A few weeks ago, we told you about the different types of machine learning through a series of examples, and we analysed how to choose one or the other based on our objectives and the datasets available to train the algorithm.

Now let's assume that we have an already labelled dataset and we need to train a supervised learning model to solve the task at hand. At this point, we need some mechanism to tell us whether the model has learned correctly or not. That is what we are going to discuss in this post, the most used metrics to evaluate the quality of our models.

Model evaluation is a very important step in the development methodology of machine learning systems. It helps to measure the performance of the model, that is, to quantify the quality of the predictions it offers. To do this, we use evaluation metrics, which depend on the learning task we apply. As we saw in the previous post, within supervised learning there are two types of tasks that differ, mainly, in the type of output they offer:

• Classification tasks, which produce as output a discrete label, i.e. when the output is one within a finite set.
• Regression tasks, which output a continuous real value.

Here are some of the most commonly used metrics to assess the performance of both types of tasks:

Evaluation of classification models

In order to better understand these metrics, we will use as an example the predictions of a classification model to detect COVID patients. In the following table we can see in the first column the example identifier, in the second column the class predicted by the model, in the third column the actual class and the fourth column indicates whether the model has failed in its prediction or not. In this case, the positive class is "Yes" and the negative class is "No".

Examples of evaluation metrics for classification model include the following:

• Confusion matrix: this is a widely used tool that allows us to visually inspect and evaluate the predictions of our model. Each row represents the number of predictions of each class and the columns represent the instances of the actual class.

The description of each element of the matrix is as follows:

True Positive (VP): number of positive examples that the model predicts as positive. In the example above, VP is 1 (from example 6).

False positive (FP): number of negative examples that the model predicts as positive. In our example, FP is equal to 1 (from example 4).

False negative (FN): number of positive examples that the model predicts as negative. FN in the example would be 0.

True negative (VN): number of negative examples that the model predicts as negative. In the example, VN is 8.

• Accuracy: the fraction of predictions that the model made correctly.  It is represented as a percentage or a value between 0 and 1. It is a good metric when we have a balanced dataset, that is, when the number of labels of each class is similar. The accuracy of our example model is 0.9, since it got 9 predictions out of 10 correct. If our model had always predicted the "No" label, the accuracy would be 0.9 as well, but it does not solve our problem of identifying COVID patients.
• Recall: indicates the proportion of positive examples that are correctly identified by the model out of all actual positives. That is, VP / (VP + FN). In our example, the sensitivity value would be 1 / (1 + 0) = 1. If we were to evaluate with this metric, a model that always predicts the positive label ("Yes") would have a sensitivity of 1, but it would not be a very intelligent model. Although the ideal for our COVID detection model is to maximise sensitivity, this metric alone does not ensure that we have a good model.
• Precision: this metric is determined by the fraction of items correctly classified as positive among all items that the model has classified as positive. The formula is VP / (VP + FP). The example model would have an accuracy of 1 / (1 + 1) = 0.5. Let us now return to the model that always predicts the positive label. In that case, the accuracy of the model is 1 / (1 + 9) = 0.1. We see how this model had a maximum sensitivity, but has a very poor accuracy. In this case we need both metrics to evaluate the real quality of the model.
• F1 score: combines the Precision and Recall metrics to give a single score. This metric is the most appropriate when we have unbalanced datasets. It is calculated as the harmonic mean of Precision and Recall. The formula is F1 = (2 * precision * recall) / (precision + recall). You may wonder why we use the harmonic mean and not the simple mean. This is because the harmonic mean means that if one of the two measurements is small (even if the other is maximum), the value of F1 score is going to be small.

Evaluation of regression models

Unlike classification models, in regression models it is almost impossible to predict the exact value, but rather to be as close as possible to the real value, so most metrics, with subtle differences between them, are going to focus on measuring that: how close (or far) the predictions are from the real values.

In this case, we have as an example the predictions of a model that determines the price of watches depending on their characteristics. In the table we show the price predicted by the model, the actual price, the absolute error and the squared error.

Some of the most common evaluation metrics for regression models are:

• Mean Absolute Error: This is the mean of the absolute differences between the target and predicted values. Since it is not squared, it does not penalise large errors, which makes it not very sensitive to outliers, so it is not a recommended metric in models where attention must be paid to outliers. This metric also represents the error on the same scale as the actual values. Ideally, its value should be close to zero. For our watch pricing model, the mean absolute error is 107.8.
• Mean Squared Errors: One of the most commonly used measures in regression work. It is simply the mean of the differences between the target and the predicted value squared. By squaring the errors, it magnifies large errors, so use it with care when we have outliers in our data set. It can take values between 0 and infinity. The closer the metric is to zero, the better. The mean square error of the example model is 50113.8. We see how in the case of our example large errors are magnified.
• Root Mean Squared Srror: This is equal to the square root of the previous metric. The advantage of this metric is that it presents the error in the same units as the target variable, which makes it easier to understand. For our model this error is equal to 223.86.
• R-squared: also called the coefficient of determination. This metric differs from the previous ones, as it compares our model with a basic model that always returns as prediction the mean of the training target values. The comparison between these two models is made on the basis of the mean squared errors of each model. The values this metric can take range from minus infinity to 1. The closer the value of this metric is to 1, the better our model is. The R-squared value for the model will be 0.455.
• Adjusted R-squared. An improvement of R-squared. The problem with the previous metric is that every time more independent variables (or predictor variables) are added to the model, R-squared stays the same or improves, but never gets worse, which can be confusing, because just because one model uses more predictor variables than another, it does not mean that it is better. Adjusted R-squared compensates for the addition of independent variables. The adjusted R-squared value will always be less than or equal to the R-squared value, but this metric will show improvement when the model is actually better. For this measure we cannot do the calculation for our example model because, as we have seen before, it depends on the number of examples and the number of variables used to train such a model.

