14 major engineering problems that are related to human survival, which must be solved by AI

At Google I/O 2019, in addition to consumer-oriented feature improvements and developer tool innovations, Google also focused the conference on how to use AI to solve the thorny problems facing humanity.

The National Academy of Engineering, the highest academic organization in the U.S. engineering community, released the 14 major engineering challenges facing humanity in the 21st century in 2008. Jeff Dean, Google's scientific research leader, told Silicon Star that Google AI, which he leads, is working to address these challenges.

If the problems behind these challenges are not alleviated to a certain extent, it may mean that humans will not be able to continue healthily into the 22nd century. These problems are reflected all over the world and in all walks of life, but Google AI is trying to overcome them (or contribute to them) with its own expertise.

NAE lists a total of 14 engineering challenges. The ones marked in red are the ones that Google AI is helping to solve. Image/Jeff Dean

There is no doubt that what Google AI excels at is deep learning.

Why can deep learning be used to solve problems in energy, transportation, diagnosis, medicine, security, and many other areas, and become a versatile scientific exploration tool?

Jeff Dean believes that this is because deep learning can learn from raw, heterogeneous, and noisy data, and even if developers do not have special domain knowledge, they can develop neural networks that reach or even exceed the highest level in the field.

Moreover, machine learning technology is advancing with each passing day, with 90 papers published on ArXiv every day; new deep learning techniques are also emerging in an endless stream, enabling neural networks to master more and more complex capabilities that were previously recognized to be mastered by humans only.

Now, humanity has to deal with major engineering challenges and seek to solve them within this century. Jeff Dean believes that deep learning can be a great tool.

Restore & Improve Urban Infrastructure

Transportation is a critical part of urban infrastructure and one of the areas facing the greatest pressure as populations grow and urbanization increases. While society changes, roads often remain the same. As a result, transportation is particularly prominent in the engineering challenges of the 21st century.

Google's approach is to improve road traffic safety and efficiency, and the most effective way, besides putting down your phone while driving, may be autonomous driving.

Waymo's self-driving cars have been in development for nearly 10 years. As of last year, they have traveled a total of 8 million miles and achieved the lowest accident rate among all self-driving testers in the world.

Jeff Dean pointed out that deep learning is the hero of Google/Waymo's achievement, allowing the autonomous driving system to integrate and learn raw data from different sensors, draw high-precision maps, "understand" where surrounding vehicles, pedestrians and obstacles are, and even predict the direction of travel of other road participants, assisting autonomous vehicles in making decisions.

He introduced that Waymo now has more than 100 test self-driving cars in Arizona, which can pick up passengers to their destinations without a safety driver. Many industry insiders believe that in theory, if the proportion of self-driving cars in a region's total vehicles is higher, the accident rate in the region will be lower and the traffic efficiency will be higher.

In addition to autonomous driving, machine learning can also improve traffic efficiency in other ways. For example, in Southeast Asian countries and regions where motorcycles are popular, Google has added a "two-wheel mode" to map navigation, allowing the system to aggregate multiple data sources and recommend shortcuts and side roads for motorcycle drivers to avoid peak congestion.

Advance Health Informatics Using Deep Learning to Revolutionize Healthcare Informatics

As one of the complications of diabetes, diabetic retinopathy (DR) erodes patients. It usually takes 10 years for lesions to appear, leading to blindness. This disease is actually preventable, and experienced ophthalmologists can often observe the precursors through retinal fundus scans. However, taking India as an example, there is a shortage of about 120,000 ophthalmologists in the country, and DR patients often have no medical treatment, resulting in a large number of people losing their eyesight.

Dr. Lily Peng, a researcher at Google AI (formerly Google Research), led a team to build a DR prediction model based on a convolutional neural network, using scans labeled by ophthalmologists as training data.

Previously, PingWest had interviewed and reported on this technology. At that time, the model scored higher than humans in terms of sensitivity in detecting symptoms (98.8) and accuracy in judging symptoms (99.3) (in statistics, this score is called F-score, the score of ordinary ophthalmologists is 0.91, and the model is 0.95).

The good news is that Jeff Dean tells us that after two years of development, the new model has now gone a step further and scores on par with dedicated retinal ophthalmologists.

This is not the end. The potential of this technology is far more than just diagnosing DR. Jeff Dean revealed that Lily Peng's team has made more outstanding scientific achievements on this model. It is precisely because of the strong versatility of deep learning that they can now use the same fundus scan image to predict the incidence of gender, age, blood pressure, bone age and other diseases with extremely high accuracy.

This is a major breakthrough in medical informatics because it can supplement key information that cannot be obtained due to medical conditions. The short-term and direct effect is to provide more reference data for ophthalmologists' diagnosis and treatment recommendations, and in the long run, it can also predict and diagnose more diseases (such as cardiovascular disease) in advance. Although this is not a professional diagnosis, it is still enough to save ordinary people now and future patients 5 or even 10 years in advance.

General AI: Engineer the tools for Scientific Discovery

Many well-known scientific breakthroughs, such as penicillin and X-rays, have a certain degree of chance. Even so, people have never stopped trying to find a "perpetual motion machine" that allows scientific breakthroughs to continue to occur.

Jeff Dean pointed out that with more powerful computing power, deep learning can be more easily applied to more fields. Therefore, deep learning has the potential to become such a tool. Because as mentioned above, deep learning techniques emerge in an endless stream, allowing neural networks to master more and more complex abilities that were previously recognized to be mastered by humans.

TensorFlow is a representative example. This deep learning project led and open-sourced by Google is now being used in many industries such as agricultural planting and breeding, industrial production, the Internet, and medical finance, and continues to promote efficiency improvements in the three major industries. One example is a farm in Europe, where the farmer used TensorFlow to build breeding monitoring technology, using sensors such as cameras and motion capture to track and analyze the health status and movement trajectory of livestock at all times, significantly improving the market rate.

As for scientific breakthroughs, the aforementioned retinal fundus scan can also serve as an example.

Two years ago, Google announced the Neural Architecture Search (NAS)/AutoML technology, which can be likened to "designing and training neural networks with neural networks", and it has surpassed the results of hand-tuned neural networks in some fields including image recognition.

But now, Google AI is no longer satisfied with the achievements it has made. Jeff Dean said that they are thinking about a completely new form of neural network: a large model, but sparsely activated.

This new neural network may have more parameters than existing neural networks. However, when it processes different tasks, it only needs to activate a few nodes on the path, not all of them. The purpose of this design is to enable a neural network to perform a variety of different tasks - from hundreds to millions, so as to significantly reduce the amount of computation and time required to design, build and train the neural network, and achieve greater versatility.

Jeff Dean told Silicon Star that the new neural network he described is indeed somewhat similar to the "general AI" that people have been talking about but think will not be realized in the short term. However, he emphasized that Google AI's claim is that even in this new huge and sparsely activated network, training is still self-supervised.

In 2017, he and several colleagues (including top Google AI scholars such as Geoff Hinton and Quoc Le) and external research partners jointly submitted the first paper in this direction, titled "Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer", which presented a huge neural network architecture with more than 137 billion parameters and thousands of sub-networks. In scenarios such as language modeling and machine translation, it surpassed the current highest level of neural networks with less computing power.

Jeff Dean demonstrated Google AI's future vision for this technology: in addition to optimizing the network structure, Google may also develop new machine learning supercomputers optimized for this network structure (just like they designed TPU for TensorFlow). By then, the new computing paradigm will bring more help to Google AI in solving the great engineering challenges of the 21st century.

"Deep learning is helping us solve many major problems, and it will make great contributions to scientific breakthroughs and many areas of human development," said Jeff Dean.