Knowledge in humans is almost poetic. How we acquire knowledge to create our dreams, ideals, and ideas is a way for us to express ourselves completely. What is in our minds comes to life through expression on pages, music, dance, art, and much more. It is a way to say, "this is what is in my mind, and I have wielded it out." Learning how to express this knowledge is the action, and we love observing this action.
“Learning is the mental activity by means of which skills, habits, ideas, attitudes, and ideals are acquired, retained, and utilized, resulting in the progression of adaptation and modification of behavior” (William A. Kely, Educational Psychology, 4th ed., Milwaukee Publishing, Bruce Publishing Co., 1956, p. 238). Learning in humans can be independent and lead to action or not. It is self-driven (DePhillips), and overall, can occur in four ways: visual, auditory, reading/writing, and kinesthetic.
In contrast, artificial intelligence (AI) occurs in a world where it is being trained. Because the goal of AI is to objectively perform tasks that humans perform, it makes sense to start by making a machine run on the neurological principles that apply to humans. Humans learn neurologically by highly parallel networks of simple non-linear neurons that adjust the strengths of their connections (Yoshua Bengio). Deep learning addresses human learning computationally by using many layers of activity vectors as representations and learning the connection strengths that give rise to these vectors by following the stochastic gradient of an objective function that measures network performance (Yoshua Bengio). Understanding human learning processes can offer insights into developing more efficient and potentially sentient AI.
Learning in Machines: A Data-Driven Approach
Machine learning has seen enormous success with applications like ChatGPT and DALL-E 2, which have revolutionized the world with generative AI. In machine learning, algorithms are trained to find patterns and relationships in data. It looks into these large datasets and finds patterns in the historical data as input to make predictions, classify information, cluster data points, and even generate new information. Finding these relationships is key; contrastive learning, an approach in machine learning, locates similar and dissimilar data points.
In computing, the term “training” is often used rather than “learning.” Training happens via highly parallel networks, non-linear neurons, large datasets, and deep networks. Contrastive learning provides a way to represent semantic information and contextual relationships. Generative Adversarial Networks (GANs) train a generative neural network to produce contrastive samples by applying a neural network to latent samples from a known distribution. (Yoshua Bengio, Yann LeCun, and Geoffrey Hinton, recipients of the 2018 ACM A.M. Turing Award). For machine learning systems, humans provide the training data. By running a human-generated algorithm on the training dataset, the system generates rules such that it can receive input x and provide correct output y. This is why data is so vital in the context of AI. Data is the main raw material out of which high-performing machine learning AI systems are built (Greg Allen, Understanding AI Technology, Chief of Strategy and Communications, Joint Artificial Intelligence Center (JAIC), Department of Defense).
The Challenge of Data-Driven Learning
The challenge with relying on data for computation is that it can be polluted by humans. As Jaron Lanier states in "Who Owns the Future," data manipulation rises and becomes suspect, creating illusions even in large datasets. Model building is the hallmark of human-level learning, explaining observed data through the construction of causal models of the world. Compositionality and learning-to-learn are ingredients that make this type of rapid model learning possible (Brenden M. Lake et al., "Building Machines That Learn and Think Like People," Behavioral and Brain Sciences).
The Human Learning Experience: Organic and Multifaceted
Learning in humans is dynamic, comprising a multitude of activities, and this happens most effectively in a natural way. Learning naturally means one is not subjected to multiple trials of the same thing in an organized setting. In the purest form of learning, we see how children learn. A young child new to the world, seeking to understand their environment quickly for survival, may see a picture of a monkey and immediately understand all the important things that make a monkey a monkey. They may see the picture again, reinforcing that a monkey is still a monkey despite slight changes. They do not need to see a thousand pictures of monkeys to discern this.
Similarly, a child will enjoy playing with water and their brain understands how water feels on the skin and how it looks instantly. After this initial cataloging, the child viewing a still picture of a monkey can now distinguish a monkey on-screen moving in an environment with trees, water, and other animals and people. Children playing and enjoying water can now indicate to a parent that on their summer holiday, a trip to the beach is what they are seeing as water, vast amounts of it.
