Found in translation: More accurate, fluent sentences in Google Translate

In 10 years, Google Translate has gone from supporting just a few languages to 103, connecting strangers, reaching across language barriers and even helping people find love. At the start, we pioneered large-scale statistical machine translation, which uses statistical models to translate text. Today, we’re introducing the next step in making Google Translate even better: Neural Machine Translation.  

Neural Machine Translation has been generating exciting research results for a few years and in September, our researchers announced Google's version of this technique. At a high level, the Neural system translates whole sentences at a time, rather than just piece by piece. It uses this broader context to help it figure out the most relevant translation, which it then rearranges and adjusts to be more like a human speaking with proper grammar. Since it’s easier to understand each sentence, translated paragraphs and articles are a lot smoother and easier to read. And this is all possible because of end-to-end learning system built on Neural Machine Translation, which basically means that the system learns over time to create better, more natural translations.

Today we’re putting Neural Machine Translation into action with a total of eight language pairs to and from English and French, German, Spanish, Portuguese, Chinese, Japanese, Korean and Turkish. These represent the native languages of around one-third of the world's population, covering more than 35% of all Google Translate queries!

With this update, Google Translate is improving more in a single leap than we’ve seen in the last ten years combined. But this is just the beginning. While we’re starting with eight language pairs within Google Search  the Google Translate app, and website; our goal is to eventually roll Neural Machine Translation out to all 103 languages and surfaces where you can access Google Translate.

And there’s more coming today too -- Google Cloud Platform, our public cloud service, offers Machine Learning APIs that make it easy for anyone to use our machine learning technology. Today, Google Cloud Platform is also making the system behind Neural Machine Translation available for all businesses through Google Cloud Translation API. You can learn more about it here.

Today’s step towards Neural Machine Translation is a significant milestone for Google Translate, but there’s always more work to do and we’ll continue to learn over time. We’ll also continue to rely on  Translate Community, where language loving multilingual speakers can help share their language by contributing and reviewing translations. We can’t wait for you to start translating and understanding the world just a little bit better. 

Here’s How Automation Will Prevent Trump’s Job Promises

IN BRIEF

• Wadhwa believes Donald Trump will have difficulty creating jobs given the shift toward automation, with robots predicted to do some jobs for one-tenth the cost of human workers.
• He urges talks between the government and the private sector on how to distribute new wealth and cope with the fact that tech will disrupt industries and eliminate millions of jobs.

KEEPING PROMISES

One of the more resounding aspects of president-elect Donald Trump’s campaign platform was his vow of job creation — or job restoration — by bringing back jobs that American companies have shipped overseas. In his column for The Washington Post, renowned expert Vivek Wadhwaexplains why he thinks Trump will have a difficult time implementing his plans in the face of the expected shift toward automation.

“[A]dvancements in technology and how they reshape our economy may also keep him from delivering on some of the major promises that made him so popular during the campaign season,” writes Wadhwa, who’s a fellow at Stanford University’s Rock Center for Corporate Governance, director of research for the Center for Entrepreneurship and Research Commercialization at Duke, and a distinguished fellow at the Singularity University.

This isn’t because artificial intelligence (AI) is out to get Trump (or us). It’s just that technology’s been on a course toward automation, and no one can change that. The trends all point towards a future of job automation, Wadhwa explains in the article.

BIGGER THAN TRUMP

Wadhwa provides several reasons why attempting to stifle technology would be counterproductive and could even hurt the United States’ economy more.

For one, most of the jobs Trump targets are in the manufacturing sector, where automation already has a strong foundation, and other lower-skill industries are also being disrupted. “Machines are learning to do the jobs of manufacturing workers; artificial intelligence-based tools are mastering the jobs of call-center and knowledge workers; and cars are beginning to drive themselves,” Wadhwa says, the latter point referring to companies like Uber that are testing the waters with driverless taxis and trucks.

He goes on to explain in detail why automation is financially sound for business:

Some robots already cost less to operate than the salaries of the humans they replace, and they are getting cheaper and better. Boston Consulting Group predicts that, by 2025, the operating cost of a robot that does welding will be less than $2 per hour, for example. That’s more affordable than the $25 per hour that a human welder earns today in the U.S., and even cheaper than the pay of skilled workers in the lowest-income countries.

Businesses, at the end of the day, will resort to the cheaper option in the long run, and that is automation. So, Wadhwa asks, “Trump may be able to keep immigrants out, but how will he stop the advance of robots?”

Rather than fight these advancements, which Wadhwa says could also be used to “provide solutions to the problems of humanity, such as a lack of energy, food, education and health care,” he emphasizes the need for communication between the government and the private sector.

We now need a nationwide conversation on how we can distribute the prosperity we are creating. We must create equity and fairness in our legal, justice and economic systems. And, recognizing that technology will disrupt entire industries and wipe out millions of jobs, so we must ease the transition and pain for the people most affected and least prepared.

Whether Donald Trump is the man to lead this conversation remains to be seen, but if he has any hope of making good on his campaign promises, he’s going to have to try.

