Original article can be found here (source): Artificial Intelligence on Medium
Trends in Learning Algorithms in AI
Hyejeong Jeon, Research Fellow, LG Electronics
Jongwoo Han, Research Fellow, LG Electronics
Ben Eum, Principal Director North Asia, Element AI
Artificial Intelligence is impacting the world that we live in. It can already be found across the systems, services and devices we use every day. In addition, while there have been exciting advancements in AI applications, AI community lacks a common framework and language for discussing future advancements across various domains, such as transportation, home, work, and public spaces.
For this reason, LG Electronics and Element AI collaboratively studied and developed the evolution levels of AI technologies to reflect the definitions of the progress in AI, and presented them at the LG Electronics Press Conference held at CES in 2020 . Not only do the AI evolution levels pay attention to the development of the technology itself, but also reflect changes in human experiences as technology advances. As technology evolves, the quality of life of users is also enhanced.
The evolution levels of AI technologies are based on insights into cutting-edge science, engineering of artificial intelligence and imaginations involved in research, and comprise of four clear levels. Each level represents both a step change in the potential aspect of AI-powered products and services that could provide new benefits to users and society. From a perspective that the future of AI should be directed toward human-centric designs, we have coined the term AIX — Artificial Intelligence Experience, which we hope will contribute to sharpening the discourse about the trend in learning algorithms.
2. Evolution Levels of Artificial Intelligence
Factories have been incorporating machines to augment or replace human physical labor for many decades, and changes in mechanical power have continually reshaped the physical layout of plants. For instance, the shift from group drive systems (where a centralized motor turned many linked machines) to unit drive systems (where individual electric motors ran individual tools) allowed production to be disaggregated into linear assembly lines in the 1920s.
Despite several decades of progress in creating increasingly complex machines, factories with low levels of automation are still utilized for products to move along the assembly line, and stationary operators are trained to deal with and work under the constraining aspects of such machines. In contrast, more autonomous robots with perception and mobility capabilities allow more flexibility. This enables spaces and processes to be optimized for human-machine collaboration rather than simple linear flows.
In home, users do not necessarily need to work at fixed locations while operating home appliances, but they have to consider the limitations of time and space related to the functions their appliances possess. For example, if an individual wants to cook, they should preheat an oven first. If a user wants to go out, they may need to wait for their washing machine to complete its task before leaving their clothes to dry. However, if AI can remember and consider the needs of its user, life can be more comfortable for individuals — e.g., enabling AI technology to preheat an oven automatically or wash and dry laundry by itself. If AI technology can also resolve errors, such as heavy or unbalanced laundry loads in a washer, which causes the machine to stop its task, then users will have more time to invest in their hobbies, families and other interests.
In the future, humans won’t have to adapt themselves to machines as artificial intelligence should be provided in products and services in ways to adapt machines to humans.
2.1. Level 1: Efficiency
At this level, AI enhances conveniences for users by operating systems and products based on predefined orders or conditions. Applying general machine learning algorithms, specific operations are activated when users give orders with voice, or when particular conditions are met. Most of existing AI driven products are equipped with the Level 1 technologies.
Level 1 AI air-conditioning can detect a person in a space by its sensors and release cold air toward him or her, and cool at a set temperature. The user is able to control the operation of an air-conditioner by voice.
When Level 1 AI is applied to a sprinkler operating on a regular basis, its sensors can recognize the fact that it has rained recently and skip the scheduled operation, ultimately saving water.
2.2. Level 2: Personalization
Level 2 AI can be ‘personalized’ so that it is capable of pattern learning through the accumulated interactions with users. This level of artificial intelligence analyzes user’s behaviors in the past and discovers patterns, thereby predicting behaviors in the future. Even when a product or service is shared by multiple users, it can analyze the voices, faces, and methods of use of each individual, and distinguish distinct patterns. By doing this, AI is able to recognize each user even when they do not log in.
For example, a refrigerator at this level understands what kind of food a certain user enjoyed in the past, and suggests recipes they may enjoy for meals. In contrast, a Level 1 refrigerator can only make a general suggestion when ordered to “recommend recipes for spicy food” by a user, regardless of prior interactions.
A smart mirror with Level 2 AI can display a user’s biometric data collected from a variety of sensors. In addition, if a user is taking medicine, it can remind him or her not to skip it.
2.3. Level 3: Reasoning
AI at the level of ‘reasoning’ can discover and recognize specific patterns and causes of behaviors that are shown during the use of a wide range of products and services through causality learning. Based on these findings, AI will work to serve what is in the best interest of its users, even in new situations. To provide customized services to users, it is essential to collect information through various points of interaction.
While Level 2 AI focuses on individual correlations between users and products, or users and services, Level 3 AI integrates the data collected from different products and services to figure out comprehensive causal relationships.
For example, when a user turns on a heater, takes out thick clothes from a wardrobe with sensors, and brews hot coffee, AI at a ‘reasoning level’ will understand that these actions are intended to raise the temperature of the environment or the user’s body. Later when the weather forecast says that it is getting colder, the AI can turn off air-conditioning, turn on a heating system, and then suggest to the user to wear warm clothes. Additionally, it can also ask the user whether they would like hot coffee, recalling the user’s previous behavior and preference in a colder climate.
