Anima Anandkumaris a professor at the California Institute of Technology and director of research at
"The situation with the coronavirus shows that humans are far ahead of AI"
How did the artificial intelligence revolution begin?For several decades, interestingdiscoveries in this area. For a person, there is nothing more mundane than recognizing images on a screen. But for an AI, this is an extremely difficult task, because it is not born with already given abilities for this. Scientists were faced with the task of "teaching" the machine brain to identify what it saw. In the early days of research, a Stanford professor began capturing images to make it easier for a computer to classify them. It was labeled pictures that started the deep neural learning revolution.
Reaching a new level of AI development began with the inclusion of billions of parameters in the program that will allow it to recognize a specific object.The difficulty of the task is to make a deep neural networkunder conditions of invariance, she recognized, for example, a dog regardless of the difference in posture, color, breed, and so on. AI training means that during image processing, a number of layers of the frame are viewed in a hierarchical order. So, at the basic level, only lines are visible at different angles to each other. Then they combine and make up more complex shapes, for example, a dog's face in color. The shape, color and other individual characteristics are already distinguishable. It was the stage when we learned to divide the process into parts, into levels, that became a huge step forward.
Further development requires not only deep neural networks, but also huge computing power.Over the past decades we have seenthe slowdown in the growth of our computers' capabilities, with single-threaded computing speeds no longer doubling. But at the same time, we can now simultaneously calculate operations that are huge in volume and complexity. The current level of development of deep networks has led to the fact that billions of processes can now occur in them simultaneously. And their goal is only one: to determine what is shown in the photograph. This procedure is carried out using matrix animation processes and other technologies. And here, of course, everything depends on the power of the video cards.
By 2014, deep neural networks were able to recognize images better than humans, which means that today they have become even more advanced.This happened thanks to the combination of threefactors: the capabilities of labeled data, the flexibility of existing deep neural network algorithms and the enormous capabilities of computers. There are many great examples of what heights the machine brain has already achieved, for example, deep reinforcement learning helped AI beat a human in the game of Go. This player was one of the best, but he lost. In addition, artificial intelligence is now capable of generating photographs of people so realistic that this technology has already passed the Turing test. A person can no longer understand which image is real and which was created by a machine. These are great examples of AI success, but the progress certainly won’t stop there.
However, there are a number of examples where AI fell short of expectations.Let's say a situation where the key aspectis safety. For several decades now, we have seen an increase in the technology used in autonomous cars, but, unfortunately, their lack of sophistication still leads to accidents. Also, a machine cannot replace a person when it comes to content creation and editing. And the current coronavirus situation shows that humans are far ahead of AI.
What will we do in the future, developing artificial intelligence and technologies of deep neural networks?Currently actively developingrobotics: for example, one company has a robot that can do backflips. But he cannot even be compared to a dog. This robot is awkward, constantly falls, but the animal, while falling, learns how to make this or that movement next time without falling. But the robot is not capable of this, it is not trained. This raises the question, is it possible to make AI that is ready to learn and solve problems on its own?
"Understanding algorithms is beyond our capabilities."
An artificial intelligence algorithm is a collection of pre-built information and a very clear task.We determine what data and preliminarilythe given parameters should be used, as well as how to implement the decision-making process. Creating and maintaining an AI algorithm requires huge amounts of data. This is difficult because there is more and more data, for example, when recognizing a video stream, the number of recognized frames is very large. Labeling is problematic because people assign names to millions of videos, and therefore to objects on the screen.
As for the preset parameters, then again you have to deal with the fact that everything is done manually.It's easy to fool AI these days.For example, if we have a stop sign and we place several blocks on it, then the artificial intelligence will no longer understand its meaning. And a car moving without driver assistance no longer recognizes this sign as a call to stop. Our human intelligence is completely different. Maybe we can transfer our way of thinking into a computer, but we haven't been able to do that yet. As for directions, orders and instructions, we have to act very simply: we give one task - to recognize what is shown in this picture. As for the parameters for assessing the success or ineffectiveness of the algorithm, here we are very limited.
