The future of artificial intelligence

 

Writing

Kazi Iffat Haque

IBM's Deep Blue supercomputer once caused a stir by defeating the world chess champion. Recently, Google's Deep Mind Computer's Alfago program hit the headlines again, losing one of the best players in another complex board game, Gore. Excluding the game, we are searching for something with a picture on Google, typing something in Bengali and immediately reading its English translation, or giving voice commands to Siri on the iPhone, again the picture of the known people in the Facebook post is automatically tagged. Tell me where? These require intelligence. This intelligence is innate in human beings. We have to learn that language because we understand each other's oral language. Or when I recognize a picture, I learn it by looking at it and remember what the thing is. This process of human learning and understanding through thought, consciousness and experience is called cognitive process. At the root of it is our brain. When we are working on that cognitive process with the device, we are calling it artificial intelligence. The beginning of the concept of artificial intelligence or AI in the fifties. But until a few years ago, we didn't see much use of it in real life.

We now have a lot of data in our hands for the benefit of the internet. From digital processes, from every social media activity, lots of pictures, texts and many more kinds of data are being created. The availability of data has increased the importance of artificial intelligence. One of the earliest algorithms of artificial intelligence is the artificial neural network, which simulates human cognitive processes and creates models from existing data. Then that model applies to unknown data. This type of algorithm can be used for image recognition or text processing. For example, after creating a model from many familiar images, it can be used to identify new images. A neural network consists of several layers, through which various features of data are processed to form the model. The process of creating this model is called training. This requires a huge amount of data and complex computational algorithms. Layers were very limited in the early neural networks. Because, then the data was limited. Computers were not as readily available as running complex algorithms. There are many layers that need to be done properly to recognize the image. For example, the visual cortex of our brain processes the matter of sight থেকে from optical signals to different layers of neurons, the object is identified by processing various features like size, color, texture, orientation etc. And a lot of training data is needed. The more data can be trained, the more likely it is that the model will work properly. From that came the Deep Neural Network, which processes a lot of data for training through many layers. This process of modeling through deep neural networks is known as deep learning. Deep learning requires both data and computing power. In the age of big data, data is plentiful now. And it is possible to do complex calculations on this huge data using supercomputers. In 2011, work on the Google Brain Cat project was completed. Using deep learning algorithms, it has been able to accurately recognize pictures of cats by learning from 10 million images on YouTube. This is one of the outstanding achievements of Deep Learning. But it takes 2,000 CPU combined computing machines to finish this job! There are very few computers on such a large scale. So it is not possible to take full advantage of this huge potential of deep learning by relying on conventional computer hardware.

The same calculations are performed simultaneously on a lot of data in each layer of the deep neural network. We call this parallel processing. At one time, only supercomputers had the ability to perform such large-scale parallel processing. But the graphics processing unit (GPU) supercomputer has brought that computing power into our hands. The use of GPU as a computer graphics card has been around for a long time. GPU is a type of hardware accelerator that helps graphics process much faster. The GPU has several thousand parallel processing units to work on many pixels at once. So GPU can be used not only for graphics processing, but also for any parallel data processing work. And with the addition of a programming language (such as CUDA) to program GPUs, it has become easier to use GPUs for parallel processing applications. Realizing this potential of GPUs, scientists have already started using GPUs in biology or space research. But it was not until 2011 that GPUs began to be used for artificial intelligence. Following the completion of Google's 'CAT' project, its chief researcher, Andrew Ng, a professor at Stanford University, began working with GPU maker Nvidia to test the effectiveness of GPUs for deep learning. 2000 CPUs can be done with only 12 GPUs. The use of GPUs has brought deep learning to the forefront of computing. The use of deep learning in various applications began Artificial intelligence reached millions of people through Google, Facebook, Map and many more applications.

In addition to the GPU, work is also underway to make special hardware for deep learning. Google's Deep Mind computers use their own hardware accelerator, Tensor Processing Unit (TPU). TPU Google's machine learning framework ASIC chip specially made for tensor floor. Again, Microsoft is testing the use of FPGA for deep learning. However, Nvidia is one of the leading GPUs in terms of hardware for deep learning. GPUs typically operate on double-precision (64-bit) and single-precision (32-bit) data. Model training requires high precision computing. So conventional GPU will continue to be used in this case. Another use of deep learning is data intervention. It works by using a trained model, such as tagging a new image using a model made from tagged images. Low precision data for interface works. The advantage of low precision hardware is that computation can be done much faster and it is energy efficient. Energy saving hardware is important for cloud systems. So with the advancement of low precision hardware we will see more widespread use of deep learning applications in cloud systems. Google's TPU works on low precision data. Nvidia is also working on a 16-bit GPU design. The next challenge in deep learning is to run data interface applications on embedded devices, such as smartphones, tablets, cameras. With the advancement of deep learning software the importance of energy saving hardware design is therefore more in the days ahead.

 Author: Senior Chief Hardware Engineer, Oracle Corporation, USA