Humanoids 2016: "Deep Learning ist keine black box"

Patrick van der Smagt

 

(Bild: heise online / Hans-Arthur Marsiske)

… und Intelligenz kein heiliger Gral, sondern lässt sich berechnen – sagt Patrick van der Smagt, seit kurzem leitender Forscher bei Volkswagen, im Interview am Rande der Konferenz Humanoids 2016.

Am Ende hat es dann doch noch geklappt mit den Fußballrobotern. Am letzten Tag der Konferenz Humanoids 2016 in Cancún konnte Daniel Lofaro von der George Mason University endlich die Mini-Darwins auf einem kleinen Spielfeld im Flur vor den Vortragssälen antreten lassen. Wer Lust hatte, konnte sich eine Virtual-Reality-Brille aufsetzen, mit einem Gamepad die kleinen Roboter steuern und versuchen, den Ball zu treffen. Wenn nur die Perspektive der Roboterkamera zur Verfügung steht, ist das gar nicht so einfach. Und das ist auch die Idee: Lofaro möchte beim RoboCup einen Wettbewerb für ferngesteuerte humanoide Roboter etablieren, um die Aufmerksamkeit stärker auf das Problem der Wahrnehmung zu lenken.

Es ist gar nicht so einfach, aus Sicht eines Roboters den Ball zu treffen. 
Bild: heise online / Hans-Arthur MarsiskeWahrnehmung war natürlich auch in vielen Vorträgen ein Thema, insbesondere die taktile. Beim Menschen ist es offensichtlich, wie wichtig der Tastsinn für die Orientierung ist. Das zeigt sich bei so vermeintlich simplen Aufgaben wie dem Zerschneiden eines Brötchens oder Apfels oder auch dem Schälen einer Banane. Das probierte Akihiko Yamaguchi (Carnegie Mellon University) ferngesteuert mit dem zweiarmigen Baxter-Roboter – mit ziemlich matschigem Ergebnis, wie eine Videoaufnahme des Versuchs zeigt.

Um Robotern das nötige Gefühl für die Welt zu verleihen, braucht es zum einen geeignete Sensoren, zum anderen aber auch Verfahren, die anfallenden Daten zu verarbeiten. Das führt zu einem der derzeit am heißesten diskutierten Themen nicht nur in der humanoiden Robotik: Deep Learning. Es war am letzten Konferenztag Thema eines Plenarvortrags. Wir haben danach mit dem Referenten Patrick van der Smagt, der früher an der TU München gelehrt hat und vor kurzem Leiter der Forschung zu Künstlicher Intelligenz im data lab bei der Volkswagen AG geworden ist, über das Thema gesprochen.

van der Smagt: Forschungen zu maschinellem Lernen und neuronalen Netzen haben bereits in den 1940er-Jahren begonnen. Aber erst jetzt ist klar geworden, dass das komische Wort "Intelligenz" kein heiliger Gral ist, sondern etwas, das sich berechnen lässt. Es wächst das Bewusstsein, dass dieses Verkoppeln von Kausalitäten, das unser Gehirn leistet, eigentlich ein Teil der Wahrscheinlichkeitstheorie ist. Das lässt sich heute in sehr eingeschränkten Domänen bereits berechnen und nachbauen. Bildverarbeitung ist dafür ein sehr gutes Beispiel: Rechner können heute Gesichter oder Objekte besser und viel schneller erkennen als Menschen.

Die Unruhe erstreckt sich insbesondere auch auf viele Forscher, die Deep Learning als nicht nachvollziehbare "black box" bezeichnen...

Das ist es aber nicht.

Es handelt sich doch um einen rein mathematischen Prozess. Wie kann der nicht nachvollziehbar sein?

Es ist ein mathematischer Prozess, der teilweise gut verstanden ist. Diese Mathematik gab es schon immer, aber jetzt lässt sie sich sehr effizient mit neuronalen Netzen berechnen. Dadurch wird sie plötzlich anwendbar. Es ist wie zwei Werkzeuge, die ich schon immer hatte, aber jetzt auf einmal kombinieren kann. Dadurch können wir so etwas wie Sprach- und Bilderkennung realisieren oder Kausalitäten entdecken.

Wie kommt es dann zu dieser Idee der black box?

In den technischen Wissenschaften und insbesondere in der Robotik ist der Ansatz sehr verbreitet, die Lösung einer Aufgabe mathematisch auszuformulieren. Dazu werden Modelle der Welt formalisiert. Der Standardansatz, um zum Beispiel eine Tasse zu greifen, besteht darin, ein Modell dieser Tasse zu erstellen, das Höhe, Breite, Masse und andere Parameter wiedergibt. Daraus lässt sich berechnen, wo der Griff am besten ansetzen soll. Wenn ich es komplizierter haben will, kann ich auch noch den Reibungskoeffizienten und weitere Größen berechnen. Der Ansatz funktioniert wunderbar, sofern ich weiß, welche Tasse ich packen möchte. Aber bei einer anderen Tasse habe ich dann ein Problem. Das gilt ebenso für Bildverarbeitung, taktile Sensorik, Kraftsensorik, Motoren und so weiter. In einer Laborumgebung, in der ich die Welt perfekt kontrollieren und modellieren kann, funktioniert die Methode sehr gut. Ebenso in der Produktion: Wenn ich weiß, welche Teile ich habe und wo sie liegen, kann ich sie greifen und platzieren.