Conclusion

When working with supervised learning algorithms it is very important to choose a correct evaluation metric for our model. For classification models it is very important to pay attention to the dataset and check whether it is balanced or not. For regression models we have to consider outliers and whether we want to penalise large errors or not.

Generally, however, the business domain will guide us in the right choice of metric. For a disease detection model, such as the one we have seen, we are interested in high sensitivity, but we are also interested in a good accuracy value, so F1-score would be a smart choice. On the other hand, in a model to predict the demand for a product (and therefore production), where overstocking may incur a storage cost overrun, it may be a good idea to use the mean squared errors to penalise large errors.

Content elaborated by Jose Antonio Sanchez, expert in Data Science and enthusiast of the Artificial Intelligence.

Contents and points of view expressed in this publication are the exclusive responsibility of its author.

2020 is coming to an end and in this unusual year we are going to have to experience a different, calmer Christmas with our closest nucleus. What better way to enjoy those moments of calm than to train and improve your knowledge of data and new technologies?

Whether you are looking for a reading that will make you improve your professional profile to which to dedicate your free time on these special dates, or if you want to offer your loved ones an educational and interesting gift, from datos.gob.es we want to propose some book recommendations on data and disruptive technologies that we hope will be of interest to you. We have selected books in Spanish and English, so that you can also put your knowledge of this language into practice.

Take note because you still have time to include one in your letter to Santa Claus!

INTELIGENCIA ARTIFICIAL, naturalmente. Nuria Oliver, ONTSI, red.es (2020)

What is it about?: This book is the first of the new collection published by the ONTSI called “Pensamiento para la sociedad digital”. Its pages offer a brief journey through the history of artificial intelligence, describing its impact today and addressing the challenges it presents from various points of view.

Who is it for?: It is aimed especially at decision makers, professionals from the public and private sector, university professors and students, third sector organizations, researchers and the media, but it is also a good option for readers who want to introduce themselves and get closer to the complex world of artificial intelligence.

Artificial Intelligence: A Modern Approach, Stuart Russell

What is it about?:  Interesting manual that introduces the reader to the field of Artificial Intelligence through an orderly structure and understandable writing.

Who is it for?: This textbook is a good option to use as documentation and reference in different courses and studies in Artificial Intelligence at different levels. For those who want to become experts in the field.

Situating Open Data: Global Trends in Local Contexts, Danny Lämmerhirt, Ana Brandusescu, Natalia Domagala – African Minds (October 2020)

What is it about?: This book provides several empirical accounts of open data practices, the local implementation of global initiatives, and the development of new open data ecosystems.

Who is it for?: It will be of great interest to researchers and advocates of open data and to those in or advising government administrations in the design and implementation of effective open data initiatives. You can download its PDF version through this link.

The Elements of Statistical Learning: Data Mining, Inference, and Prediction, Second Edition (Springer Series in Statistics), Trevor Hustle, Jerome Friedman. – Springer (May 2017)

What is it about?: This book describes various statistical concepts in a variety of fields such as medicine, biology, finance, and marketing in a common conceptual framework. While the focus is statistical, the emphasis is on definitions rather than mathematics.

Who is it for?: It is a valuable resource for statisticians and anyone interested in data mining in science or industry. You can also download its digital version here.

Europa frente a EEUU y China: Prevenir el declive en la era de la inteligencia artificial, Luis Moreno, Andrés Pedreño – Kdp (2020)

What is it about?: This interesting book addresses the reasons for the European delay with respect to the power that the US and China do have, and its consequences, but above all it proposes solutions to the problem that is exposed in the work.

Who is it for?: It is a reflection for those interested in thinking about the change that Europe would need, in the words of its author, "increasingly removed from the revolution imposed by the new technological paradigm".

Armas de destrucción matemática: Cómo el big data aumenta la desigualdad y amenaza la democracia. Cathy O’Neil, Capitán Swing (2018)

What is it about?:  This book calls attention to the problems that can lead to the misuse of algorithms and proposes some ideas to avoid making mistakes.

Who is it for?: These pages do not appear overly technical concepts, nor are there formulas or complex explanations, although they do deal with dense problems that need the author's attention.

Data Feminism (Strong Ideas), Catherine D’Ignazio, Lauren F. Klein. MIT Press (2020)

What is it about?: These pages address a new way of thinking about data science and its ethics based on the ideas of feminist thought.

Who is it for?: To all those who are interested in reflecting on the biases built into the algorithms of the digital tools that we use in all areas of life.

Open Cities | Open Data: Collaborative Cities in the Information, Scott Hawken, Hoon Han, Chris Pettit – Palgrave Macmillan, Singapore (2020)

What is it about?: This book explains the importance of opening data in cities through a variety of critical perspectives, and presents strategies, tools, and use cases that facilitate both data openness and reuse..

