodness-of-fit tests

A large number of synthetic data sets are generated from the hypothesized power-law distribution. Then each of these distributions is fitted to their own power-law model individually and the KS statistics is calculated for each distribution. The *p-* value is defined to be the fraction of synthetic datasets where the distance (KS statistic value) is greater than the distance for given dataset. A large value of *p* (close to 1) means that the fluctuations between given data and the hypothesized model could be because of statistical fluctuations alone while a small value of *p* (close to 0) means that the model is not a possible fit to the distribution.

### Dataset generation

The generated dataset needs to be such that it has a distribution similar to the given data below $x_\min$ and follows the fitted power-law above $x_\text{min}$. Suppose the given data has *n<sub>tail</sub>* observations (where $x ≥ x_\text{min}$) and *n* observations in total. With a probability of $x_\text{min}/n$, a random number $x_i$ is generated from the hypothesized power-law distribution. With a probability of $1- n_{tail}/n$, a number is picked randomly from the given dataset with $x < x_\text{min}$. This way, the generated dataset of n elements is expected to follow powerlaw above $x ≥ x_\text{min}$ and same distribution as given data below $x_\text{min}$.

If we want the *p-* values to be accurate to within about *ε* of the true value, then we should generate at least *1/4 ε<sup>-2</sup>* synthetic data sets. 

The power law is ruled out if *p ≤ 0.1*. A large *p-* value does not mean that the power-law is the correct distribution for the data. There can be other distributions that can fit the data equally well or even better. Moreover, for small values of n, it is possible that the given distribution will follow a power law closely, and hence that the p-value will be large, even when the power law is the wrong model for the data.


## Alternate distributions

*p-* value test can only be used to reject the power-law hypothesis and not accept it. So even if *p-value > 0.1*, we can only say that power-law hypothesis is not rejected. It could be the case that some other distribution fits the data equally well or even better. To eliminate this possibility, we calculate a *p-* value for a fit to the alternate distribution and compare it with the *p-* value for the power-law. If the *p-* value for power-law is high and the *p-* value for the other distribution is low, we can say that data is more likely to be drawn from the power-law distribution (though we still can not be sure that it is **definitely** drawn from the power-law distribution).

### Likelihood Ratio Test

This test can be used to directly compare two distributions against one another to see which is a better fit for the given data. The idea is to compute the likelihood of the given data under the two competing distributions. The one with the higher likelihood is taken to be the better fit. Alternatively, the ratio of the two likelihoods, or the logarithm *R* of the ratio can be used. If *R* is close enough to zero, then it could go to either side of zero, depending on statistical fluctuations. So *R* value needs to be sufficiently far from zero. To check for this, Vuong's method [Vuong, Q. H., 1989, Econometrica 57, 307] is used which gives a *p-* value that can tell if the conclusion from the value of *R* is statistically significant. If this *p-* value is small (*p < 0.1*), the result is significant. Otherwise, the result is not taken to be reliable and the test does not favor either distribution. 

Other than the likelihood ratio, several other tests like minimum description length (MDL) or cross-validation can also be performed.

The goal is to improve the training process for a spoken dialogue system, more specifically a telephone-based system providing restaurant information for the Cambridge (UK) area. They train a supervised system which tries to predict the success on the current dialogue – if the model is certain about the outcome, the predicted label is used for training the dialogue system; if the model is uncertain, the user is asked to provide a label. Essentially it reduces the amount of annotation that is required, by choosing which examples should be annotated through active learning.

https://i.imgur.com/dWY1EdE.png

The dialogue is mapped to a vector representation using a bidirectional LSTM trained like an autoencoder, and a Gaussian Process is used for modelling dialogue success.