Aber jetzt gibt es Methoden, um das Gleiche mit ähnlicher Genauigkeit zu lernen und das gelernte Verhalten zu generalisieren. Man gibt dem System zehn Beispiele zum Lernen und danach erledigt es die weiteren Aufgaben selbstständig. Wenn es dabei zu einem Fehler kommt, kann ich allerdings nicht mehr sagen, welche Zeile im Programmcode dafür verantwortlich ist oder welche Stellschraube wie weit gedreht werden muss, um das Problem zu korrigieren. Stattdessen kann ich eine Vorhersage treffen, wie gut ich mit welcher Wahrscheinlichkeit greifen werde und kann den voraussichtlichen Fehler beim Greifen quantifizieren. Das ist eine völlig andere Herangehensweise und Denkweise. Ich kann nicht mehr zurückrechnen und den Fehler eindeutig identifizieren. Das ist zum Beispiel ein Problem beim autonomen Fahren. Ich kann sehr gut probabilistische Aussagen treffen, kann das Verhalten aber nicht bis ins letzte Detail analysieren.

Ist das vergleichbar mit dem Problem, das Anbieter etwa von Industrierobotern mit Open-Source-Software haben? Viele tun sich damit schwer, weil sie ihre Systeme dann nicht mehr hundertprozentig spezifizieren und dem Kunden garantieren können, dass nichts verändert wird.

Open Source ist etwas anders gelagert, weil man hier niemanden für einen Fehler im Code verantwortlich machen kann. Lernende Systeme dagegen sind so komplex, dass ich sie nicht mehr Regel für Regel beschreiben, sondern nur noch Wahrscheinlichkeiten benennen kann. Eigentlich war das schon immer so. Aber inzwischen sind probabilistischen Systeme deutlich leistungsfähiger als die regelbasierten Systeme. Das erfordert ein gewisses Umdenken.

Nun gibt es neuronale Netze und maschinelles Lernen schon lange. Was ist das Neue an Deep Learning?

Ich habe einen Datensatz mit einer gewissen Wahrscheinlichkeitsverteilung. Den kann ich abbilden auf einen anderen Datensatz mit einer anderen Wahrscheinlichkeitsverteilung. Mehr ist das nicht. Ich sehe zum Beispiel ein Bild und kann die Wahrscheinlichkeit ausrechnen, ob es einen Hund oder eine Katze zeigt, weil ich es vorher gelernt habe. Die dafür verwendeten Methoden der Neuronalen Netze sind in den 1950er- und 1960er-Jahren aus dem damaligen Verständnis von der Funktionsweise des Gehirns abgeleitet worden. Die Neuronen sind kleine Berechnungselemente, die ihre Aktivität aus ihrem Input berechnen. Diese Aktivität geben sie wiederum weiter in ein Netzwerk, das aus hundert, tausend oder noch mehr solcher Neuronen bestehen kann. Der Einfluss eines Neurons auf ein anderes lässt sich lernen, dafür gibt es Algorithmen. Die mathematische Basis dafür heißt back propagation. Mithilfe dieser Algorithmen lassen sich bessere Gewichtungen der wechselseitigen Beeinflussung der Neuronen finden, um eine bestimmte Abbildung von Input auf Output zu lernen.

Das heißt, das System verändert die Parameter seines eigenen Programms?

Genau, diese Parameter werden gelernt. Der Unterschied zwischen Deep Neural Networks und Neural Networks besteht darin, dass Deep Neural Networks eine höhere Komplexität von Schichten mit versteckten Neuronen haben und dadurch besser generalisieren können. Es ist keine neue Methode, kein Paradigmenwechsel, lässt sich mit den heute verfügbaren Rechnern aber effizient berechnen.

Das heißt, das "deep" in Deep Learning bezieht sich auf die Zahl der neuronalen Zwischenschichten: je mehr, desto tiefer?

Genau. Aber erst seit wenigen Jahren haben wir die Rechner, um damit effizient arbeiten zu können.

Verstehe ich es richtig, dass der Output einer Schicht dabei der Input ist für die nächste Schicht? Also die erste Schicht mag in einem Bild Kanten und Linien detektieren, die nächste fügt sie zu geometrischen Formen zusammen, in denen die wiederum nächste dann Objekte erkennt?

Richtig. Zufälligerweise macht das Gehirn es genauso.

Zufälligerweise?

Der Zufall liegt darin, dass die probabilistische Darstellung der Daten für Effizienz sorgen. Das heißt nicht, dass die Struktur der neuronalen Netze die gleiche wäre wie im Gehirn. Aber es ist eine Methode, die die gleiche Wahrscheinlichkeitsdarstellung effizient finden kann. Ich habe zwei unterschiedliche Methoden, die zum gleichen Ergebnis kommen. Wir sollten sehr vorsichtig sein mit der Behauptung, neuronale Netze würden irgend etwas vom Gehirn erklären.

Wie viel Entwicklungspotenzial sehen Sie in Deep Learning?

Wir stehen am Anfang.

Und wo kann es noch hingehen?

