The Development of AI and Custom Models - Huawei Developers

Artificial intelligence (AI) and machine learning have become hot topics in recent years. Many predict that the mobile network era will ultimately lead to a future driven by AI. But where did AI come from? How does it work? And how can it improve our lives?
To get some answers, let's look back over the history of AI.
{
"lightbox_close": "Close",
"lightbox_next": "Next",
"lightbox_previous": "Previous",
"lightbox_error": "The requested content cannot be loaded. Please try again later.",
"lightbox_start_slideshow": "Start slideshow",
"lightbox_stop_slideshow": "Stop slideshow",
"lightbox_full_screen": "Full screen",
"lightbox_thumbnails": "Thumbnails",
"lightbox_download": "Download",
"lightbox_share": "Share",
"lightbox_zoom": "Zoom",
"lightbox_new_window": "New window",
"lightbox_toggle_sidebar": "Toggle sidebar"
}
AI has a long history. We can perhaps trace its origins to the scientist Alan Turing (1912–1954), who along with his contemporaries, attempted to solve complex tasks by simulating human consciousness and mentality. The famous Turing Test was created to test whether a machine was truly "intelligent".
Later, the computer was invented, and used to store and process data. This made the creation of AI a more tangible possibility. In 1956, at the Dartmouth Workshop, Marvin Minsky clearly defined what AI was for the first time, using the neural network as a data model, an idea which had been proposed by neuroscientists. At this workshop, he also improved the programming language used for AI, which made the technology even more tangible.
The neural network, which is the basis of human thinking, essentially relies on function and feedback between neurons. The question of how to simulate the human brain has long been a focus of AI experts. In 1958, a computer scientist named Frank Rosenblatt introduced an algorithm called a perceptron. This is the simplest form of neural network, consisting of just two neuron layers, and was used for the binary classification of data. The scientific world realized then that AI had a bright future. More and more people became aware of, and interested in developing, this technology.
However, AI still had a long way to go. In 1969, Minsky proved that the perceptron could only handle linear classification, and struggled to handle even the simplest XOR problems. This issue became a huge stumbling block for AI specialists at that time, and as a result, the potential of AI was not recognized by many, and the field stagnated for nearly 20 years.
In 1986, Geoffrey Hinton advocated the backward propagation of errors (or backpropagation for short). This method was useful for dealing with nonlinear classification and was widely applied in multi-layer neural network structures. It once more kindled some enthusiasm for deep learning. To obtain more precise results, increasing numbers of layers were added to the network structure. But this came at a price: deeper layers gradually lost the effective learning that shallower layers had achieved. Because of backpropagation's vanishing gradient problem, people chose to use shallow learning methods instead when solving real-life problems.
Deep learning did not find favor again until 2006, when Hinton proposed a solution to the vanishing gradient problem. Meanwhile, passion for deep learning began to spread from the academic community to industry. More and more companies and institutions started to apply it to things like voice recognition and image classification. It was in these sectors that deep learning began to show the significant advantages it has over traditional, shallow learning.
Since 2012, neural network structures and optimization algorithms have been emerging in quick succession. This has led to a dramatic improvement in the performance of deep learning. However, despite optimized algorithms and computing power, deep learning still puts many people off because training data requires long periods of time.
Another hurdle the technology had to overcome was the question of how to reuse acquired knowledge for sectors which had limited training data. Transfer learning was developed as a solution to this problem. With this method, knowledge gained during a task is transferred to a target task. Because this method effectively utilizes the model parameters it obtains, it requires much less time to train models. Transfer learning is now considered to be the future of AI algorithms.
Over the course of AI's long history, algorithms have been developed which have the ability to solve a wide range of complex problems. To truly unleash the power of AI, we need to minimize the cost of using deep learning neural networks, so we can resolve problems in different sectors.
It's highly possible that in the coming years, AI will profoundly change the world. In fact, the technology's integration with the Internet has already changed all aspects of our daily life. But AI will also transform all industrial sectors, including IoV (internet of vehicles), home appliances, health care, agriculture, and manufacture. To solve intricate tasks within these sectors, we need methods which can provide more accurate results.
This is where HUAWEI ML Kit's custom model service comes in. This service utilizes the transfer learning method to help you quickly customize your models. By simply providing it with a small amount of data for your chosen sector, you can obtain a model that is applicable to that sector.
Hopefully, in the future, we will be able to fully harness AI. Instead of being a tool which is useful to just a handful of people, it can be applied in all aspects of society, bringing us services which are intelligent and tailored to our needs.