Who is it for?: Perfect for those integrated in the data value chain in cities and those who have to develop open data strategies within the framework of a smart city, but also for citizens concerned about privacy and who want to know what happens - and what can happen- with the data generated by cities.

Although we would love to include them all on this list, there are many interesting books on data and technology that fill the shelves of hundreds of bookstores and online stores. If you have any extra recommendations that you want to make us, do not hesitate to leave us your favorite title in comments. The members of the datos.gob.es team will be delighted to read your recommendations this Christmas.

Spain already has a new National Artificial Intelligence Strategy. The document, which includes 600 million euros for measures related to artificial intelligence (AI), was presented on December 2 at the Palacio de la Moncloa.

The National Strategy for Artificial Intelligence (known as ENIA) is component 16 of the Plan for the Recovery, Transformation and Resilience of the Spanish economy, and one of the fundamental proposals of the Digital Spain Agenda 2025 in its line 9 of action, which highlights AI as a key element for boosting the growth of our economy in the coming years. In addition, the new strategy is aligned with the European action plans developed in this area, and especially with the White Paper on Artificial Intelligence.

Objectives and lines of action

The ENIA is a dynamic and flexible framework, open to the contribution of companies, citizens, social agents and the rest of the administrations, which was created with 7 objectives: scientific excellence and innovation, the projection of the Spanish language, the creation of qualified employment, the transformation of the Spanish productive fabric, the creation of an environment of trust in relation to AI and the promotion of an inclusive and sustainable AI that takes into account humanist values.

To achieve these objectives, 6 lines of action have been created, which bring together a total of 30 measures to be developed in the period 2020-2025:

In short, the aim is to create a national ecosystem of innovative, competitive and ethical artificial intelligence. And to do this, it is essential to have large volumes of quality and interoperable data and metadata, which are accessible, complete, secure and respectful of privacy.

Open data in the National Strategy of Artificial Intelligence

The availability of open data is essential for the proper functioning of artificial intelligence, since the algorithms must be fed and trained by data whose quality and availability allows continuous improvement. In this way we can create value services that impact on the improvement of society and the economy.

The National Strategy for Artificial Intelligence highlights how, thanks to the various initiatives undertaken in recent years, Spain has become a European benchmark for open data, highlighting the role of the Aporta Initiative in promoting the openness and reuse of public information.

In strategic axis 3 of the document, several key areas are highlighted where to act linked to AI data platforms and technological infrastructures:

• Developing the regulatory framework for open data, to define a strategy for publication and access to public data from administrations in multilingual formats, and to ensure the correct and safe use of the data.
• Promote actions in the field of data platforms, models, algorithms, inference engines and cyber security, with the focus on boosting research and innovation. Reference is made to the need to promote Digital Enabling Technologies such as connectivity infrastructures, massive data environments (cloud) or process automation and control, paying special attention to Strategic Supercomputing Capabilities (HPC).
• Promote the specific development of AI technologies in the field of natural language processing, promoting the use of Spanish in the world. In this sense, the National Plan of Language Technologies will be promoted and the LEIA project, developed by the Royal Spanish Academy for the defense, projection and good use of the Spanish language in the digital universe, will be supported.

In the specific case of open data, one of the first measures highlighted is the creation of the Data Office at the state level that will coordinate all public administrations in order to homogenize the storage, access and processing of data. To strengthen this action, a Chief Data Officer will be appointed.  In addition, a multidisciplinary working group on open data in the state public sector will be set up to highlight the efforts that have been made in the field of data in Spain and to continue promoting the openness and reuse of public sector information.

The strategy also considers the private sector, and highlights the need to promote the development of accessible repositories and to guide companies in the definition of open or shared data strategies. In this sense, shared spaces of sectorial and industrial data will be created, which will facilitate the creation of AI applications. Furthermore, mention is made of the need to offer data disaggregated by sex, age, nationality and territory, in such a way as to eliminate biases linked to these aspects.

In order to stimulate the use and governance of public and citizen data, the creation of the Data for Social Welfare Project is established as an objective, where open and citizen-generated data will play a key role in promoting accountability and public participation in government.

Other highlights of the ENIA

In addition to actions related to open data, the National Strategy of Artificial Intelligence includes more transversal actions, for example:

• The incorporation of AI in the public administration will be promoted, improving from transparency and effective decision-making to productivity and quality of service (making management and the relationship with citizens more efficient). Here the Aporta Initiative has been playing a key role with its support to public sector bodies in the publication of quality data and the promotion of its reuse. Open data repositories will be created to allow optimal access to the information needed to develop new services and applications for the public and private sectors. In this sense, an innovation laboratory (GobTechLab) will be created and training programs will be carried out.
• The aim is to promote scientific research through the creation of a Spanish Network of Excellence in AI with research and training programs and the setting up of new technological development centers. Special attention will be given to closing the gender gap.
• A program of aid to companies for the development of AI and data solutions will be launched, and the network of Digital innovation Hubs will be reinforced. A NextTech Fund for public-private venture capital will be created.
• Talent will be promoted through the National Digital Skills Plan. AI-related elements will be introduced in schools and the university and professional AI training offer will be boosted. The SpAIn Talent Hub program will be developed in coordination with ICEX Invest to attract foreign investment and talent.
• A novelty of the strategy is that it takes into account ethical and social regulation to fight discrimination. Observatories for ethical and legal evaluation of algorithmic systems will be created and the Digital Rights Charter, currently under revision, will be developed.