Ich denke schon, dass wir eine Intelligenz ähnlich der des Menschen schaffen können. Daran habe ich schon während meines Studiums nie gezweifelt, habe aber auch nicht erwartet, dass das in absehbarer Zeit gelingen könnte. In Teilbereichen, etwa in der Bildverarbeitung oder Bilderkennung, sind wir schon weit gekommen. Von allgemeiner Intelligenz sind wir dagegen noch weit entfernt. Ob es fünf Jahre sind oder fünfzig, kann ich nicht sagen. Ich hoffe ein bisschen auf fünf, vielleicht auch zehn Jahre, dann könnte ich noch mitmachen. Aber früher oder später kommt es, davon können wir ausgehen. Wie die Gesellschaft dann aussieht, ist eine andere Frage. Die Notwendigkeit, 40 Stunden pro Woche zu arbeiten, dürfte es dann nicht mehr geben. Es gibt viele Jobs, die für Roboter viel besser geeignet sind als für Menschen. Übergänge sind immer schmerzhaft, aber danach könnte es viele Vorteile bringen.

Um noch einmal an die erste Frage anzuknüpfen: Ist die Unruhe angesichts von Deep Learning gerechtfertigt?

Wenn man sich vor gesellschaftlichen Veränderungen fürchtet, ja. Aber nur dann. Smartphones und Tablets haben in der menschlichen Interaktion auch sehr viel verändert, sicherlich nicht immer im positiven Sinn. Aber sie haben die Welt deutlich kleiner gemacht und das globale Denken innerhalb weniger Jahre komplett geändert. Künstliche Intelligenz, die ja heute schon in diesen Geräten steckt, wird das nur weiter beschleunigen. Ich freue mich darauf. (anw) 

Brace yourselves! Blighty is turning to AI for help

AI will improve UK government's digital services, apparently

The UK government is looking at ways to revamp its digital services with the help of machine learning and AI, according to a new report released by the Government Office for Science.
AI is the biggest trend taking over technology, and with the looming presence of autonomous vehicles and robots, governments are beginning to take an interest too.
The report was written by Chief Scientific Adviser Sir Mark Walport, and looks at ways the government could exploit AI to increase productivity and revenues, and improve decision making.
It’s quite wishy-washy however, and the government hasn’t made any concrete promises. It might decide to use machine learning, which could make existing services such as health, social care, or emergency services more efficient.
Or it could make it easier for officials to make better-informed decisions that are “more transparent” through better data analysis. And customer service may be improved by using that data to “help departments better understand the groups they serve, in order to be sure that the right support and opportunity is offered to everyone.”
Understanding how AI makes decisions is quite difficult, however. The government wants AI to identify key topics for policy makers and generate advice through data, but it won’t be able to understand how the system arrived at its decision very clearly, as AI isn't very good at explaining things.
Transparency is a problem that many researchers are trying to solve, so the government won’t be able to rely on it entirely and has vowed all decisions will have a “human in the loop.”
Although the government isn’t quite sure of the effects that AI could have on its digital services, it is willing to experiment by setting up “sandbox areas” for government analysts to test deep learning in a “safe and controlled environment.”
The two biggest issues are data protection and privacy. “Teams making use of artificial learning approaches need to understand how these existing frameworks apply in this context. For example, if deep learning is used to infer personal details that were not intentionally shared, it may not be clear whether consent has been obtained,” the report said.
The UK government doesn’t have a great reputation for managing its IT and data services, or having a clear direction for AI either. The “right form of governance for artificial intelligence, and indeed for the use of digital data more widely, is not self-evident,” it admitted.
“But the important task is to set out what needs to be done before considering how it is to be achieved,” the report said.

Eine tückische Blackbox

Das "Deep Learning" verhilft der KI derzeit zum Durchbruch. Doch die Sache hat einen Haken: Die Technik ist nahezu undurchdringlich - und verblüffend leicht zu täuschen.

Dean Pomerleau kann sich noch gut daran erinnern, als er sich zum ersten Mal mit dem Blackbox-Problem herumschlug. Man schrieb das Jahr 1991, und Pomerleau versuchte, einem Computer das Autofahren beizubringen. Inzwischen ist das dank Forschung an autonomen Fahrzeugen alltäglicher geworden. Seinerzeit war es eine echte Pioniertat.

Pomerleau klemmte sich hinter das Steuer seines speziell umgerüsteten Humvees und lotste den Militärtruck durch die Straßen der Stadt. An Bord hatte der Forscher, damals Robotik-Doktorand an der Carnegie Mellon University in Pittsburgh, einen Computer, den er darauf programmiert hatte, durch eine Kamera zu blicken, die Geschehnisse auf der Straße zu interpretieren und jedes Manöver, das Pomerleau machte, abzuspeichern. Die Hoffnung war, dass das Gerät irgendwann genügend Assoziationen hergestellt haben würde, um das Auto selbstständig zu lenken.

Bei jeder Fahrt trainierte er das System ein paar Minuten lang, dann ließ er es eigenständig fahren. Alles schien prima zu laufen – bis der Humvee eines Tages vor einer Brücke plötzlich seitwärts ausscherte. Nur durch einen schnellen Griff ins Lenkrad konnte Pomerleau einen Crash verhindern.

