Author: СТРЕЖНЕВ МИХАИЛ АЛЕКСЕЕВИЧ | STREZHNEV MIKHAIL
Annotation
The use of quantum technologies in the future has a high potential for training neural networks, since machine learning is carried out much more efficiently by reducing resource costs.
Keywords: quantum computer, neural network, machine learning, artificial intelligence, qubit
Introduction
Various scientists have long had ideas about creating machines that could think like humans. There have been many successful attempts in history to make mechanisms copy human movements (wind-up dolls have existed since the 17th century), but they could not be made to “think”.
The first significant step on this path was taken in 1943, when Warren McCulloch and Walter Pitts published an article [1] in which they described the process of neural networks, as well as the structure of an artificial neuron (Fig. 1).
In 1958, Frank Rosenblatt created a single-layer perceptron [2] capable of recognizing some letters (this is considered the world's first neural network).
The topic of neural network development is of interest, in part because it is one of the most dynamic areas of technological progress. For example, the first neural network capable of effectively recognizing objects in a photo appeared only in 2012 [3] (AlexNet). Currently, this technology has achieved such success that there are many unmanned vehicles and delivery robots that use algorithms to identify objects around them. Moreover, the ChatGPT-3 natural language processing model even passed the Turing test [4], that is, neural networks are no longer inferior to humans in text generation.
This field of research was chosen, among other things, because we use the achievements of artificial intelligence every day (for example, FaceID technology or T9 in our smartphones). Moreover, this topic is directly related to the direction of study at the university: the design of neural networks.
In the process of designing neural networks, we face the problem that neural network training is a long and costly process.
Hypothesis: quantum computers can be used in the future to optimize the learning process of neural networks.
The purpose of the study is to identify ways to optimize the learning process of neural networks using quantum computers.
Research objectives:
to empirically show the relevance of the prospects for the use of artificial intelligence in the year 2100;
to identify the difficulties of organizing neural network training;
to study the current state of development of quantum computing technologies and uncover the potential of using quantum computers in the development of artificial intelligence.
Prospects for the use of neural networks
In order to empirically verify that the area of neural networks will really be relevant in the future, it was decided to create a model for segmenting birthmarks and malignant tumors in photographs. The daily use of this technology would reduce the time to detect dangerous formations on human skin and thereby possibly save his life, because it is very important to detect it in time.
For these purposes, the UNet architecture was used [5] (Fig. 2), since its analogues are successfully used in medicine. Moreover, from the point of view of deep learning engineering, it has more advantages than its predecessors (SegNet [6]).
The ADDI project photo database was used for training and testing the model. The code of the learning function is shown in Fig. 3.
As we can see, the loss function is decreasing. This means that the effectiveness of our model is growing. Figure 5 shows the results of segmentation of skin tumors performed by the model after training.
As we can see, the neural network successfully identifies the skin areas of interest to us (see the bottom row of images). It excludes random objects that get into the frame (for example, hair or the black edges of the lens in the corners of photos), and focuses directly on neoplasms.
The difficulties of organizing neural network training
In the process of creating the model, it turned out that the process of learning it is quite long. Taking into account the need to select the optimal hyperparameters, more than two hours were spent on training each time the program was launched.
For comparison, to train such large models as ChatGPT, it takes several months of continuous operation of tensor processors and more than $4 million [7], because the maintenance of electronics and energy costs are enormous. But not only training is resource-intensive, but also the further operation of neural networks. For example, for OpenAI (the ChatGPT developer company), this amounts to about 700 thousand dollars per day [8]. For several large companies (for example, Microsoft, Google, Adobe [9]) engaged in the development of artificial intelligence, this type of activity is unprofitable.
The potential of using quantum computers
in the development of artificial intelligence
Speaking in simple terms, such computers are based on the use of qubits [10] (quantum bit). In ordinary computers, bits are used, the value of which is always known: 0 or 1 (Fig. 6). In qubits, the principle of superposition is used - it is impossible to correctly find out what their value is at a certain point in time.
This approach has an important advantage in comparison with the calculations we are used to: you can iterate through many options at the same time. For example, this allows you to quickly study possible scenarios of drug-body interaction. Pharmaceutical companies are already developing this idea (in particular, Boehringer Ingelheim since 2021 [11]). Moreover, Volkswagen Group has introduced a traffic management system based on quantum computing [11], applying it in Beijing (2016) and Lisbon (2019).
The question arises: what is the obstacle to learning on quantum supercomputers of neural networks? In fact, qubits have a significant drawback: the state of qubits can be influenced by factors such as temperature, radiation, and even neighboring qubits – this is called the phenomenon of decoherence [12]. Because of this, a quantum processor requires a large refrigeration unit and reliable protection from external factors. Therefore, at the moment such technologies are very expensive, and this can be done mainly only by large IT companies, for example, IBM, Google, Honeywell, IonQ, PsiQuantum and D-Wave [13].
If we talk about the potential of their application, then at the moment these installations cannot be used for absolutely all computational tasks: their functionality is very limited. Nevertheless, they are extremely efficient – a 1000-qubit computer from D-Wave is 100 million times faster than conventional computers [14].
The experience of teaching machine learning models based on quantum computing is still small, but such experiments have also been conducted. A group of researchers from the Ural Federal University was able to train a neural network to effectively solve problems on a quantum computer [15]. Other scientists were able to find a way to apply the traditional clustering algorithm [16] (k-means is a well–known machine learning algorithm), and specialists from ETH Zurich and IBM Q [17] used a simple fully connected neural network on the Iris dataset and proved that learning on quantum computers is significantly faster (Fig. 7).
Conclusion
The relevance of the prospects for using artificial intelligence in the year 2100 is experimentally confirmed by the performance of a neural network trained to recognize neoplasms on the skin. The ideas of using neural networks are also confirmed by a number of scientific studies.
The organization of neural network training is accompanied by the following difficulties: duration, energy consumption and high cost.
The study of the current state of development of quantum computing technologies has shown their high efficiency compared to conventional computers, since the calculations themselves are performed much faster.
Currently, supercomputers are not used on a daily basis to train AI models, because it is expensive and technologically difficult, but the fact that it is possible, and in recent years humanity has made significant steps in this direction, suggests that in a few decades it will certainly become an integral and familiar part of people's lives.
Thus, it can be argued that the use of quantum computers in the development of artificial intelligence has high potential and is indeed feasible in practice in the foreseeable future.
The full document is attached