[Editor’s Note] It has been 21 years since the man-machine war started in a flash. People expect ai to bring more possibilities for human development in various fields, rather than beating human beings in chess and Texas Hold ’em again and again. The author deep senior investment a queue in the field of artificial intelligence, applied artificial intelligence application into the physical world, the applications of the digital world, physical world of critical applications, the digital world of critical applications four categories, combined with the annual artificial intelligence (AAAI) on eight technical director of the artificial intelligence technology companies in their respective areas of technical progress carried on the thorough analysis, The difficulties in the practical application of artificial intelligence and the development opportunities for start-ups in related fields are pointed out.

Giiso Information, founded in 2013, is a leading technology provider in the field of “artificial intelligence + information” in China, with top technologies in big data mining, intelligent semantics, knowledge mapping and other fields. At the same time, its research and development products include information robot, editing robot, writing robot and other artificial intelligence products! With its strong technical strength, the company has received angel round investment at the beginning of its establishment, and received pre-A round investment of $5 million from GSR Venture Capital in August 2015.

This article is from “Silicon Valley spy”, author: AI serious said, 100 million euro forward, for the industry reference.

From the perspective of practical applications, critical applications are almost not allowed to occur errors, once the failure may cause personnel and property losses, so the reliability requirements of the overall system including hardware and software are very high, and the difficulty of implementation is also increased.

For the application of physical world, the system is required to be robust and able to deal with various uncertainties and complexities in the physical world.

Therefore, the application of ARTIFICIAL intelligence can be divided into four categories from the two dimensions of digital/physical world and critical/non-critical applications:

The first category is critical applications that take place in the physical world, such as driverless cars, where human life matters.

The second category is critical applications that occur in the digital world, such as problems involving finance and computer security, which may directly cause property losses.

The third category is non-critical applications that take place in the physical world, such as sweeping robots.

The fourth category is non-critical applications that take place in the digital world, such as recommendation systems. In terms of business application routes, the general rule is to start from non-critical applications in the digital world and gradually penetrate to critical applications in the physical world.

Critical applications that take place in the physical world

In general, critical applications that take place in the physical world are technically difficult, such as high-level autonomous driving, and are areas that require a long time to cultivate and wait.

One of the speakers at AI in Practice, Vincent Vanhoucke from Google and Dimitri Dolgov from Waymo, shared some of the lessons learned in developing robots and driverless cars.

Vincent’s Google Brain team is currently working in three areas: speech recognition, computer vision, and robotics. And Dimitri cites the California Department of Motor Vehicles (DMV) ‘s 2016 data on driverless miles and discommunication (when they require a human driver to drive) as a basic proof of how difficult it can be.

Google had the lowest failure rate, with one failure every 5,128 miles. You can imagine that Google’s test data was still in a limited environment, and Tesla’s failure rate was one failure every 3 miles.

Even with Google’s current data, driving at such high frequencies is still a long way from being fully autonomous (Lever4or5).

According to industry practice for a long time, it is often much harder to improve reliability from 90% to 99% than from 0% to 90%. However, it is much harder to improve reliability from 99% to 99.99% than from 90% to 99%, and we may require more than 99.9999% reliability for driverless vehicles.

High-level driverless cars are still a long way off, and the cycle from design to production will take an additional three to five years, which can be very long for startups in the industry. It’s worth noting, however, that driverless or assisted driving in limited Settings (such as highways) still makes sense.

The last part of Vincent Vanhoucke’s talk focuses on robots! He said that before he got into robotics, he would laugh at the various robot falls in the DARPA Challenge, and then stopped laughing after he got into the real thing. Machine learning researchers assume that robots are already using machine learning techniques on a large scale, that robot and environmental states are perfectly known, that samples are sufficient, and that computer simulations are close to the real physical world. From this perspective, the field of robotics offers many interesting topics for machine learning.

First: how to coordinate perception and execution is the key to robots.

Robot perception is software level, execution is mechanical level. Algorithms do not understand machinery, machinery do not understand software is often a common problem faced by the industry.

Second, how to improve the effective use of samples has once again become the core issue.

It is often very difficult to obtain training samples in the physical world. Take the experiment of random grasping of objects by robot arm as an example. In order to obtain training samples, Google can only obtain training data with 14 mechanical arms day and night. How to obtain samples efficiently, or how to use samples efficiently, will be the core problem.

Third, the field of robotics involves reinforcement learning, unsupervised learning and active learning.

As for intensive Learning, the core technology in the field of robotics, especially Deep Reinforce Learning, almost all the engineers we visited, including Vincent, agreed that the technical implementation is very difficult.

Fourth: closed-loop control systems are essential to improve performance.

Fifth, new data structures are needed to represent Kinematic chain, convolution of images and motion trajectories. In a question-and-answer session afterward, he had high hopes for the transfer.

The old IBM is diversified and mostly oriented to the enterprise services market (to B). Therefore, Michael Witpock proposes a more systematic and traditional approach to the field of artificial intelligence.