Wieder zurück im Labor, versuchte er zu ermitteln, was den Computer auf Abwege geführt hatte. "Das war Teil meiner Doktorarbeit: die Blackbox zu öffnen und herauszufinden, was das Gerät denkt", erklärt Pomerlau. Aber wie? Er hatte dem Rechner ein "neuronales Netz" einprogrammiert – ein Verfahren der künstlichen Intelligenz (KI), das sich am Aufbau des menschlichen Gehirns orientiert und gemeinhin besser mit komplexen Alltagssituationen umgehen kann als die Standardalgorithmen jener Zeit.

Ebenso undurchsichtig wie das Gehirn

Leider sind solche Netze aber auch ebenso undurchsichtig wie das Gehirn. Statt das Gelernte fein säuberlich als Datenblock im digitalen Speicher abzulegen, verteilen sie die Informationen in einer Art und Weise, die außerordentlich schwierig zu entschlüsseln ist. Erst nachdem Pomerleau die Reaktion seiner Software auf verschiedene visuelle Reize eingehend untersucht hatte, konnte er das Problem ausfindig machen: Das neuronale Netz hatte nämlich die grasbewachsenen Seitenstreifen als Orientierungshilfe für den Straßenverlauf benutzt. Als dann unvermittelt eine Brücke auftauchte, war die Irritation groß.

Heute, 25 Jahre später, wird es immer schwieriger, aber auch immer wichtiger, das Rätsel der Blackbox zu lösen. Die Technologie der KI hat sich, was Komplexität und Anwendungsfelder angeht, geradezu explosionsartig weiterentwickelt. Pomerleau, der an der Carnegie Mellon University inzwischen Robotik in Teilzeit lehrt, beschreibt sein kleines System auf dem Transportfahrzeug als "Arme-Leute-Version" der riesigen neuronalen Netze, die heutzutage auf den Rechnern laufen. Für die Technik des "Deep Learning", bei dem neuronale Netze mit Hilfe riesiger Datenbanken trainiert werden, gibt es inzwischen sogar erste kommerzielle Einsatzbereiche – von selbstfahrenden Autos bis zu Internetseiten, die aus dem Browserverlauf eines Benutzers Produktempfehlungen generieren.

Auch in der Wissenschaft deutet sich die Allgegenwart von Deep Learning an. Künftige Radioteleskope werden ohne die Hilfe der KI kaum sinnvolle Signale in ihren enormen Datenmengen ausmachen können; in Gravitationswellendetektoren soll Deep Learning noch die kleinsten Störquellen ausfindig machen und eliminieren helfen; Verlagshäuser werden Millionen wissenschaftlicher Veröffentlichungen und Bücher durchforsten und per KI verschlagworten. Und schließlich, so prognostizieren es zumindest einige Forscher, könnten mit Deep Learning ausgestattete Computer sogar Vorstellungsvermögen und Kreativität entwickeln. "Man würde dem Computer einfach große Datenmengen vorsetzen und bekäme dann die Naturgesetze ausgespuckt", meint der Physiker Jean-Roch Vlimant vom California Institute of Technology in Pasadena.

Google's neural network learns to translate languages it hasn't been trained on

First time machine translation has used true transfer learning

The gap between human and machine translators could be narrowing as researchers find a new way to improve the learning capabilities of Google Translate’s neural network.
On the same day that Google announced its translation services were now operating with its Neural Machine Translation (NMT) system, a team of researchers released a paper on arXiv showing how its NMT could be pushed one step further.
An NMT is a large single neural network that learns to translate by being trained on a pair of languages. Google Translate has been around for ten years, and before an NMT was used, it often provided a clumsy attempt at translation. Sentences could be lost in translation because individual phrases were translated together instead of whole sentences.
First, the encoder breaks down the sentence into its constituent words and represents the meaning as a vector.
The system interprets the whole sentence, and the decoder begins to translate each word by looking at the weighted distribution over the encoded vectors and matching them up to the most relevant words in the target language.
Google now promises to provide better translations that are “smoother and easier to read” for eight language pairs: English and French, German, Spanish, Portuguese, Chinese, Japanese, Korean and Turkish.

Photo credit: Google

For many NMTs, the machine encodes the sentence into vectors by focusing on each word individually – something called the “attention mechanism.” Several attention mechanisms were required to cope with different languages, and multiple encoders and decoders are used for translation.
But when a single attention mechanism is used for multiple languages, the system can be coaxed into learning to translate between pairs of languages it hasn’t encountered before – something the researchers call “zero-shot translation.”
“For example, a multilingual NMT model trained with Portuguese→English and English→Spanish examples can generate reasonable translations for Portuguese→Spanish, although it has not seen any data for that language pair,” the paper said.
Interestingly, the researchers found evidence that the network was learning “some sort of shared representation,” where sentences with similar meaning were represented in similar ways regardless of language.
Training the model across multiple languages can enhance the performance of translating into individual languages, and the system can even cope better with sentences that are written in a mixture of languages.
Although the researchers only used a maximum of 12 language pairs for a single model, it would be easy to add more, and simple to use, as it operates on the same architecture as the NMT currently being used by Google Translate.
Zero-shot translation is the first time that “true transfer learning” has been shown to work for machine translation, according to the researchers.
Transfer learning is considered key to developing true machine intelligence, as systems must learn to deal with new situations based on knowledge gained from previous experiences. ®

Set the Machines free to learn

A decade ago straight through processing was a buzz word and speed to market was critical. The progress financial institutions have made in moving almost all aspects of their transaction foot print digital has left little to leverage on the transaction side.