In short, we are facing a necessary strategy to boost the growth of AI in Spain, promoting our society and economy, and improving our international competitiveness.

Can you imagine an AI capable of writing songs, novels, press releases, interviews, essays, technical manuals, programming code, prescribing medication and much more that we don't know yet? Watching GPT-3 in action doesn't seem like we're very far away.

In our latest report on natural language processing (NLP) we mentioned the GPT-2 algorithm developed by OpenAI (the company founded by such well-known names as Elon Musk) as an exponent of its capabilities for generating synthetic text with a quality indistinguishable from any other human-created text. The surprising results of GPT-2 led the company not to publish the source code of the algorithm because of its potential negative effects on the generation of deepfakes or false news

Recently (May 2020) a new version of the algorithm has been released, now called GPT-3, which includes functional innovations and improvements in performance and capacity to analyse and generate natural language texts.

In this post we try to summarize in a simple and affordable way the main new features of GPT-3. Do you dare to discover them?

We start directly, getting to the point. What does GPT-3 bring with it? (adaptation of the original post by Mayor Mundada).

1. It is much bigger (complex) than everything we had before. Deep learning models based on neural networks are usually classified by their number of parameters. The greater the number of parameters, the greater the depth of the network and therefore its complexity. The training of the full version of GPT-2 resulted in 1.5 billion parameters. GPT-3 results in 175 billion parameters. GPT-3 has been trained on a basis of 570 GB of text compared to 40 GB of GPT-2.
2. For the first time, it can be used as a product or service. For the first time it can be used as a product or service. That is, OpenAI has announced the availability of a public API for users to experiment with the algorithm. At the time of writing this post, access to the API is restricted (this is what we call a private preview) and access must be requested.
3. The most important thing: the results. Despite the fact that the API is restricted by invitation, many Internet users (with access to the API) have published articles on its results in different fields.

Content elaborated by Alejandro Alija, expert in Digital Transformation and Innovation.

Contents and points of view expressed in this publication are the exclusive responsibility of its author.

Content prepared by Julián Valero, professor at the University of Murcia and Coordinator of the Research Group "Innovation, Law and Technology" (iDerTec).

Contents and points of view expressed in this publication are the exclusive responsibility of its author.

Machine learning is a branch within the field of Artificial Intelligence that provides systems with the ability to learn and improve automatically, based on experience. These systems transform data into information, and with this information, they can make decisions. A model needs to be fed with data to make predictions in a robust way. The more, the better. Fortunately, today's network is full of data sources. In many cases, data is collected by private companies for their own benefit, but there are also other initiatives, such as open data portals.

Once we have the data, we are ready to start the learning process. This process, carried out by an algorithm, tries to analyze and explore the data in search of hidden patterns. The result of this learning, sometimes, is nothing more than a function that operates on the data to calculate a certain prediction. 

In this article we will see the types of machine learning that exist including some examples.

Types of machine learning

Depending on the data available and the task we want to tackle, we can choose between different types of learning. These are: supervised learning, unsupervised learning, semi-supervised learning and reinforcement learning.

Supervised learning

Supervised learning needs labeled datasets, that is, we tell the model what we want it to learn. Suppose we have an ice cream shop and for the last few years we have been recording daily weather data, temperature, month, day of the week, etc., and we have also been doing the same with the number of ice creams sold each day. In this case, we would surely be interested in training a model that, based on the climatological data, temperature, etc. (characteristics of the model) of a specific day, tells us how many ice creams are going to be sold (the label to be predicted). 

Depending on the type of label, within the supervised learning there are two types of models:

• Classification models, which produce as output a discrete label, that is, a label within a finite set of possible labels. In turn, classification models can be binary if we have to predict between two classes or labels (disease or not disease, classification of emails as "spam" or not "spam") or multiclass, when we have to classify more than two classes (classification of animal images, sentiment analysis, etc.). 
• The regression models produce as output a real value, like the example we mentioned of the ice cream. 

Unsupervised learning

Unsupervised learning, on the other hand, works with data that has not been labeled. We do not have a label to predict. These algorithms are mainly used in tasks where it is necessary to analyze the data to extract new knowledge or group entities by affinity. 

This type of learning also has applications for reducing dimensionality or simplifying datasets. In the case of grouping data by affinity, the algorithm must define a similarity or distance metric that serves to compare the data with each other. As an example of unsupervised learning we have the clustering algorithms, which could be applied to find customers with similar characteristics to those who offer certain products or target a marketing campaign, discovery of topics or detection of anomalies, among others.  On the other hand, sometimes some datasets such as those related to genomic information have large amounts of characteristics and for various reasons, such as reducing the training time of the algorithms, improving the performance of the model or facilitating the visual representation of the data, we need to reduce the dimensionality or number of columns in the dataset. Dimensionality reduction algorithms use mathematical and statistical techniques to convert the original dataset into a new one with fewer dimensions in exchange for losing some information. Examples of dimensionality reduction algorithms are PCA, t-SNE or ICA.