He refers to large-scale modeling of the world, moving from explicit, symbolic, and decomposed modeling to invisible, statistical modeling. For example, the previous modeling and solving of nonlinear variables such as friction in the dynamics equation of robots is not easy

IBM emphasizes the importance of symbolism and believes that knowledge expression and logic are very important in solving complex problems.

The traditional knowledge representation based on logic deserves to be Rethink.

IBM’s research advantage in this area is that it has both hardware and software. Throughout his presentation, Michael Witpock, a researcher at IBM, proudly described IBM’s past advances in ARTIFICIAL intelligence, which have been widely deployed in a variety of areas, including human resources.

In a domestic case, when talking about the application of ARTIFICIAL intelligence in the field of human resources, the head of HUMAN resources of a dairy giant dismissively talked about how human resources work is humane and how cold machines should deal with it. Isn’t hiring on LinkedIn part of human resources? The seemingly simple answer to how traditional industries face the progress of high technology, be subverted or actively integrate, is not easy to land on the ground.

It is worth noting that, as artificial intelligence applications invade traditional industries, it is often necessary to model the physical environment of the controlled object, which is a broader world than the Internet, with more opportunities, and of course, more difficult.

Non-critical applications that take place in the digital world

In terms of implementation difficulty, non-critical applications occurring in the digital world are most likely to occur. In fact, recommendation system is a good example. On the one hand, people are relatively tolerant of the accuracy of recommended products.

Non-critical applications that take place in the digital world are crowded with Internet giants, and it’s hard for startups to make a difference in this space, and there may be some opportunities in verticals. How to break through the limitations of talent, data and computing resources and find living space is worth further discussion.

Xavier Amatriain, appearing on behalf of Quora, answers exactly that question. Quora is a small and medium-sized start-up company and an American q&A website (similar to Zhihu in China). Quora has only 85 technical engineers, including just two researchers. There’s not enough talent, there’s not enough computing and storage resources, there’s not enough data,

How can startups avoid some technological detours and get ai right? Xavier summed up some of the lessons of his years of actually working on machine learning.

1. More data or better algorithms?

Xavier thinks better algorithms are more important;

For small companies, the amount of data is small, and the cost of acquiring labeled data is extra. Small companies tend to pile up more data than big companies, so it’s often more efficient to focus on optimizing algorithms than it is to focus on getting the data, which small companies also need to accumulate.

2. Complex model or simple model?

Xavier believes that model and feature selection need to match;

The more complex the model, the better. In startups, it doesn’t matter if a cat is black or white. According to the defined problem, the model matching the feature is selected.

3. When is supervised or unsupervised learning used?

Xavier argues that unsupervised learning can reduce dimensions and make engineering breakthroughs to features. In some cases, combining supervised and unsupervised learning works surprisingly well;

4. A combination of multiple algorithms or a single algorithm?

Xavier argues that it is important to use combinative algorithms as much as possible. Rather than focusing on original academic research, startups need to be “copyists”. It is wise to try different combinations of algorithms whenever possible to improve accuracy.

5. Do not take the output of one model as input to another system

Xavier warns that this could be a system design nightmare.

Critical applications occurring in the digital world and non-critical applications occurring in the physical world

Critical applications occurring in the digital world and non-critical applications occurring in the physical world are areas of opportunity for startups. For example, ai can be used in finance and security. Or sweeping or toy robots are typically non-critical applications that occur in the physical world. These are the two areas where startups have the most opportunity.

A disruptive theoretical breakthrough in ARTIFICIAL intelligence still has to wait

The search for new theoretical breakthroughs in ARTIFICIAL intelligence, through the intersection of neuroscience and other disciplines with computational science, remains at the theoretical stage.

Although deep learning has made a lot of progress, people still know it in many fields but do not know why. In fact, theoretical breakthroughs in artificial intelligence are still made.

For deep learning, now based on probability and mathematical statistics, Gary Marcus, a neuroscience professor at New York University, hopes to find a breakthrough in artificial intelligence from a neurobiological perspective. He just joined the new Uber AI Lab. We quote him here, out of context, that “the biggest worry about AI right now is that the technology is stagnating”! That’s what we’re worried about.

Giiso information, founded in 2013, is the first domestic high-tech enterprise focusing on the research and development of intelligent information processing technology and the development and operation of core software for writing robots. At the beginning of its establishment, the company received angel round investment, and in August 2015, GSR Venture Capital received $5 million pre-A round of investment.

In the wave of deep learning in recent years, advances in ARTIFICIAL intelligence have been more engineering advances than theoretical breakthroughs, especially massive data and super-scale brute force calculations. As Peter Norvig once remarked about Google’s remarkable performance in ARTIFICIAL intelligence, “We don’t have better algorithms, we just have more data.”

As for Artificial General Intelligence, Gary continues to criticize the past few decades of stagnation. Intelligence at this stage cannot read, understand and reason like a human, and driverless cars are not as safe to be believed…