Thanks to recent advances in machine learning, artificial intelligence is currently rocking the markets as the most revolutionary technology of the fourth industrial revolution. AI, according to the business world, will revolutionize our world forever, and it has already done so in many ways.
In general, there are two types of machine learning: learning from precedent, or inductive learning, and deductive learning. Since the latter is commonly referred to the field of expert systems, the terms “machine learning” and “learning by precedent” can be considered synonymous. This method of training is currently popular, however expert systems are in crisis. The underlying knowledge bases are difficult to reconcile with the relational data model, so industrial DBMS cannot be effectively used to populate the knowledge bases of expert systems.
In turn, precedent-based learning is divided into three main types: supervised learning, unsupervised learning, and reinforcement learning. In addition to the above, other methods of learning are being developed: active, multitasking, multivariate, transfer, etc. Especially successful in recent years has been the development of “deep learning”, which can successfully combine learning algorithms with and without a teacher.
Supervised learning – This method of learning is used in cases when there are large amounts of data, for example, thousands of photos of pets with markers (tags, labels): this is a cat, and this is a dog. It is important to develop an algorithm that would allow a machine to identify whether a photo it had never seen before depicted a cat or a dog. In this situation, the “teacher” is the person who prepares the markers ahead of time. The machine itself chooses the signs by which it distinguishes cats from dogs.
Therefore, in the future, the found algorithm can be quickly reconfigured to solve another problem, for example, to recognize chickens and ducks. Again, the machine itself will do the difficult and painstaking work of extracting the features by which it will distinguish these birds. And a neural network that has been trained to recognize cats can quickly be taught to process CT scan results.
Unsupervised learning – Although there is quite a lot of labeled data, there is much more unlabeled data. These are images without captions, audio recordings without comments, texts without annotations. The task of the machine in unsupervised learning is to find connections between individual data, identify patterns, select patterns, organize data or describe their structure, and perform data classification.
Unsupervised learning is used in recommendation systems, for example, when an online store offers the buyer products that are more likely to interest him than others based on an analysis of previous purchases. Or when, after watching a video on the YouTube portal, the visitor is offered dozens of links to videos that are somewhat similar to the one they watched.
Reinforcement learning – Such learning is a special type of supervised learning, but the teacher, in this case, is the “environment”. The machine (often called an “agent” in this situation) has no prior information about the environment but has the ability to perform any action in it. The environment responds to these actions and thereby provides the agent with data that allows it to respond and learn. In fact, the agent and the environment form a feedback system.
Reinforcement learning is used to solve more complex problems than supervised and unsupervised learning. It’s used in robot navigation systems, for example, where robots learn to avoid accidents with obstacles through experience and feedback on each accident. Reinforcement learning is also used in logistics, scheduling and task planning, and in machine learning logic games.
Machine learning and, in particular, neural networks are advisable to use for solving business problems in cases when:
- A lot of different data has been accumulated, but there are no programs to process and systematize it;
- The available data are distorted, incomplete, or not systematic;
- Data are so different that it is difficult to identify the relationships and patterns that exist between them.
Business problems that can be solved by machine learning and neural networks:
- Forecasting: demand, sales volume, filling the warehouse, loading of equipment and other resources, further development of the enterprise, etc.
- Identification of: trends, hidden relationships, anomalies, repeating elements, etc.
- Identification of: photo, video, audio content, fraud attempts, lies, internal threats, external attacks on the security system, etc.
- Automation of: operators in online chats, phone operators, etc.
- Classification of: analysis of the composition of buyers, clients, customers, and their segmentation by various parameters.
- Clustering of: classification based on parameters that were unknown before.
- Development of: chatbots.
The latest research shows that 67% of CEOs see AI as a useful tool for automating processes and improving efficiency. Ordinary consumers view technology as a strong tool for improving social justice. “Over 40% of them believe AI will improve poor people’s access to the most vital services, such as medical, legal, and transportation,” say technical developers from EssayMap company.
The speed with which this remarkable transformation of automation processes took place could have been much faster, but it is being slowed by various difficulties. Despite so many benefits of machine learning in a business what are the main obstacles to the spread of AI and machine learning?
High prices
Machine learning and AI are becoming more widely used in a variety of industries. Partly, because these new technologies offer consistent savings by simplifying problem-solving, lowering operating costs, and improving process efficiency. However, new AI-based solutions are quite expensive. Prices range from $ 6,000 to $ 300,000 per year for fully customized solutions. Although the costs of innovative solutions are decreasing with time, they remain a substantial barrier for most businesses.
Most small and medium-sized businesses, in particular, lack the necessary working capital to cover the upfront costs of an AI-based solution.
Even obstacles in the form of indirect costs can slow down AI penetration into specific sectors. In many countries, for example, despite the incredible efficiency and innovative potential of AI-based services, the cost of unlimited mobile data remains rather exaggerated. The inevitable spread of 5G, on the other hand, may be able to solve this issue. As AI technologies become more efficient and widespread, their cost will become more and more affordable.
Lack of knowledge
Machine learning is a new and old technology at the same time. True experts in this discipline, with significant depth of expertise, are in short supply. Despite the fact that the number of AI majors has increased by 19%, the supply still falls short of the demand.
Many organizations are aware of their limits. No more than 20% believe that their own IT professionals have the knowledge necessary to work with AI.
Unavailable data and privacy protection
Before AI can learn anything from advanced machine learning algorithms, it needs to be saturated with data. A great deal of data.
However, in most cases, the data is not ready for use, especially unstructured data. Data aggregation is a time-consuming and complicated operation, especially when the data is kept separately or processed by several systems. All of this necessitates the involvement of a well-trained team of professionals from various fields.
Data extraction is often useless if it contains a large amount of confidential or personal information. Although categorizing or encrypting such data allows it to be utilized, these time-consuming procedures necessitate more time and resources. Confidential data that needs to be anonymized should be placed in a separate repository as soon as possible after collection to avoid this problem. When different types of data overlap, criminals can identify people. Anonymization remains a challenging task.
Confidence and credibility
Not everyone is flexible. If it is not possible to explain how the deep learning algorithm works in simple words to someone who is not a programmer or an engineer, the number of people interested in using AI to find new business prospects may begin to fall. This is particularly true for some of the older industries. In most circumstances, there is no previous data, thus the algorithm must be tested on real data to ensure its usefulness.
Many companies lagging behind in digital transformation may need to revolutionize their entire infrastructure to make meaningful use of AI. Visible results will take a long time to emerge, as the data must be collected, consumed, and digested before the experiment can bring results. Launching a large-scale machine learning project without a clear ROI necessitates a level of flexibility, resources, and boldness that many corporate leaders lack.
The first step should be to identify the owner of the machine learning project, who will lead its implementation in the company. Several established data and analytics teams simply divide their work into many smaller projects in businesses where numerous established data and analytics teams must synchronize their actions. Such pilot projects can contribute to a common understanding of machine learning science but often prevent the main business from achieving effective automation.
Conclusion
Surprisingly, many of the obstacles to AI development are caused by human nature and actions, rather than technological limits. There are no ready-made answers for those who still have doubts about the potential of machine learning. You must venture off the beaten path here, which necessitates experimentation during the development period.