In today’s day and time while most organizations are busy revamping their Policy administration systems which were long ready to be replaced a decade ago, what will set companies apart will be the organizations that start considering Machine Learning and Artificial intelligence(AI) for their core systems. 

If you look at the fundamentals of any kind of insurance, at the core, insurance offerings are about risk pooling and the ability of the insurer to price products in a manner such that over time the premium revenues outstrip the claims experience. In every type of insurance product the claims experience influencing the pricing and risk aggregation decision making done by the insurer. Historically all this analysis has been done by people and rightly so as we lived in a world that was not connected and human intervention to analyze outside factors was critical.

Fast forward to current time, all the data is on some kind of digital medium and more often than not connected and accessible. What is missing is the Machine Learning and Artificial Intelligence integration into the different facets of the insurance life cycle and the software platforms that are used to manage and maintain the data . The amount of data that needs to be analyzed and the patterns that are needed to be determined are so humongous that relying on data analysis by a person alone may not be the best approach. 

If you use google often you would have noticed that now google can predict to certain level of accuracy what you are searching and what you are looking for based on data it collects on your location, your emails, you past transaction etc. Over time there will be a cognitive angle to the search capabilities exhibited by google. If you apply the same thumb rule to underwriting, insurance pricing , risk aggregation, why would we not want to leverage machine learning in a similar manner. For this to happen we need to start building software systems with not just automation in mind but also a consideration of how can system design extend the realms of machine learning. If the dots are connected and the data patterns understood and logic applied there are certain decision making aspects that can move away from people to machines and over time evolve to largely autonomous ecosystem. 

What will differentiate the Market leaders from the laggards is investment in this aspect. These changes will come in the next decade or maybe even sooner and the Underwriting and Actuarial aspects will lean to machine learning and AI assisted functions and the next wave would lead to a totally autonomous eco system.

The picture simplistically highlights the possibilities of embedding Machine learning in the software ecosystem that we see in today's insurance landscape. This is a generalized view agnostic of the domain or line of business. Insurance carriers would need to start thinking out of the box to translate this into software platforms of the future , enhancing changing and pushing current roles into those that co exist and or radically change them as we know them today.

So before we set the drones to fly and change the commercial insurance ecosystem, Machine learning and AI need to be adopted into mainstream core software platforms. The emerging market in the foreseeable future will be opened to the players that will NOT be consumed with dev-ops and pushing the realms of delivery automation but by those firms investing in infusing Machine learning and artificial intelligence into core platforms enabling UW and actuarial functions to be supplemented by machines.

Insurance has traditionally followed and adopted what has been tried and tested in the banking space. For a change this may be an opportunity for insurance carriers to take the leads and beat the banks and other financial institutions to set free the machines and change the way products are conceived and priced and premiums calculated.

About the author:

Christopher Fernandes has worked in insurance companies managing business operations as part of start up operations and subsequently moved to technology consulting helping insurance & health care companies improve process controls and innovate business operations.

6 Incredible Ways Businesses are Using Artificial Intelligence Today

In science fiction, artificial intelligence (AI) systems are often bent on overthrowing human civilization, or they are benevolent caretakers of our species. In reality, machine-learning is already with us, evolving out of search engines like Google and seeping into our everyday lives without much fanfare.

To highlight exactly how AI is debuting in a less-than-Hollywood (but still quite impactful) style, Business News Daily spoke with six companies rolling out AI products that will recalibrate and streamline everyday business operations. From conducting marketing research to monitoring for nefarious activity on stock exchanges, these systems can do it all. [See Related Story:AI Comes to Work: How Artificial Intelligence Will Transform Business]

SparkCognition – Monitoring the London Stock Exchange

In stock exchanges around the world, plenty of people are trying to tilt the table in their favor to ensure their investments pay off. Monitoring and policing this sort of activity is a key part of maintaining investors' faith in the stock exchange, but traditional methods have proved inadequate to address all  the malfeasance that's out there.

"Markets can be manipulated and, unfortunately, humans are constantly trying to come up with new ways to do so," Amir Husain, founder and CEO of SparkCognition, said. "The idea behind AI-based monitoring is to a) know about all the existing methods of market manipulation and guard against those and b) to learn about new forms of market manipulation and alert analysts to potential abuse."

Husain's team believes their system will beef up the monitoring power of regulators who oversee exchanges without inundating them with false positives and constant alerts. However, the system is designed to quickly catch on to both old and new techniques of gaming the market. Husain said that SparkCognition's AI could have even prevented larger wrongdoings, like the London Interbank Offer Rate (LIBOR) rigging scandal.