Semi-supervised learning

Sometimes it is very complicated to have a fully labeled dataset. Let's imagine that we are the owners of a dairy product manufacturing company and we want to study the brand image of our company through the comments that users have posted on social networks. The idea is to create a model that classifies each comment as positive, negative or neutral and then do the study. The first thing we do is dive into social networks and collect sixteen thousand messages where our company is mentioned. The problem now is that we don't have a label on the data, that is, we don't know what the feeling of each comment is. This is where semi-supervised learning comes into play. This type of learning has a little of the two previous ones. Using this approach, you start by manually tagging some of the comments. Once we have a small portion of tagged comments, we train one or more supervised learning algorithms on that small portion of tagged data and use the resulting training models to tag the remaining comments. Finally, we train a supervised learning algorithm using as labels those manually tagged plus those generated by the previous models.

Reinforcement  learning

Finally, reinforcement learning is an automatic learning method based on rewarding desired behaviors and penalizing unwanted ones. Applying this method, an agent is able to perceive and to interpret the environment, to execute actions and to learn through test and error. It is a learning that sets long term objectives to obtain a maximum general reward and achieve an optimal solution. The game is one of the most used fields to test reinforcement learning. AlphaGo or Pacman are some games where this technique is applied. In these cases, the agent receives information about the rules of the game and learns to play by himself. At first, obviously, it behaves randomly, but with time it starts to learn more sophisticated movements. This type of learning is also applied in other areas as the robotics, the optimization of resources or systems of control.

Automatic learning is a very powerful tool that converts data into information and facilitates decision making. The key is to define in a clear and concise way the objective of the learning in order to, depending on the characteristics of the dataset we have, select the type of learning that best fits to give a solution that responds to the needs.

Content elaborated by Jose Antonio Sanchez, expert in Data Science and enthusiast of the Artificial Intelligence.

Contents and points of view expressed in this publication are the exclusive responsibility of its author.

El Bosco, Tiziano, El Greco, Rubens, Velázquez, Goya ... The Prado National Museum has more than 35,000 works of unquestionable value, which make it one of the most important art galleries in the world.

In the year of its bicentennial, we approached the Prado joint with Javier Pantoja, Head of the Digital Development Area, to know what innovative projects has been implemented to enrich the visitors experience.

1. In recent years, the Prado Museum has launched a series of technological projects aimed at bringing its collections closer to citizens. How did this idea come about? What is the digital strategy of the Prado Museum?

The Prado Museum teams does not only seek to develop a digital strategy, but we want “the digital” to be part of the global museum's strategy. Therefore, the last 3 strategic plans have included a specific line of action for the digital area.

When we talk about digital strategy in the Prado, we refer to two aspects. On the one hand, we talk about processes, that is, the improvement of internal management. But, on the other hand, we also use digital tools to attract a higher audience to the Prado.

Our intention is considering technology as a tool, not as an end. We have a strong commitment to technologies such as linked data or artificial intelligence, but always without losing the objective of spreading the

2. What kind of technologies are you implementing and how are you doing it? Can you tell us some concrete projects?

The most interesting point of the project was the launch of the current Prado website at the end of 2015. We were looking to create a semantic web, conceiving collections and the related information in a different way, that allow us to go further, and generate projects as augmented reading or the timeline that we have recently launched using Artificial Intelligence.

It has been a many-years process. The first task was to digitize and document the collection. For that we needed internal applications that allowed all areas of the museum to nurture the databases with their knowledge in a homogeneous way.

Then, we had to make that information available to the public using semantic web models. That is, we needed to define each element semantically: what a “technique”, a “support”, a “painter”, a “sculptor”, etc. are. It was about creating relationships between all the data to result in knowledge graphs. For example, we want to link Goya to the books that speak about him, the kings who ruled in his time, the techniques he used, etc. The work was tedious, but necessary to obtain good results.

This allowed to create a faceted search engine, which helped bring the collection closer to a much more users. With a boolean search engine, you have to know what you are looking for, while a faceted one is easier to find relationships.

All this informative work of updating, reviewing and publishing information open the way for a second part of the project: it was necessary to make the information attractive, through an aesthetic, beautiful and high-usability web. Do not forget that our goal was "position the Prado on the Internet." We had to create an easy-to-use data web. In the end, the Prado is a museum full of works of art and we seek to bring that art closer to the citizens.

The result was a website that has won several awards and recognitions.

3. And how did the augmented reading projects and the timeline come about?

They come up from a very simple matter. Let me give an example. An user accesses to the La Anunciación by Fra Angelico work sheet. This sheet mentions terms such as Fiésole or Taddeo Gaddi. Probably, the majority of the users does not know what or who they are. That is, the information was very well written and documented, but it did not reach the entire audience.

Here, we had two solutions. On the one hand, we could make a basic-level worksheet, but creating a text adapted to anyone is very complex, regardless of their age, nationality, etc. That is why we chose another solution: What does a user do to find something he/she didn´t know? In this situation, users search on Google and click on Wikipedia.

As the Prado Museum cannot be an encyclopedia, we took advantage of the knowledge of Wikipedia and DBpedia. Mainly for 3 reasons:

• It has the precise knowledge structure
• It is a context encyclopedia
• It is an open source

4. What challenges and barriers have you found when reusing this data?

First of all, we had to write down all the entities, cities, kings… one by one. The work was impossible, since there are currently 16,000 worksheets published on the web. In addition, the Prado is constantly studying and reviewing publications.

That is why we used a natural language recognition engine: the machine reads the text and the Artificial Intelligence understands it as a human, extracts entities and disambiguates the terms. The machine is processing the language, which it understands based on the context. In this task, we use the knowledge graph we already had and the relationships between the different entities through DBpedia.

