curiosity artificial intelligence (curiosity AI)
Curiosity artificial intelligence (curiosity AI) is the simulation of human curiosity in artificial intelligence. Curiosity AI is also known as artificial curiosity, AI curiosity, curious algorithm and algorithmic curiosity.
Curiosity is what drives most self-directed learning in humans. When we encounter a gap in our knowledge, our interest may be sparked, creating a desire to seek out the missing information. Emulating the behavior of a curious human in an algorithm could enhance the potential for self-directed machine learning so that an AI system would be driven to seek out or develop solutions to unfamiliar problems.
Narrow AI vs. curiosity AI
The capacity to emulate human curiosity is a defining component of strong or general-purpose AI, which more closely replicates human intelligence than narrow AI systems.
Narrow AI systems, also known as weak AI, are often capable of outperforming humans at their particular tasks. ROSS, an expert system sometimes referred to as an AI lawyer, can take over many tasks of a human legal assistant, many of them at a level far beyond what a human is capable of. ROSS -- a narrow AI -- can mine data from about a billion text documents. ROSS analyzes information to provide precise responses to complicated questions in less than three seconds.
Narrow AI systems are limited, however, by the fact that they require specific instructions and lack the human capacity to independently develop new approaches to novel problems, which tend to stop them in their tracks.
Curiosity can help AI systems behave more like humans in new situations by incorporating behaviors associated with curiosity into algorithmic models. A curiosity AI system might, for example, prioritize exploration by reinforcing behavior that yielded new information about its environment. Behaviors supporting the ability to explore can be reinforced and those that prevent or limit exploration can be punished.
How does curiosity AI work?
Reinforcement learning (RL) is the process of motivating AI to perform desired behaviors and punishing it for undesired ones. It seeks observations and experiences that will provide a programmatic reward, a feedback signal that informs the AI that it is making a step toward its goal. It is a carrot and stick approach to training the AI with either positive or negative feedback.
RL is used in many AI applications currently, from robotic autonomous arms programs that can play a video game better than human world champions. However, RL has its limitations. It needs a lot of feedback to work properly -- when it has little feedback to go on, it struggles. The goal of curiosity AI is to solve this problem -- to keep the AI searching when there is no clear sign indicating what to do.
A curiosity AI does well in situations where not much feedback is given to the AI on its progress, which mirrors a lot of real-world scenarios. In Google's latest study of curiosity AI, called "Curiosity and Procrastination in Reinforcement Learning," researchers used a variation of the following example to explain curiosity.
A person is walking around the supermarket, looking for spinach dip. At every step, the person passes several aisles that show no indication of having spinach dip. If the person operated on a reinforcement learning logic, they would walk in circles re-observing the same set of aisles, operating purely off of memory and getting nowhere. Introduce curiosity to the person's way of thinking and they are driven to recognize that they are walking in circles and arbitrarily choose a new direction that might provide an indication of where the spinach dip eventually will be found. A curious person, and similarly a curiosity AI, has the ability to recognize that redundant pattern and find a new circle to walk in that will have different stimuli and perhaps a route to the desired goal. Reinforcement learning purely works with what it knows, whereas a curiosity AI seeks to find environments that it doesn't know yet.
To get the AI to experience new things in its environment, to be curious, the RL algorithm is modified to work this way:
- The algorithm adds observations to memory.
- The algorithm computes a reward based on how far its current observation is from the most similar memory it has.
- The algorithm gets a larger reward for observing things that are not yet represented in memory.
By structuring the algorithm this way, it is able to avoid being trapped in a feedback loop, where all inputs are equally goal-related and the algorithm cannot differentiate between them. The curiosity AI gets two types of rewards:
- Goal-related observations. It is rewarded when it can compare current observations against memory and find goal-related stimuli, like a classic RL.
- Novel observations. It is rewarded more for finding novel environments that do not yet exist in memory.
Both rewards can be combined as well, so that the biggest reward a curiosity AI can receive is when observing something that is both completely new and goal-related.
Why is curiosity AI important?
Experts estimate that the majority of all data produced in 2020 will be generated directly by machines. With such a large volume of data, advanced artificial intelligence will be necessary on the processing side as well as the data-generation side. Curiosity AI ideally possesses the large capacity for data that current AI has, but with the added ability of sifting through it and creating relevant patterns within it when there aren't many goal-related indicators to do so. In this way, it can hold and look at more data than a human is capable of, while being able to view that data in a more human-like, critical way.
Applications of curiosity AI
Curiosity AI is currently used in a number of business automation areas, including:
- Data analysis
- Human resources
- Collaboration and productivity tools
- Customer service tools
Additionally, curiosity AI is expected to improve industrial automation in a number of areas, including:
- Supply chain management
- Factory and warehouse process optimization
- Quality control
- Fault detection
- Predictive maintenance
Advanced AI currently sees less use in the industrial sector than in business, partially because the margin for error is much smaller than in a typical business application, and the data that automated devices handle in an industrial setting is far more complicated and harder to process.
However, curiosity AI is expected to improve industrial automation because it is capable of handling more complex inputs with a consistently high level of success required in industry. It can handle uncertainty and can generalize and explain complex data in a way that narrow AI cannot.
Industrial machines are also not typically designed with usability in mind, so curiosity AI could enable them to be more flexible without a human needing to get involved.
Some potential industrial use cases include:
- Data center energy and air conditioning management. A curiosity AI could predict optimal cooling and maintenance times and react to environmental changes. It could also help recycle waste heat and switch to renewable resources in the data center.
- Manufacturing lines. A curiosity AI would perform anomaly detection and root cause analysis to identify problems and suggest potential solutions. This would reduce unexpected failures and increase the amount of uptime for the manufacturing line.
- Data processing and edge computing. Many industrial systems utilize internet of things (IoT) devices like sensors that collect complex physical sense data, and curiosity AI would provide flexible and powerful processing tools on those edge devices.
Some other use cases in addition to industrial processes include:
- Self-driving cars. These vehicles need to make high-level decisions to control a vehicle and react to unpredictable situations on public roads.
- Knowledge work. Knowledge work cannot be automated by narrow AI to a satisfactory level in many cases, but curiosity AI would be able to create an artificial intuition that could change that.
- Medical. Curiosity AI could help healthcare providers develop an automated medical suggestion tool that is able to use its problem-solving ability to recommend medical solutions and diagnoses based on each patient's individual unique way of living and set of healthcare specifics. One project that aims to do this is Google's Project Nightingale.
Curious AI company
Curious AI is also the name of a Helsinki-based organization that builds curiosity AI tools for control and optimization of complex industrial processes. They describe one of their core products -- Curious Engine -- as a "toolbox for deep-net modeling, systems control and model-based optimization."
They also are invested in providing "digital coworkers," which can be thought of like chatbots but with enhanced problem-solving skills. The company describes the difference between the two terms system 1 and system 2 thinking, which is another way of viewing the difference between narrow and curiosity AI:
- System 1 thinking is procedural, process and rote-task oriented. In the context of a chatbot, a system 1 thinker might listen to, respond and think of solutions based on the responses it is capable of or have worked from in the past.
- System 2 thinking is more focused on pattern recognition and high-level problem-solving -- the ability to understand and use novel situations practically. In the context of a chatbot, a system 2 thinker can think more critically about the way the user communicates and develop its answer to suit the novel communication style of a given user. It can also use curiosity to venture new methods of communicating.
Sentian.AI is a company in the same field -- they create industrial curiosity AI solutions.