Metamind won 8 million dollars in investment gold, uncover the mystery of deep learning

Editor's note: The concept of deep learning stems from the study of artificial neural networks. As a kind of artificial intelligence, "depth learning" is a training system that can handle massive amounts of information from audio, images and other input signals, and if new information is presented to the system, it will respond in the form of inferences. Technology companies such as Google and Facebook have made technological advances and mergers and acquisitions in this area, and "deep learning" start-ups are springing up.

Richard Socher, a graduate student at Stanford University, created Metamind after graduating, and four months later won 8 million dollars from the venture capital Khosla Venture and Salesforce chief executive Marc Benioff.

(Metamind founder Richard Socher, and Sven Strohband in his office in California)

Richard Socher never thought he would get into the cutting edge of artificial intelligence, he just wanted to combine the math and language he liked.

But things happened one after another, and he developed an excellent technology "recursive neural network" (Recursive neural NX), and now he started the Enterprise Metamind from university officially launched, and won several well-known enterprises financial support.

Only four months after the company was founded, Socher and his team have sought to demonstrate that Metamind's ability to process images and text in "depth learning" is superior to any current technology. To this end, in addition to the announcement from Khosla Venture and Salesforce chief executive Marc Benioff to receive 8 million of dollars of funds, Metamind also on the official website demonstrated their various technical capabilities.

As a kind of artificial intelligence, "depth learning" is a training system that can handle massive amounts of information from audio, images and other input signals, and if new information is presented to the system, it will respond in the form of inferences. Technology companies such as Google and Facebook have made technological advances and mergers and acquisitions in this area, and "deep learning" start-ups are springing up.

But Socher believes people can appreciate its advantages when using Metamind, which has two core technologies, a "convolution neural network" developed by Professor Yann LeCun, a New York University hired by Facebook, which has made a breakthrough in image mining. There is Soher's own "recursive Neural network", which has achieved great success in text processing.

"We are at the forefront of this technology and have a higher level of openness with other technologies, and we want to use scientific language as much as possible," Socher said in an interview with technology blogger VentureBeat. ”

As simple as drag and drop

It's as simple as a mouse drag and drop, which means that almost anyone can now do "deep learning".

"You don't have to be a programmer," as Socher demonstrates, users can train metamind some simple text, and then receive several lines of code that can be embedded directly into the application without the need for a data center or even a public cloud like Amazon Web Services, Everything Metamind.

Metamind Web site There are many such demonstrations, some can point out how similar to the semantic two sentences, and some users search for a keyword to show the positive or negative degree of push text. Metamind can also form a classifier: the user uploads a spreadsheet with a text label so that the system knows what to search for and then gives it some text for dynamic analysis.

Metamind can also classify images as long as they are trained with a set of related pictures. When it "digests" some of the food pictures, you can drag a picture of a fish or potato chip and confidently say that it is a fish or potato chip.

Socher also shows how metamind extracts images that match the text entered in the TextBox.

Or enter a few words.

Socher the word "bird" (bird) with a laptop, and the system shows some images, each with a bird. Then he entered the plural form of the bird (birds), and the image changed, with many birds on each sheet. "The coolest thing is that it really has a sense of semantic synthesis--how words form the meaning of longer sentences," he says. He typed the bird on the water, and then the system gave a picture of the bird flying on the water.

This kind of work requires many types of "depth learning", "convolution neural network" after scanning a large number of images can extract features, and "Recursive neural network" can extract meaning from the sentence, the two technologies can work together. Google and Microsoft have recently announced that they have mastered the technology to handle text and images at once, but Socher completed the study independently last year and published two papers this February. "We've had this technology for months," he said. ”

Socher came a long way, he was a German, and during college he studied natural language processing (NLP), but he felt that there were too few of them, so he began to study computer vision at the graduate level, and although the math was too much, it was still imperfect because it was too easy. He then went to the United States to read a PhD and specialize in machine learning at Stanford University. There he listened to Professor Andrew Ng's report on "deep learning" and its application in computer vision.

"I think that's a great idea, but they're still not very good for natural language processing," Soher says. "I've created some ' deep learning ' new patterns that can be applied to natural language processing. ”

His "Recursive neural network" analyzes the connection between the two adjacent words, and then it analyzes the relationship between the two words and the word on their left, and so on, and so on, the word "recursive" means until it can understand the meaning of all the linguistic elements in a sentence.

He presented these models for the first time in 2011, sparking interest in academia, after which he published several papers proving the feasibility of a "recurrent neural network".

He had wanted to do research, but earlier this year he realized he didn't want to go this way.

Socher, a teaching assistant for more than 300 students at Stanford University's machine learning program, sees how much you want to apply this technology to all data types.

"There's a whole new project in a completely different area every five minutes," he said. "I like it, machine learning has a bright future, and its importance will be further revealed." ”

Over the years, he has turned down the job offers from big companies, and he wants to bring them into a wider area where more individuals and businesses can use them.

The establishment of the enterprise

He needed money to build a team to achieve these ideas, and finally he went to Khosla to see the company's chief technical officer, Sven Strohband. Now, Strohband has joined Metamind as CEO.

At the same time, Khosla founder Vinod Khosla (also co-founder of Sun Microsystems) as the start-up consultant, Salesforce's chief executive Benioff and the University of Montreal leader in the field Yoshua Bengio joined the company's top brass.

Metamind has set up a 10-person team and has started to attract paid subscribers, who serve small businesses and work for big Fortune 500 companies. They provide licenses for Metamind systems running in the Enterprise data center, as well as additional consulting services for companies using metamind power systems.

Specific business includes the extraction of hidden keywords in financial Analysis reports, or analysis of people seeking customer service help chat records. More extensive applications and according to the X-ray prediction of disease and so on.

Although Metamind already has many of these applications, they are still recruiting new projects, showing an open mind, in order to let the world tell them what is the best option, which is right for a start-up.

Strohband said: "We believe it should be used by more people because we think it has many uses, and frankly we really can't predict what people will do with it." "(Translate | Tracey)

(Responsible editor: Mengyishan)