The work is carried out together with Telefónica -sponsoring company- and GNOSS -developing company-; and the degree of reliability was very high, between 80% and 90%. Even so, there were complex issues. For example, when we talk about “the virgin's announcement”, we do not know if we refer to the concept, to the church in Milan, to some of the paintings related to this subject... In order to ensure that everything was correct, the documentation service reviewed the information.

5. What was the next step?

We already had an increased reading. At this point, we asked ourselves why not making these relationships visible. And so that, the Timeline emerged: a cluster of new knowledge graphs that allowed the user to see in a simpler way the relationships between concepts, result of the exploitation of the linked data web.

Timelines tend to had just one layer, but we wanted to go further and create a multilayer structure that would allow us to understand and deepen the concepts: a layer of history, another layer of literature, architecture, philosophy, performing arts ... In this way we can easily see, for example, what works were created during the 100-year war.

For this, we had to review the datasets and select the concepts according to the interest they generate and their context into the Prado collection.

6. These types of projects have great potential to be used in the educational field ...

Yes, our tool has a marked informative intention and has great potential to be exploited in educational environments. We have tried to make it easy for any teacher or disseminator to have at a glance the entire context of a work.

But it could also be used to learn history. For example, you can teach the Spanish Independence War using “The second of May” and “The third of May”, paintings elaborated by Goya. In this way, more active learning could be achieved, based on the relationships between concepts.

The tool has an appropriate approach for secondary and high school students, but could also be used at other stages.

7. What profiles are necessary to carry out such a project?

We create multidisciplinary teams, composed of designers, computer scientists, documentaries, etc. to carry out the entire project

There are very good specialists doing semantics, working on linked data, artificial intelligence, etc. But it is also necessary to have people with the ideas to join the puzzle into something usable and useful for the user. That is to say, linking wills and knowledge, around a global idea and objectives. Technological advances are very good, and you have to know them to take advantage of their benefits, but always with a clear objective

8. Could you tell us what are the next steps to be followed in terms of technological innovation?

We would like to reach agreements with other museums and entities so that all knowledge is integrated. We would like to enrich the information and link it with data from international and national cultural and museum institutions.

We already have some alliances, for example, an agreement with the National Film Library and the RTVE visual archive, and we would like to continue working on that line.

In addition, we will continue working with our data to bring the Prado collection closer to all citizens. The Museum has to be the first reuser of its data sources because it is who best knows them and who can get a good result from them.

Data science is an interdisciplinary field that seeks to extract actuable knowledge from datasets, structured in databases or unstructured as texts, audios or videos. Thanks to the application of new techniques, data science is allowing for answering questions that are not easy to solve through other methods. The ultimate goal is to design improvement or correction actions based on the new knowledge.

The key concept in data science is SCIENCE, and not really data, considering that experts has even begun to speak about a fourth paradigm of science, including data-based approach together with the traditional theoretical, empirical and computational approachs.

Data science combines methods and technologies that come from mathematics, statistics and computer science. It include exploratory analysis, machine learning, deep learning, natural-language processing, data visualization and experimental design, among others.

Within Data Science, the two most talked-about technologies are Machine Learning and Deep Learning, both included in the field of artificial intelligence. In both cases, the objective is the construction of systems capable of learning to solve problems without the intervention of a human being, including from orthographic or automatic translation systems to autonomous cars or artificial vision systems applied to use cases as spectacular as the Amazon Go stores.

In both cases, the systems learn to solve problems from the datasets “we teach them” in order to train them to solve the problem, either in a supervised way - training datasets are previously labeled by humans-, or in a unsupervised way - these data sets are not labeled-.

Actually, the correct point of view is to consider deep learning as a part of machine learning so, if we have to look for an attribute to differentiate both of them, we could consider their method of learning, which is completely different. Machine learning is based on algorithms (Bayesian networks, support vector machines, clusters analysis, etc.) that are able to discover patterns from the observations included in a dataset. In the case of deep learning, the approach is inspired, basically, in the functioning of human brain´s neurons and their connections; and there are also numerous approaches for different problems, such as convolutional neural networks for image recognition or recurrent neural networks for natural language processing.

The suitability of one approach or another will rely on the amount of data we have available to train our artificial intelligence system. In general, we can affirm that, if we have a small amounts of training data, the neural network-based approach does not offer superior performance than the algorithm-based approach. The algorithm-based approach usually come to a standstill due to huge amount of data, not being able to offer greater precision although we teach more training cases. However, through deep learning we can have a better performance from this greater amount of data, because the system is usually able to solve the problem with greater precision, the more cases of training are available.

None of these technologies is new at all, considering that they have decades of theoretical development. However, in recent years new advances have greatly reduced their barriers: the opening of programming tools that allow high-level work with very complex concepts, open source software packages to run data management infrastructures, cloud tools that allow access to almost unlimited computing power and without the need to manage the infrastructure, and even free training given by some of the best specialists in the world.

All this issues are contributing to capture data on an unprecedented scale, and store and process data at acceptable costs that allow us to solve old problems with new approaches. Artificial intelligence is also available to many more people, whose collaboration in an increasingly connected world is giving rise to innovation, advancing increasingly faster in all areas: transport, medicine, services, manufacturing, etc.