2.4. Level 4: Exploration
At the ‘exploration’ level, AI can make the user’s life much more enriching through experimental learning, which is a process where AI reasons logically, formulates and verifies hypotheses, and discovers better solutions by itself. AI repeats meaningful experiments so that it can continue to come up with new ideas and acquire more knowledge, allowing novel information to further enhance a user’s life.
For example, when AI recognizes that humans can sleep well at 17 degrees Celsius, it can suggest to the user to turn on the ceiling fan to help circulate fresh air and maintain a comfortable body temperature and then ask the user whether they want the fan turned on during sleep.
On a larger scale, a smart city powered by Level 4 AI would optimize traffic systems by collecting information through vehicles and traffic sensors. Through the process of optimizing itself, a smart city could help its residents to live a more efficient and safer life.
3. Four Dimensions of AI Levels
To clarify what the differences between these evolution levels actually mean, the basic dimensions were invented to represent the distinct characteristics of each level. The ways which AI at each level can be safe and trustworthy were also considered. AI evolution along these dimensions is necessary to make sure that products and services can embody higher levels of AI technologies.
[Figure 6] illustrates the features of the dimensions at each level.
Users or operators can control devices either directly or indirectly. For example, a pilot using an autopilot function would set the desired altitude and the system would use onboard sensors and its own control algorithms to meet the specified parameters. A smart washer can select appropriate amount of water according to the level of cleaning predetermined by a user. In contrast, when it comes to a self-driving car, a passenger only sets a final destination while the vehicle will disassemble the objective into the sub goals that translate the goal into direct parameters, such as velocity and directions, and missions, like maintaining lanes. Similarly, a smart grid is able to make suggestions to minimize the frequency of complete or partial blackouts through a number of strategies to reduce energy consumption and facilitate efficient distribution and its own control algorithms to meet the specified parameters. Eliminating the need for the person to directly control all dimensions can be more efficient as it enables the user to easily program more complex tasks. More autonomy can be achieved by allowing users to provide higher-level, more abstract inputs as control, gain better understanding of the vagueness of the context, and translate it into contextual relevance for difficult compromises. Since an objective can be achieved through various plans and strategies with different amount of costs and risks, the highly advanced, fully-evolved AI should help a user achieve a broader objective exploring contradictions between short-term and long-term goals. People may have interests and influences that span months, years, or even longer, so AI would need to be able to take into account information that is relatively far in the past (like people’s childhood memories), or far in the future (such as future job prospects), to solve problems like providing truly personalized life-long education.
3.2. Environmental Awareness
The simplest work is the one that is done repeatedly in the unchanging environment. At this level, it is sufficient to build an AI system that is equipped with procedural knowledge of the task at hand and basic mapping of the environment. For example, a smart infrastructure recognizes that it has snowed and automatically heats up the road or sprinkles salt to improve the road conditions. However, one of the advantages that an AI system can offer is the ability to learn and adapt. An AI-powered system can be of more assistance if it learns that roads become more slippery when they are wet. For instance, if the road is wet because of the snow, AI system will understand it has become harder for vehicles to stop quickly and prolong the duration of yellow traffic lights.
As the systems evolve, AI can learn about the user’s environment as much as it can intervene successfully in the social or biological processes. For example, if a smart kitchen system can understand the processes like food spoilage and fermentation, it will be able to start unfreezing ingredients in a timely manner, automatically apply salt to preserve food, and even marinate meat and vegetables at the right time to help the meal preparation.
Achieving this type of adaptability requires a system to have a variety of sensors and power of understanding in relation to ordinary physical phenomena or internal models of causality. As the patterns found in social and biological systems are complicated, AI should evolve like children, who can generalize and reason with limited knowledge and experience, and apply the newly acquired knowledge to totally different areas.
At the most evolved level, this understanding plays a role merely as a foundation for more exploration and further development. Just like an honor student or a successful explorer, the highly advanced AI will strive to discover new insights to test what is understood and what is hypothesized.
3.3. User Understanding
Today AI-powered products and services are personalized in terms of behavioral patterns of people including what they liked in the past or what they are likely to click on. Humans can make a better choice through successful interactions due to their ability to infer from the internal states of others, such as what they think and feel, and how they behave. For example, once a driver notices that other drivers are careless at a certain intersection, he or she will be able to predict and take precaution at such locations. Likewise, AI can also evolve to make similar responses, not only reacting to behaviors of its users, but also considering their mental conditions and emotional moods. For instance, a cooking control system in the kitchen may notice that the user is distracted and is not aware of the hot cooking surface, so it could then give warning to the user to take precaution, or lower the temperature of the cooking surface.
To become a true companion capable of interpreting and understanding social relations, AI should be able to learn and comprehend the lessons that could be hidden deep inside. For instance, in some countries, proper business manners dictate that newly exchanged business cards should stay visible on the table during the meeting while in other regions, it is expected to put them aside immediately after the exchange. People do not score points or receive rewards for what they do, but they learn from others what the objectives of ‘the games’ really are, and more advanced AI needs to possess the ability to recognize those social nuances.
