Do not be afraid of machines - AI technologies are not yet ready to think independently
Despite Ilon Mask’s warnings
made by him this summer, we don’t have too many reasons to lose sleep, worrying about Skynet and terminators. Artificial Intelligence (AI) is still far from being transformed into a malicious and omniscient force. The only apocalypse that shines for us is the too-strong hope of people for machine learning and expert systems, as tesla owners, who removed their hands from the steering wheel, proved.
Examples of what today they are trying to pass off as AI — technologies such as expert systems and machine learning — are great for creating software that can help in the areas of sequence recognition, automatic decision making, and human-computer communication. These technologies have existed for several decades, and the quality of their work depends on the quality of the input data. Therefore, it is unlikely that in the near future, AI will replace people's opinions on important issues that require a more complex answer than “yes” or “no”.
Expert systems, also known as rule-based or knowledge-based systems, are computers programmed using explicit rules developed by human experts. Computers can apply these rules, only much faster and without interruption, to come to the same conclusions as human experts. Imagine asking an oncologist how she diagnoses cancer, and then programmed the medical software to go through the same steps. In the case of a specific diagnosis, an oncologist can study which of the rules were activated during the process of setting it, in order to confirm the correctness of the expert system's work.
However, to create and maintain a system of such rules requires a lot of time and special knowledge, and the correctness of the work of too complex expert systems is difficult to confirm. And, of course, such systems do not work outside their rules.
A horse who knows one trick
Machine learning (MO) allows computers to come to a solution - but without prior explicit programming. Instead, they are shown hundreds or thousands of examples of data sets, and are told how they can be broken down into categories: “cancer / not cancer”, “stage 1 / stage 2 / stage 3”.
Sophisticated algorithms train on these data sets and learn how to make the correct diagnoses. Machine learning can be trained on data sets, even if the human expert cannot explain how the decision was made. By increasing the quantity and quality of data collected by various organizations, machine learning promotes AI technologies in an ever-expanding set of application applications that promise to revolutionize the industry - if used correctly and wisely.
But MO has inherent weaknesses. For example, it is impossible to carry out reverse engineering
with the algorithm. One cannot ask how a definite diagnosis was made. And it will not work to ask the MO in the area in which she did not train.
A classic example of MO is the classification of images by the types of "cat / dog / both / none of them." After this treatment, you can’t ask the MO if there is a poodle or a dog in the picture - it cannot adapt to new questions without training or adding another MO level to the system.
AI technologies, when viewed as automation, can add much to business productivity. In some problem areas, the AI works great, especially where the required solution is fairly straightforward, and does not contain too many nuances.
Here begins to emerge pattern
One of the most commonly used types of MOs is pattern recognition based on clustering and data categorization. Amazon users have already felt how MO-based analytics can be used in sales: Amazon's recommendation engine uses clustering based on customer purchases and other data, identifying products that might interest him.
This kind of analytics has been used in physical retail stores for years - some grocery stores place products on the display case next to the most frequently purchased items. But MO can automate such tasks in almost real time.
MO perfectly works with the recognition of all laws - in medical image processing, financial services (is it a fraudulent credit card transaction?), And even in IT management (if server load is too high, try these settings until the problem disappears).
This kind of data-based automation is used outside the retail world to work with other routine tasks. For example, the Apstra startup has tools that use MO and real-time analytics to automatically adjust and optimize the performance of data centers, which not only reduces the need for administrative staff, but also reduces the need to upgrade equipment.
Another startup, Respond Software
, offers expert systems that can be used in corporate security centers to automatically recognize and deal with security incidents. And Darktrace
uses MO to identify suspicious activity on the networks — the Enterprise Immune System is looking for actions that do not coincide with those that were carried out on the network before, and warns security personnel about things that might interest them. The Antigena module can automate the response to detected problems by disconnecting network connections that are similar to malicious ones.