For some reason, data scientist has been called the sexiest job of the 21st century.

Content prepared by Jose Luis Marín, Head of Corporate Technology Startegy en MADISON MK and Euroalert CEO.

Contents and points of view expressed in this publication are the exclusive responsibility of its author.

Not all press headlines on Artificial Intelligence are true, but do not make the mistake of underestimating the changes in the coming years thanks to the development of AI.

Artificial intelligence has not yet changed our lives, but we are experiencing a cycle of over-exposure about its results and applications in the short term. The non-specialized media and the marketing departments of companies and other organizations foster a climate of excessive optimism regarding the current achievements in the development of AI.

Undoubtedly, we can all ask our smartphone some simple questions and get reasonable answers. In our online purchases (for example on Amazon.com), we experience the recommendation of personalized products based on our profile as buyers. We can search for events in our digital photographs library hosted in an online service such as Google Photos. With just writing, "mum's birthday" we will get a list of photos of that day with a relatively high precision in seconds. However, the time has not yet come for AI modifying our daily experience when we consume digital or physical products and services. Up to now, our doctor does not use AI to support his diagnosis and our lawyer or financial advisor does not use AI to prepare a resource or better invest our money.

The hot news floods us with photorealistic infographics of a world full of autonomous cars and virtual assistants with whom we talk as if it were a person. The risk of overexploitation in the short term is that, in any development process of a disruptive and exponential technology, technology necessarily fails. This leads to a state of disappointment that entails a reduction of investments by companies and states, which in turn produces a slowdown in the development of such technology. The development of AI is not unconnected to this process and has already known two stages of slowdown. These historical stages are known as the "winters of artificial intelligence". The first winter of the AI took place in the early years of the 1970 decade. After the birth of what today we know as IA - in the decade of 1950 - at the end of the sixties, Marvin Minsky got to assure in 1967 that "... in the course of a generation... the problem of creating an artificial intelligence will be practically solved...". Only three years later, Minsky himself said: "... in three to eight years we will have achieved a machine with an artificial intelligence superior to human beings ..."

Currently, no expert would dare to specify when (if so) this could happen. The second winter of AI arrived in the early years of the nineties. After a period of overexploitation of the then known as "expert systems". Expert systems are computer programs that contain logical rules that code and parameterize the operation of simple systems. For example, a computer program that encodes the rules of the chess game belongs to the type of programs that we know as expert systems. The coding of fixed logic rules to simulate the behavior of a system is what is known as symbolic artificial intelligence and was the first type of artificial intelligence that was developed.

Between 1980 and 1985, companies had invested more than one billion US dollars each year in the development of these systems. After this period, it was demonstrated that these turned out to be extremely expensive to maintain as well as not scalable and with little results for the investment they entailed. The historical journey of t AI development from its beginnings in 1950 to the present is a highly recommended and exciting reading.

What has artificial intelligence achieved so far?

Although we say that we must be prudent with the expectations regarding AI applications in the short term, we must punctuate here what have been the main achievements of artificial intelligence from a strictly rigorous point of view. For this, we will base on the book Deep Learning with R. Manning Shelter Island (2018) by Francois Cho-llet with J.J. Allaire.

• Classification of images at an almost-human level.
• Recognition of spoken language at an almost-human level.
• Transcription of written language at an almost-human level.
• Substantial improvement in the text conversion to spoken language.
• Substantial improvement in translations.
• Autonomous driving at an almost-human level.
• Ability to answer questions in natural language.
• Players (Chess or Go) that far exceed human capabilities.

After this brief introduction of the current historical moment in regarding to AI, we are in a position to define, in a slightly more formal way, what artificial intelligence is and which fields of science and technology directly are affected by its development.

Artificial intelligence could be defined as: the field of science that studies the possibility of automating intellectual tasks that are normally performed by humans. The reality is that the scientific domain of AI is normally divided into two sub-fields of computer science and mathematics called Machine Learning and Deep Learning. The representation of this statement can be seen in Figure 1.

Figure 1. Artificial Intelligence and sub-fields such as Machine Learning and Deep Lear-ning.

We are not going to speak here about technical considerations about the difference between machine learning and deep learning (you can learn more about these concepts in this post). However, perhaps the most interesting is to understand the paradigm shift from the first development of AI based on logical rules to the modern conception of our days, where machines find out those rules that govern a certain process (for example, the risk of suffering a myocardial infarction) based on input data and previous experiences that have been recorded and analysed by a certain algorithm. Figure 2 represents that change of paradigm between the classic programming of logical rules and the machine learning approach.

Figure 2. The paradigm shift to Machine Learning from classical programming.

The fast development of AI in our days is the result of multiple factors. However all experts agree that two of them have been and are key to enhance this third wave of artificial intelligence development: the reduction of computing cost and the explosion of available data thanks to the Internet and the connected devices.

Just as the Internet has impacted, and more than it is going to impact, in all areas of our lives - from the end consumer to the processes and operating models of the most traditional companies and industries - the IA will radically change the way in which humans use our capacities for the development of our species. There is a lot of news about the negative impact that AI will have on jobs positions in the next 10-20 years. While there is no doubt that the automation of certain processes will eliminate some jobs that still exist today, it is no less true that, precisely, the first jobs to disappear will be those jobs that make us "less human". Those jobs that employ people as machines to repeat the same actions, one after another, will be replaced by agents equipped with AI, either in the form of software or in the form of robots equipped with AI.