Learning as described in these cases is not the learning of a new skill; it is observing. In the popular conscious competency model, the focus is on learning a new skill. Martin Broadwell’s model moves through four stages of learning: unconscious incompetence, conscious incompetence, conscious competence, and unconscious competence (Management of Training Programs by DePhillips, Frank A., Berliner, William M., and Cribbin, James J.). This model shows how a learner moves from being unaware of their ignorance to being independent and applying their knowledge.
This observation allows one to think about learning in the capacity of being "organic" as much more efficient than running a training module for 50,000 cycles or more, or the famous experiment of cataloging a cat as a cat through thousands of images by Google. The notion of learning this way, similar to humans through repetition, can produce favorable results, though not efficiently. Learning by the person receiving the information is one that chooses the method, understands the importance of acquiring this knowledge, and is not a set of organized rules adjusted to a set of differentiated learning strategies. Each person possesses natural strengths and preferences, which affect their attitudes and commitments towards learning, as well as their abilities in developing competence in different disciplines (https://www.businessballs.com/self-awareness/conscious-competence-learning-model/).
Bridging the Gap: Lessons from Human Learning for AI Development
There seems to be a gap between learning for AI and humans. Students notoriously complained about being victims of a learning environment during the COVID-19 pandemic, where learning happened online. The correlation that being online all day for entertainment purposes to learning in a structured way was frustrating. The notion that they were no longer free on their devices was met with hindered learning capabilities.
The ability to keep learning environments organic has several benefits, such as time, where the ability to grasp the nature of things almost instantly occurs, critical thinking evolves, and empathy grows—all things teachers cannot fully achieve in a classroom. Regulated learning reduces understanding and enjoyment, influencing how much a human learner wishes to undertake the task and dedicate their full attention to it.
Machine learning advances the concept of learning through agility. Computers can perform the same computations as humans, simply faster, and digest more information from subject matter experts than a person can in their lifetime, to find relationships and patterns in vast amounts of information through statistics and probability. Yet agility, though fascinating and successful, does not and will not lead to true intelligence like humans.
In contrast, human learners, take for example the KIPP program (The Knowledge is Power Program), spend more time grasping a topic. As Malcolm Gladwell writes, KIPP students spend twenty-five minutes collectively solving a math problem, resulting in better retention and understanding of the process. Currently, our AI systems are training individually and not as a collective. Skills such as motivation and grit are more difficult to program than endurance, reward, and self-control.
AI Benefiting from Human-Like Learning
In human learning, the parent or caregiver acts as a facilitator, providing a rich and stimulating environment, offering guidance when needed, and encouraging independent exploration. This process allows children to develop a sense of agency and self-efficacy, which are crucial for lifelong learning and critical thinking. A child can be self-driven and give themselves their own directives to carry out play. This is driven by some other inner desire as they seek to understand their world around them.
In contrast, machine learning is often viewed as a purely technical endeavor, requiring specialized knowledge and expertise from computer scientists and engineers. While this is true to some extent, the current paradigm overlooks the potential benefits of incorporating more "human-like" learning strategies.
The limitations of this approach are becoming increasingly evident. AI systems trained in highly structured environments often lack the ability to think outside the box, generalize to new situations, or exhibit the kind of creative problem-solving that humans excel at.
By focusing solely on technical aspects and neglecting the broader principles of learning and cognition, we may be inadvertently hindering the development of truly intelligent machines.
To bridge this gap, we need to consider how to empower machines to learn more autonomously and self-directedly. This could involve:
Open-ended exploration: Allowing AI systems to explore vast amounts of data without specific goals or constraints, much like a child playing with blocks or exploring a new environment.
Intrinsic motivation: Designing AI systems with an innate drive to learn and discover, rather than relying solely on external rewards or punishments.
Curiosity-driven learning: Encouraging AI systems to actively seek out new information and experiences that challenge their current understanding of the world. Providing it with a sense of self.
Learning from mistakes: Incorporating mechanisms within training models for AI systems to learn from errors and failures, rather than simply discarding them as incorrect outputs.
Emotional intelligence: Allowing AI systems to include accounts of emotions in their memory and associate them with current events. This builds on experiences, which humans excel at for their decision-making.
By incorporating these principles, we could potentially create AI systems that are not only more intelligent but also more adaptable, resilient, and capable of independent thought. This shift in approach could open up new avenues for innovation and lead to AI that truly complements and enhances human capabilities.