"Better oversight and protection to prevent manipulation helps everyone, including the institution in whose name a lone actor may be engaging in illegal and improper trading practices," Husain said. "If you are a business owner … [of a] publicly traded [company], AI-based surveillance systems will protect you by preventing scenarios where your stock is used in a scam. If you are an average investor, these systems will give you more reason to trust the value of your portfolio."

Equals3 – Lucy, market research companion

Proper analysis of keyword trends, algorithmic changes, point-of-sale data and so on is key to marketing success, but poring over the ever-increasing mountain of data is time-consuming and expensive. To help improve the efficiency of the process, Equals3 developed their product known as Lucy, a marketing research companion powered in part by the famous Jeopardy! champ, IBM Watson.

Incorporating a blend of Watson's services and Equals3's own services developed in-house, Lucy is a centralized AI that can pull up exactly the information an analyst requires. (Think Alexa on steroids and with a deep fixation on marketing and advertising.)

"Lucy is a combination of our own stuff and some amazing services from IBM's Watson to create this unique solution that brings together all the owned and licensed content, structured and unstructured, to our customers," Scott Litman, managing partner at Equals3, said. "You can ask her, 'How much did BMW spend month by month?' Lucy knows. 'What was the daily web traffic for Ford.com?' Lucy knows. I can ask her for keyword research and she knows."

The initial license for Lucy costs $50,000 for a year and allows for five users to operate within one brand or marketing category. Lucy comes with a basic set of knowledge and understanding from the get-go – U.S. Census Data, website analytics, social media management, keyword research – but will learn and adapt to the specific uses for which it is applied. For Litman, it appears inevitable that companions like Lucy will become the norm very shortly, and Equals3 wanted to become an early option in the marketing space.

"We truly believe AI is going to drive a competitive advantage and those that adopt early will win out over those that don't," Litman said.

Daisy Intelligence – Decision-making assistance

Try as we might, humans often make poor decisions. Machines, however, excel at examining vast sets of data and research to determine, quantifiably, the best course of action. That's the idea behind Daisy Intelligence'sdecision-making assistant. Primarily geared toward retail and insurance industry applications, this SaaS product offers itemized lists of decisions, ranging from the most optimal course of action to the least.

"You send us your raw data, our machine analyzes it and spits out decisions," Gary Saarenvirta, CEO of Daisy Intelligence, told Business News Daily. "We don't change your process at all, except that now you have a new piece of information to consider … It's about making smarter decisions; these are all mathematical quantitative things that computers do better than people. If you can make smarter decisions it can drive your profitability."

For example, Daisy Intelligence can determine what deals a retailer should offer, or what the featured product should be during the current ad campaign. Saarenvirta claims Daisy could double retailers' profit-margins if used regularly. In insurance, Daisy operates to detect fraud and provide underwriting recommendations. By identifying suspicious claims, or those that are vastly different than the average claim of its type, the machine signals a human to investigate further.

"We're saving them time and giving them new insight they've never had before," Saarenvirta said. "Our machine, unlike human beings that get tired and bored, puts as much effort into the 199th item as the   first item. Daisy doesn't get tired [or] bored."

Chime – AI for real estate

Any line of work that operates strongly on referrals can be difficult to grow; converting the lead into actual business is never a given. Chime is aiming to use machine-learning to connect real estate agents with leads that are more likely to pay off, and then streamline the transaction paperwork afterward.

The idea behind Chime was to "bring the real estate industry into the next generation," Matt Murphy, co-founder of the company, said. Between scouring the web for information while developing the leads and then triggering automated ad campaigns to make first contact with prospective buyers or sellers, Chime attempts to help agents take the first steps toward cultivating a promising lead.

"We created an intelligent lead-routing system, so if someone comes in looking in a certain area, at a certain income range, with certain housing requirements, we can send that lead to the appropriate agent," Murphy told Business News Daily. "If that agent doesn't take action, it will just round-robin it to someone else and [notify the manager]. There's a lot of intelligence there to optimize workflows in the agents' lives to help increase productivity and minimize time spent on mundane tasks."

Chime has been deployed for only several months, but there are already hundreds of clients using the system, Murphy said. So far, the feedback has been positive: conversions on Chime leads are significantly higher than the industry average.

"Top agents on the platform are converting at 8 to 9 percent, [which] … shows the machine-learning and intelligence is working to direct higher quality leads," Murphy said. "Of course, it takes agents to actually develop the relationship, but the combination of a great agent and great technology shows that it's working."

Restless Bandit – Human resources and hiring assistance

Have you ever brought a candidate in for an interview and then realized they were a terrible fit? Restless Bandit is a SaaS product that automates the candidate-screening process and even reaches out to qualified people for vacant positions. By mining the company's existing talent pool, Restless Bandit aims to reduce the work an organization needs to do before bringing a qualified candidate in for a face-to- face interview.

"It will make recruiters more efficient," Steve Goodman, co-founder and CEO of Restless Bandit, said. "They won't have to weed through the 80 percent of candidates that are clearly not a good fit. It will also increase quality [of the candidates] because the only candidates we reach out to are at least minimally qualified."