MO can be used to analyze more human communication. After extensive work carried out by data processing specialists and developers, MO algorithms have relatively well learned to recognize the “mood” of the text
- to determine whether it sounds positive or negative. It began to apply to the "text processing" of social networks
and image processing.
Microsoft Project Oxford has
created an API for checking the emotions of people in photos, and also created an API for word processing that identifies its mood. IBM Watson also works with this analysis in its Tone Analyzer
product, which is able to rate tweets, emails and other texts by emotionality.
Such technologies are integrated into user experience systems that define user complaints about products and services and send information to people for a response. IBM collaborated with Genesys to integrate its Watson system into the Customer Experience Platform
to allow people to answer their questions directly and connect people with complaints right away with those experts who can best answer them. The system has to learn from people on the fly, but it is constantly improving the quality of its answers - although its effectiveness is still to be tested.
Even the perennial field of human activity, human resource management, benefits from AI in measuring employee productivity and performance, evaluating performance and even introducing smart chatbots to help employees set vacation dates or express their concerns about managing spoken language. AI startups optimize HR routine operations: Butterfly
offers coaching and mentoring, Entelo
helps recruiters to browse social networks in search of candidates, Textio
helps to write more effective job descriptions.
But AI does not work well with uncertainty, including deviations in training data or expert rules. Different doctors can make different diagnoses or recommend different treatments. What should be done expert diagnostic system?
The widely discussed application of MO is the filtering of applications entering college. The AI was trained on data from several years of applications, including marks for academic performance, reviews from the school, and even essays, and told which students were accepted and which were refused.
The purpose of the experiment was to reproduce the work of people taking students - and the system worked, but it also reproduced their shortcomings, for example, a tendency toward certain racial groups, socio-economic classes, and even such activities as participation in sports teams. The result: technical success, and the rest - an epic failure.
Until there is a breakthrough in the processing of ambiguities or contradictions in the rules and aptitudes in the training data, the AI will be difficult to develop.
To improve performance, MO systems need training on good data. But in order to understand this data, people in many cases need to pre-process the information — assign the necessary metadata and formatting, and also direct the MO algorithms to the right parts of the data to get the best results.
Many of the advances in the application areas of MOs and AIs are due to the work done by human experts in many areas to provide better data in large volumes.
Low-cost historical satellite data and improvements in weather data
make it possible for MOs to predict crop problems in developing countries. Descartes Labs
, using data from the LANDSAT 8 satellites, build a 3.1 trillion pixel mosaic denoting the arable land of the planet and track plant growth. By combining this data with meteorological data, the company's DoD has been able to accurately predict US and soybean yields in individual counties. With the increase in low-cost satellite imagery and data from pervasive weather sensors, predictive systems will continue to improve their accuracy - with the help of data processing experts and other human experts.
Predictions of another variety may change the work of business enterprises. A recent study
from Naiang Technological University of Singapore showed that forecasts based on MOs using neural networks are able to more accurately predict production requests, which allows companies to better plan production than using expert systems or other predictive technologies based only on historical data. This is especially noticeable in industries with unstable demand, for which demand is either too high or too low, and very rarely - medium. Such systems can find patterns themselves, without prior explanations about the method of data modeling.
Such systems, becoming more complex and using more and more data types, can give enterprises and organizations the ability to find patterns in ever larger data sets. But if we can use AI to help make decisions in areas in which we already know what needs to be done, we cannot send AI agents into the unknown, without providing them with supervision from people who can draw up expert rules or create new training data sets from scratch.
And although some AI systems, such as IBM Watson or Amazon Alexa, can suck up huge amounts of unstructured data from the Internet and use them to search the text or build a knowledge base to answer questions, it will not help you create new training data sets for pattern recognition - for now. The sci-fi image of computers that independently find their own data sets is unattainable for today's AI — just like for tomorrow. Solutions, and questions, will continue to be created by people.Alan Zeychik - Head of the Camden Associates analytical group operating in Phoenix, pcs. Arizona. A former developer and systems analyst, he was the founder and chief editor of the Software Development Times magazine.