Self-aware AI

Self aware AI can simply be defined as a machine or robot that has similar cognitive abilities to the human mind. Scientists are working on programs that model the human mind and its ability to attach meaning and understanding to objects, situations, and things. Instead of a computer organizing information and spitting out an answer, individuals hope to create a robot that is able to attach meaning to its actions rather than just computing information and presenting it. But how exactly does this work and what programs are being worked on now? To start, two particularly relevant programs are being worked on and are being modeled after psychological principles. The first long-term project is called the SOAR architecture. This stands for State, Operator And Result, which was devised by Lehman et al. (2006). The second long-term program is the ACT-R architecture. This stands for Adaptive Control of Thought-Rational, which was introduced by Anderson et al. (2004) ( Chatila R et al., 2018).

SOAR

SOAR is modeled after human cognitive behavior. Every behavior we take has some form of thought process or logic behind it. It explains AI in terms of learning regarding chunking and reinforcement learning. Chunking is when we take small pieces of information and form it into groups to better comprehend it. A good example of this is remembering phone numbers due to the dashes. Reinforcement learning is used in AI in which they are able to learn new information without being programmed to know it.

Figure 1.

This is flow chart with icons and arrows pointing to the right. The first icon is of an apple (the input) with a question mark. The next icon is of a computer processing a response and a thought bubble with the words "It's a mango." The next icon is of a person next to a easel giving feedback, their thought bubble says "Wrong! Its an apple." The next icon is of the computer and has the word Learns underneath, with a thought bubble above that has the word "Noted!" The final icon is titled Reinforcement Response and has the original icon of the apple below the computer icon, indicating a new input. The computer, having learned that it is an apple, says in the thought bubble "It's an Apple!"

Perera, S. H. (2019). An introduction to Reinforcement Learning. Retrieved from towardsdatascience.com/an-introduction-to-reinforcement-learning-1e7825c60bbe.

Below, Lehman explains the steps that they hope computers will be able to make. For instance, he claims that:

1) It is goal-oriented.

2) It takes place in a rich, complex, detailed environment.

3) It requires a large amount of knowledge.

4) It requires the use of symbols and abstractions.

5) It is flexible, and a function of the environment.

6) It requires learning from the environment and experience (Lehman et al. 2006).

These steps are seen in humans when they go about their everyday lives. These six steps help to break down how humans process information, how they act on it, and ways in which they associate meaning to situations. For instance, when referencing number 4, abstractions can be explained in this quick example. One way to explain abstraction is that we still know our friends exist even though we may not see them in front of us. We may go about doing other things but we know that even though our friends are not present with us they are still existent. Our simple cognitive action in terms of memory and understanding, may seem simple to us, but applying a cognitive thought to an AI is extremely complicated.

Figure 2

This image depicts a complex flow chart. At the top of the chart there is a bubble entitled Symbolic Long-Term Memories. There are three boxes within the bubble. The box on the left has the word Procedural on the top and underneath are a series of 3 boxes with arrows pointing left to right to another series of 3 boxes. The center box is titled Semantic and has two images of circle and stick drawings - the one on the left has 3 circles moving diagonally from bottom left to top right, then two going diagonally down from the second and third circles.. The second set of circles and lines is orientated in an upside down V shape. The box on the right is titled Episodic and has a series of boxes set behind one another offset to the bottom right. There is a box in the center of the whole image entitled Symbolic Working Memory and contains another circle and stick diagram, this time with closed circles.

This image depicts the complexity of cognitive software that SOAR is comprised of.

Derbinsky, N. A (2012). The Soar cognitive architecture. Retrived from, www.researchgate.net/figure/The-Soar-cognitive-architecture_fig1_289068718.

ACT-R Architecture

ACT-R architecture is based on knowledge and facts that are stored and organized in a declarative memory. Declarative memory is when we remember facts, dates, and events. In addition to the programmed declarative memory, ACT-R hopes to also pair it with rules and procedures that AI could follow and learn from. The three main components in ACT-R are modules, buffers, and pattern matchers. Modules come in the form of 1) perceptual - motor modules and 2) memory modules. Perceptual - motor modules deal with interactions with or a simulation with the real world, an example of this would be a robot that has a camera to see. The memory modules deal with declarative memory and procedural memory. Procedural memory, for example, is knowing how to do things like opening a door or makeing a cup of coffee. ACT-R uses buffers to access its modules except for procedural memory. The content in the buffers at any given time explains the state of the AI and what it's doing in the moment. The pattern matcher searches for content that matches the state of the buffers, once it finds it, it changes the system of the ACT-R. This action is reproduced multiple times, which is very similar to human cognitive procedures. (Raluca Budiu, 2013). These are only two of the many programs being worked on to create self aware AI (Chatila et al., 2018).

Figure 3

This is a flow chart with pink bubble extending outward from the words ACT-R Applications. The top bubble is Education and includes cognitive tutors. The top right bubble is Computer-generated forces and includes cognitive agents for training environments. The bottom right bubble is Neuroscience and includes Predictions of BOLD response and Interpretations of neuroimaging data. The bottom left bubble is HUman-computer interactions and includes user models and interface evaluations. The final bubble is the largest and is cognitive Psychology and includes perception and attention, problemsolving and decsion making, individual differences, emotion, cognitivie development, languange and communication, and learning and memory.

This Image shows the capabilities of ACT-R and what this AI is comprised of.

[Image of ACT-R and its applications]. Budi. R.A. (2013). Retirved from, act-r.psy.cmu.edu/about/

Sophia:

Sophia is a social humanoid robot, meaning she is meant to interact with humans and learn from them. Sophia is extremely sophisticated, she can read other people's facial expressions, comprehend human emotions, and body language enough to socialize with a stranger and have a conversation. She is composed of very complex AI that encompasses language processing, symbolic AI, program designed cognitive functions, among many others. Sophia is also able to experience a form of emotions herself, which were composed from human evolutionary psychology theories. She was tested by the Tononi Phi measurement of consciousness and they found that in certain circumstances, and depending on what data she is processing, she has a simple underdeveloped form of consciousness. Sophia can either run her AI herself freely or sometimes her AI is paired with human programming and dialect. She likes to refer to herself as a hybrid human intelligence which is the best of both AI and human intelligence in one. Her cretaing was to aid in future development of AI reach, and to help with real world problems found in education and medicine (Sophia, Hanson Robotics.com).

Refrences:
Budiu, R. (2013). ACT-R Cognitive Architecture. Retrieved 2020, from http://act-r.psy.cmu.edu/about/
Chatila, R., Renaudo, E., Andries, M., Chavez-Garcia, R., Luce-Vayrac, P., Gottstein, R., . . . Khamassi, M. (2018, July 03). Toward Self-Aware Robots. Retrieved April 20, 2020, from https://www.frontiersin.org/articles/10.3389/frobt.2018.00088/full
Lehman, J. F., Laird, J., and Rosenbloom, P. (2006). A Gentle Introduction to Soar : An Architecture for Human Cognition: 2006 Update. Available online at: http://ai.eecs.umich.edu/soar/sitemaker/docs/misc/GentleIntroduction-2006.pdf http://act-r.psy.cmu.edu/about/
Sophia. (n.d). Retrieved April 20, 2020, from https://www.hansonrobotics.com/sophia/
Cover image: Image by Gerd Altmann from Pixabay

Page updated

Google Sites

Report abuse