We sometimes cannot understand how successful the current algorithm is, because it is beyond our understanding.In addition, there are several problems associated withthe fact that the data we have is mainly on light-skinned men. For this reason, AI misidentifies dark-skinned women. There are other errors in facial recognition. The problem arises because evaluation of the effectiveness of artificial intelligence is extremely limited. We must not forget about the paradigm that says that we need huge amounts of data, and all of it must be labeled. The preset parameters must be clear to the algorithm, and the task itself must be simple and logical.
First, you need to ensure that you do not need to label the data.AI must work without human assistance tothe computer itself found concepts, formed ideas, and understood the features of a particular image. Is it difficult? Yes, very much, but people do it, and with ease. As for the pre-set data, here we need to create very clear images, to show what is what of the data that we “feed” to the system. And here you can learn a lot from the human brain. And finally, the tasks that we give to the algorithm. AI must be more adaptive, because now we train our system from scratch every time, but we need to make it so that it can adapt and change, perform different tasks. So now we are training to make artificial intelligence flexible.
How to understand that there is a cat in front of us?
We recognize a cat, even if it is a blurry picture, because our brain is constantly trying to give a certain sharpness to the blurry image for further analysis.There are many theories, and one of the most famoussays that we not only look at some object, but at the same time the brain selects options for what it can be. Deep neural networks do the same. We have a priori data on how the cat should look. And we are trying to match this picture with the idea of what cats look like. It is important to understand this during development to ensure consistency of image identification.
How can we achieve stability in the identification of objects by artificial networks?This naturally occurs due torepeatability. We take some external picture and look at it, and the signal goes to the brain. There is also top-down feedback. Using information about what a cat looks like, the brain forms a specific perception. How can we make these complex processes in our brain implemented by AI? It is necessary to combine a good classifier for the concept of "cat", which will "feed" the neural network, with an excellent generator of these images. At the same time, the classification of the concept and training of the neural network will take place. Feedback will be obtained for standard neural networks. And this connection will provide an opportunity to receive generative feedback. In other words, when trying to process the incoming signal, a person tries to mark the image. And then there is feedback when we try to generate a perception based on what we see. These two processes must be interconnected.
A standard neural network, as a rule, cannot recognize fuzzy images, but our model, thanks to a feedback mechanism, makes the images clearer and then can recognize them.We see that such a scheme has proven itseffectiveness, so we can be inspired by how humans see when creating computer vision. Based on already developed models, you can create more advanced algorithms that will have high performance. But we also need an efficient infrastructure that can handle AI processes at scale. We do not work with algorithms individually. You take some data and you need to visualize it, which is a very complex process. Therefore, you need a powerful processor capable of processing impressive amounts of information. In addition, we use certain frameworks (CLARA) for various applications, including medicine. Now, given the COVID-19 pandemic, there is a need to train machine learning models at scale. The goal of this is to develop vaccines and cures for the virus. The CLARA tool can work with three-dimensional structures and various algorithms, essentially being the coordinator of their work.
Another AI learning opportunity is to use stimulated data rather than real data.We have a range of robots that can becomechefs in our kitchens. Such machines are capable of opening and closing a drawer, taking an object, mixing or whipping something. These operations, which are simple for humans, are very difficult for robots, since it is problematic to train them to do this. But with the help of simulation processes we will be able to open non-existent, programmed boxes. And thus the robot learns similar operations. Programs allow us to do this in parallel and on a large scale, which allows us to overcome the limitations that data imposes on us. But such a training system means that very complex algorithms need to be developed that will take the machine from the world of simulation to the real world, opening up completely new, exciting prospects for working with AI. There is a program that allows you to add simulation to those models when artificial intelligence is trained on real data. This is another example that we have good infrastructure and can work on very complex problems. It became possible to create new algorithms and models, as well as test them much faster than previously done.
The future of AI must be comprehensive and be embodied in different spheres so that we have a highly adaptive, constantly learning tool.To do this, we now need to rethink approaches todeep learning. Self-directed learning is key to success, so we need to find ways to build unsupervised learning programs into systems. And if we talk about convolutional neural networks, the feedback system makes them more stable. And this is the first step towards creating a real base for the next generation of AI.
See also:
Treasure hunter finds 3000-year-old treasure in Scotland
Perseids Meteor Shower - 2020: Where to See It, Where to Look and How to Take a Photo
Look at the 3D map of the Universe: it was compiled for 20 years and it has already surprised scientists