Additionally, democratizing the process of teaching machines to learn could have far-reaching benefits. By making AI development more accessible to a wider range of people, we could tap into a broader pool of creativity and diverse perspectives, potentially leading to breakthroughs that wouldn't be possible in a purely technical, expert-driven environment.
Ultimately, the goal is not to replicate human learning perfectly but to draw inspiration from its most effective aspects to create AI systems that are capable of learning, growing, and evolving in ways that were previously unimaginable.
Decision-Making Through Emotional Intelligence
Humans learn about their world through the experience of emotions. The ability of humans to think in a more advanced way occurred only between seven thousand to thirty thousand years ago, when humans began to think more intelligently, creatively, and sensitively. This is what scientists call the “Cognitive Revolution,” possibly due to some genetic mutation, though confirmation is still needed. This major inflection point for humans is referred to as the “Tree of Knowledge,” where humans through language found success in migrating, building tools, producing art, and understanding vast amounts of information and simplifying it within a threshold number of people to be 150, as Yuval Noah Harari writes in his book "Sapiens: A Brief History of Humankind."
The notion to communicate and develop linguistics arises from a need for survival, which current AI has yet to develop. Humans' need for survival needs the support of others in their community. Longevity can arise from supportive communities, achieving positive health. Emotional well-being is key in developmental intelligence in a positive way to produce survival. Interpersonal flourishing is a core feature of quality living across cultures and time (Carol D. Ryff). Skills such as empathy allow humans to understand and connect with others, fostering collaboration, cooperation, and social cohesion. This builds on their personal needs of survival for one through survival for all. The need for this survival mechanism may differ for machines, as humans have physical constraints (the body) while machines' physical constraints lie in their circuitry and stability of their electronic devices.
Emotional Intelligence in AI
Developing AI systems with emotional intelligence presents challenges and potential benefits. Affective computing aims to create machines that can recognize, interpret, and respond to human emotions. Emotionally intelligent AI agents can read, understand, express, and regulate emotions, developing behavioral expressions that seem emotional and using them effectively (Karimi Aliabadi et al., 2021).
Applications for emotionally intelligent AI include healthcare, education, and customer service. In healthcare, AI could provide empathetic support to patients, improving mental health outcomes. In education, emotionally aware AI could tailor learning experiences to students' emotional states, enhancing engagement and retention. In customer service, AI with emotional intelligence could offer more nuanced and effective support, improving customer satisfaction.
The Quest for Sentience
As AI develops, it is crucial to keep in mind that humans are self-driven in their learning, a mechanism caused by an underlying principle for the survival of self. This arises because one is conscious they are alive and self-aware. Achieving this in AI systems would mean the world of AI has reached sentience. The ethical implications of this are concerning, with geopolitical considerations, technical rigor, and economic, financial, and social changes to consider. Sentient AI would be the apex of achieving human-like intelligence. However, AI is defined by its own intelligence, and that may suffice as another endpoint.
Currently, computer scientists focus heavily on coding, but few focus on device development. No matter how complex, eloquent, and good the code is, it is still subject to the device's physical constraints and the laws of physics, which are not often discussed in AI education. Machines achieving this level of sentience will be limited by the amount of stress and action a device can withstand to store and work with such coding.
Humans learn not just by observing data and ingesting information but by interacting with the world through touch, smell, taste, sight, feel, and emotions. What can be learned from interaction with data is far beyond what AI today can comprehend because it is limited in how it interacts with data. The neurological complexity of human thinking and wiring goes far beyond advanced parallel neural networks. AI systems need to learn in environments that allow physical interactions with the world and combine that with affective computing. This could involve developing AI with intrinsic motivations, values, and a sense of purpose to be truly autonomous or living.
Conclusion
The evolution of AI holds tremendous potential, but it must be steered by the nuanced understanding of human learning principles. Incorporating elements of human-like learning, emotional intelligence, and intrinsic motivation can lead to more robust, adaptable, and empathetic AI systems. This journey towards sentient AI demands a balance of technical innovation and ethical consideration, ensuring that as machines grow more intelligent, they also become more aligned with the human experience. The ultimate goal is not to create AI that mimics humans but to forge AI that enhances human capabilities, fostering a future where technology and humanity progress together.
Take a look at some insights found in the development of A.I.
Key Trends to pay attention to for A.I. development
Commentaires