Using artificial intelligence, Restless Bandit finds, matches and ranks qualified candidates against the open positions in your organization. Recruiters and hiring managers are sent daily lists of matches and rankings so they can determine with whom they'd like to follow up for each open position. Goodman says the system will not only save your recruitment staff time and enhance their productivity, but will help deliver higher quality talent to your company as well. A subscription to Restless Bandit costs between $25,000 to $100,000.

"The human evaluator brings an unconscious bias to their work," Goodman said. "These algorithms look at the entire resume and job description as part of a larger model, and then tease out fits based on career arc, title climbing, industry associations, competitor employers, schools and degrees that really matter for the job, all while never bringing a job seeker's ethnicity, name, religion or location into play to determine value."

OutsideIQ – Automated regulatory compliance

The amount of time and money spent on filing the proper reports and paperwork with the appropriate levels of government can seem insurmountable and, perhaps, even prohibitive to many companies. However, failing to do so can lead to untold amounts in fines and legal fees, making regulatory compliance a necessary evil. But OutsideIQ is aiming to change that by automating the entire compliance process through artificial intelligence.

"We have clients who have fallen further and further behind on the regulatory demands they're facing, and we have clients who have tried to do it manually with armies of people," Dan Adamson, CEO and founder of Outside IQ, told Business News Daily. "It's a great relief to them that this system can make it doable in an efficient manner."

OutsideIQ's major products includes Due Diligence IQ (DDIQ), which automates due-diligence-reporting requirements for both clients and corporations, as well as use in anti-bribery and anti-corruption systems, Adamson said. InsuranceIQ, another product developed by OutsideIQ, is trained by commercial underwriters to look for the same things humans do when assessing risk; it then determines the appropriate level of coverage and offers a recommendation, and it's auditable, so humans can always confirm its methods. 

"We need the added context that AI can provide, but we have to be very open box; we have to explain everything that's been done with our system so regulators can understand, so banks can be confident with how it's being used," Adamson said. "Broadly, the exciting part is we're seeing actual commercialization of AI now. There's a real opportunity for AI to change the dynamics of humanity; it's very exciting and it's very humbling."

The rise of artificial intelligence in 6 charts

Artificial intelligence received $974m of funding as of June 2016 and this figure will only rise with the news that 2016 saw more AI patent applications than ever before

62 per cent of organisations will be using Artificial Intelligence (AI) by 2018 according to Narrative Science. At present, predictive analytics is the most used form of AI in enterprise and companies are focusing on innovation, patenting their AI developments at a faster rate than ever before.

Join us as we explore the rise of artificial intelligence in six charts including the top investors in AI and the most used AI enterprise solutions.

As of June 2016, artificial intelligence received $974m of funding. This year’s funding is set to surpass 2015’s total and CB Insights suggests that 200 AI-focused companies have raised nearly $1.5 billion in equity funding.

AI isn’t limited to the business sphere, in fact the personal robot market, including ‘care-bots’, could reach $17.4bn by 2020. Care-bots could prove to be a fantastic solution as the world’s populations see an exponential rise in elderly people. Japan is leading the way with a third of government budget on robots devoted to the elderly.

Fujitsu holds the most Artificial Intelligence patents, its human-centric AI Zinrai incorporates sensory-media technologies that can process people’s feelings, knowledge processing that can provide medical decision-making support and mathematic technology that is able to assist in mitigating air traffic.

IBM comes in close second with 88 AI patents. IBM inventors have developed new technologies that can help machines learn, reason, and process diverse data types, for example, IBM Research’s China lab patented a system that helps machines interpret emotion-laden words so they can converse in a more natural manner.

RELATED ARTICLES:
AI could rescue failing cyber security sector
Rise of the machines
Investing in artificial intelligence - is it the right time to buy?

Accel, a venture capital firm headquartered in Palo Alto, announced a $500 million fund for early stage companies and $1.5billion fund for late stage businesses. Of their nine areas of focus for the future, Artificial Intelligence topped the list of priorities.

New Enterprise Associates, also a worldwide venture capital firm, is the second top investor in Artificial Intelligence. The firm has invested in WellTok’s CafeWell.com website which combines social networking technology with knowledge of healthcare, making recommendations based on health goals allowing users to take control of their wellbeing.

The most AI based venture capital deals were in deep learning/machine learning. Naveen Rao, founder and Chief Executive of Nervana Systems explains, “the method allows computers to process tasks through ‘neural networks’, systems that mimic the structure of the human brain in order to learn, improve and, in the case of AlphaGo, become capable of decision-making that resembles human intuition more than it does brute computation.”

Programs that rely on deep learning are commonplace, for example Facebook’s newsfeed relies on deep learning to organise the news that is of most relevance to you personally and auto-tagging also uses this type of AI.

Narrative Science’s survey of over 200 CEOs, CIOs, CTOs and data scientists revealed that voice recognition and response solutions are the most widely used AI solutions in business followed by machine-learning.

Virtual personal assistants were used by 15 per cent of those surveyed. Virtual assistants, a standard feature on smartphones including Google Now and Apple Siri, are a one-stop solutions provider for business leaders and consumers alike with the ability to set up meetings, answer web searches and suggest the least congested travel routes.

Read more from our special report on Artificial Intelligence for Business