“Google searches reflect the infinite curiosity of the human being,” he says Pandu Nayak. Indian origins, about sixty years old (how many exactly you don’t even find out by asking Google), he taught at Stanford University and worked for NASA. And today it is Vice President in charge of Search of Mountain View: “Ours is a simple job to explain, it’s about organizing all the information in the world and making it universally accessible and useful”, he told Italian Tech in this conversation exclusively for Italy.
Won’t that be too ambitious a goal?
“It is as if we wanted to compile an index of a book, but with important differences: meanwhile, the Web has billions and billions of pages, and even the index is monstrous, so much so that if it were printed it would cover 12 times the distance between the Earth and the Moon and back. Also, every day 15% of searches that come to Google are completely new, never done before. In addition, we also take into account information created by people, such as reviews and places of interest, news, data from companies and institutions, as well as those produced with machine learning, such as when a restaurant is crowded, or the weather forecast. We organize all this information and present it appropriately. Or at least we try “.
What is the most difficult challenge?
“The understanding of language. In order to answer, we first need to understand what the user is looking for. At first the searches were very simple, but one was recurring and it drove us crazy. Many wanted to know if sole was a suitable food for children, only that the answers provided by Google were all about shoe soles. In English the 2 words are spelled in the same way (sole, ed), so we learned that we need to consider the context and use. In 2018 we introduced Bert, a technology for natural language processing based on a neural network, and the improvements are evident: we are often able to process even the 10% of researches that have errors or are formulated in an unclear way “.
Artificial intelligences between racism and the ethical question
by Emanuele Capone
Meanwhile, the automatic transcription system we are using for the interview (Google Transcript), with a certain irony suggests a third possibility: instead of “sole”, sole, he writes “soul”, soul, opening new horizons of meaning. Engineer Nayak, how important is randomness in research?
“It helps us satisfy people’s curiosity. This is because Google does not provide a single result, but a list, and therefore allows you to explore the world from different perspectives and interpretations. For example, let’s say that someone is looking for information on the Taj Mahal: they will read about the Agra monument, but then by chance they may discover the link of a new Indian restaurant with the same name, which has recently opened in its neighborhood. They are very different things, yet they could arouse interest in the same person, and an opportunity for pleasure or knowledge could arise from chance ”.
But sometimes the search returns questionable results: like when, years ago, the Guardian noticed that at the top of the Holocaust results was a denier site.
“It was a mistake, and that’s not the case today. We take into account several parameters to establish the order of the results research, such as relevance and reliability, but also freshness and originality of the contents. We are constantly working to improve the algorithm: in 2020 alone, there were 4887 changes ”.
So 13 times a day?
“Yup. And we also conducted over 60,000 direct comparisons between the solution in use and the one to be introduced; worldwide, over 10 thousand validators they analyze how the changes affect the results and tell us whether to move forward or not ”.
Also in Italy?
“Yes, there are several hundred”.
YouTube has a new tool in the fight against fake news, hatred and conspiracy theories
Bruno Ruffilli
Does having more data in English affect the quality of searches in other languages, such as Italian or Polish?
“In principle no, because most of the techniques we use are based on statistics, therefore applicable all over the world. In addition to the similarities we study the differences between the various languages, such as compound words in German or the segmentation of Chinese terms. We train our artificial intelligence with machine learning methods, to grasp the specific nuances of various languages. Furthermore, we are beginning to create more powerful linguistic models, which can be transferred from one language to another ”.
Today you can search for an object or a monument by framing it with the smartphone camera, soon it will be possible to ask Google to find a precise moment of a video or to recognize a song by humming it. Is there a question destined to remain unanswered?
“I would say no, but in case we point out that we don’t have enough data and explain that further research is needed.”
So are we going to keep doing one search after another?
“An average of 8 searches are needed for a complex task such as preparing a trip. With the new Multitask Unified Model technology, only one will be enough to have all the useful information: Mum not only understands language, it generates it also. And it is multimodal, so in the future it will extend to audio and video, as well as to languages other than that of research ”.
Will Google give us the answer to the fundamental question about life, the universe and everything?
“The answer has already been given by Douglas Adams: it is 42. And obviously on Google there is”.
.