The iterative and predictable tasks will fall within the exclusive domain of the machines. On the contrary, those tasks that require leadership, empathy, creativity and value judgments will continue to belong only to humans. So, is it done? Do we have nothing else to do but wait for the machines to do part of the work while we do the other part? The answer is no. The reality is there is a big space in the middle of these two extremes.

The space in which the human is amplified by the AI ​​and in return the AI ​​is trained by humans to feed back this intermediate cycle. Paul R. Daugherty and H. James Wil-son, define in his book Human + Machine this space as "The missing middle”. In this intermediate space, AI machines increase human capabilities, make us stronger with exoskeletons capable of lifting heavy loads or restore us the capacity to walk after paralysis; they make us less vulnerable by working in potentially harmful environments such as space, the bottom of the sea or in a pipeline with deadly gases; We increase our sensory capabilities in real time with glasses that carry cameras and sensors that complement our field of vision with overlapping textual information.

With the help of machines equipped with an improved AI, it is not difficult to imagine how the processes and maintenance tasks will change in a factory, the different way in which our fields will be cultivated, how the warehouses of the future will work and the efficiency of our self-managed homes.

In conclusion, let's stop looking at the machines and AI as rivals that will dispute our jobs, and start thinking about how we will merge our most human skills with the superior capabilities of the machines regarding resistance, speed and precision.

Let's do different things now to do different things in the future.

Content prepared by Alejandro Alija, expert in Digital Transformation and innovation.

Contents and points of view expressed in this publication are the exclusive responsibility of its author.

The commercial adoption of any new technology and, therefore, its incorporation into the business value chain follows a cycle that can be moulded in different ways. One of the best known models is the Gartner hype cycle. With regard to artificial intelligence and data science, the current discussion focuses on whether the peak of inflated expectations has already been reached or, on the contrary, we will continue to see how the promises of new and revolutionary innovations increase.

As we advance in this cycle, it is usual to find new advances in technology (new algorithms,  in the case of Artificial Intelligence) or a great knowledge about their possibilities of commercial application (new products or products with better price or functionalities). And, of course, the more industries and sectors are affected, the higher expectations are generated.

However, the new discoveries do not only remain on the technological level, but it usually also go deeper into the study and understanding of the economic, social, legal or ethical impact derived from the innovations that are arriving on the market. For any business, it is essential to detect and understand as soon as possible the impact that a new technology will have on its value chain. This way, the company will be able to incorporate the technology into its capabilities before its competitors and generate competitive advantages.

One of the most interesting thesis recently published to model and understand the economic impact of Artificial Intelligence is the one proposed by Professor Ajay Agrawal with Joshua Gans and Avi Goldfarb in his book "Prediction Machines: The Simple Economics of Artificial Intelligence”. The premise is very simple: at the beginning, it establish that the purpose of artificial intelligence, from a merely economic point of view, is to reduce the cost of predictions.

When the cost of a raw material or technology is reduced, it is usual for the industry to increase their use, first applying this technology to the products or services it was designed for, and later, to other product or services that were manufactured in another way. Sometimes it even affects the value of substitute products (that fall) and complementary products (that rise), or other elements of the value chain.

Although these technologies are very complex, the authors were able to establish a surprisingly simple economic framework to understand the AI. But let's see a concrete case, familiar to all of us, in which the increase of the accuracy of the predictions, taken to the extreme, could mean not only to automate a series of tasks, but also to completely change the rules of a business .  

As we all know, Amazon uses Artificial Intelligence for the purchase recommendation system that offers suggestions for new products. As mentioned by the authors in his book, the accuracy of this system is around 5%. This means that users acquire 1 out of every 20 products that Amazon suggests, which is not bad.

If Amazon is able to increase the accuracy of these predictions, let's say to 20%, that is, if users acquire 1 out of every 5 suggested products, Amazon would increase its profits enormously and the value of the company would skyrocket even more. But if we imagine a system capable of having a precision of 90% in the purchase predictions, Amazon could consider radically changing its business model and send us products before we decide to buy them, because we would only return 1 out of every 10. AI would not just automate tasks or improve our shopping experience, it would also radically change the way we understand the retail industry.

Given that the main substitute for AI predictions are human predictions, it seems clear that our value as a predictive tool will continue decreasing. The advance of the wave of automations based on artificial intelligence and data science already allows us to see the beginning of this trend.

On the contrary, company data would become an increasingly valuable asset, since they are the main complementary product necessary to generate correct predictions. Likewise, the public data necessary to enrich the companies data, and thus make possible new use cases, would also increase its value.

Following this line of reasoning, we could dare to establish metrics to measure the value of public data where they were used. We would only have to answer this question: how much improves the accuracy of a certain prediction if we enrich the training with determined open data? These improvements would have a concrete value that could give us an idea of the economic value of a public dataset in a specific scenario.

The repeated mantra "data is the new oil" is changing from being a political or marketing affirmation to being supported by economic science, because data are the necessary and indispensable raw material to make good and valuable predictions. And it seems clear that, to continue reducing the predictions cost, the data value should increase. Simple economy.

Content prepared by Jose Luis Marín, Head of Corporate Technology Startegy en MADISON MK and Euroalert CEO.

Contents and points of view expressed in this publication are the exclusive responsibility of its author.