Some cases of such sophisticated AI include sending reminders to a small minority of population, such as dementia patients or the visually-impaired, to take their medicine. Other examples include AI that can provide services such as turn-by-turn directions in cities. For AI to perform more socially sophisticated tasks, users should believe that the suggestions and decisions of the AI take their best interests into account, and be given legitimate reasons to support a broader range of objectives.
For example, a user might be comfortable sharing food allergies with AI in order to receive safe recommendations of recipes and food items that they can consume. However, that same user may not be comfortable sharing food restrictions that are based on religious or cultural practices. Therefore, a highly evolved AI must be able to be more intuitive of human emotion and be able to understand, interpret the thoughts and feelings of the user and then communicate in a manner that will best persuade the user to consider its recommendations
While machines can become more capable and functional on their own, the discussions and benefits related to the specialization would mean that most homes, workplaces, and other spaces will be packed with numerous kinds of smart devices. At the same time, a number of tasks and missions would require coordination and adjustments across countless things and systems. For instance, keeping homes or offices safe would be a virtual-actual challenge related with the protection of various physical and digital assets and points of entry.
Moving from individual operation to a successful collaboration with machines or groups of machines, artificial intelligence should consider not only its own actions, but also how it can influence others to accomplish its missions. At this evolution level, it is essential to develop a system-wide understanding that can support a flexible communication method to deliver relevant concepts and orders, instead of using strict protocols.
In some cases, it will make tasks easier to cooperate with humans. For instance, it would reduce costs and promote efficiency for a swarm of less sophisticated AI-driven trucks to follow a human truck driver than operating fully-automated trucks. However, generally speaking, cooperating with humans and teams of machines would require a more evolved AI system. It is because AI should be able to observe and interpret teammates’ signals on their status, intentions, and goals to help manage attention and coordination costs within the team, as well as share learning objectives. The most evolved AI system would recognize the gaps amongst these shared abilities, and strive aggressively to find ways to make the group more effective.
4. Implications to AI Researches
The latest trends in the major theses related to AI evolution levels are as follows:
The research on the existing deep learning methods to improve the efficiency of Level 1 AI is mainly about hyper-parameter optimization to enhance recognition capabilities and increase calculation speed.
In the past, researchers themselves had to test and experiment with a number of different cases and discover optimal hyper-parameters to solve problems that arose. Recently, there has been a remarkable development in the field of Auto-ML. Particularly, if the methods of network architecture research, which employs reinforcement learning, are applied , it is possible to learn models with high efficiency and enhanced performance.
Achieving the Level 2 of personalization would require the transfer learning of the model learned through the entire datasets. In real life conditions, there are frequently situations that are not included in training data, so it is necessary to retrain with the data learned from the user’s actual environment. The learning across the two domains like this is called domain adaptation . In addition, since it is often impossible to gather various sets of data in user environment than in a training environment, it is important to develop learning methods, like meta-learning, which use only a small amount of data .
Reasoning at Level 3 is mainly about causal inference trying to find causes for certain consequences, and it is a topic that has drawn much attention in the field of data science. If a causality is clearly comprehended, it is possible, based on the causality, to easily predict what is happening in domains other than the training domain. Therefore, causal inference has been often applied in the field of machine learning ~, which has helped to improve the efficiency of domain adaptation and reinforcement learning.
There hasn’t been much research around the topic of exploration at Level 4. It is similar to active learning in that it collects and selects data that can be helpful to the current learning status, but it can only reuse the acquired training data, and unable to generate newly needed data. In model-based reinforcement learning , additional learning can be done either at random or with the data with much uncertainty, based on an assumption that the models are perfect, but its concept is not the same as the concept of causality-based exploration suggested at Level 4 as it is based on statistics such as Bayesian method . Rather, a theory of learning models that are not affected by adversarial attacks  is more similar in that it analyzes what makes the errors more serious and actually generates data for training. Further research could be directed at dramatically improving the efficiency of learning by making an assumption of ‘what if?’ along with the existing statistics-based research methods.
Unlike the previous wave of automation that reconstructed human space and workflow to integrate with machines, the true potential of AI, which would change world, is not to reduce the role of humans, but rather to make them become more attentive and spend more time on activities they want to do. For example, we believe that future products and services in home powered by artificial intelligence will not only reduce the amount of household chores, but will also help to improve their health, education, and entertainment in new ways. By creating devices, systems, spaces, and infrastructure that adapt to people and their needs, more and more sophisticated AI will lead to a new wave of productivity and well-being.
In light of this, if we examine the future research direction of AI learning algorithms, we can expect that there will be more opportunities to be applied to products and services. A number of papers have been published at leading international AI conferences to improve the performance of previous algorithms or to suggest new learning algorithms. The industry will analyze trends and levels of learning algorithms in terms of whether they are applicable to products and services, and whether they contain recipes that solve the difficult challenges. In the future, I would like to pay attention to the higher-level learning algorithms that can change the experience of customers and provide them with new values, from the perspective of AI evolution levels.