Exploring the Relationship Between Human Cognition and Artificial Intelligence
“Mathematics is about abstracting away from reality, not about replicating it. And it offers real value in the process. By allowing yourself to view the world from an abstract perspective, you create a language that is uniquely able to capture and describe the patterns and mechanisms that would otherwise remain hidden. And, as any scientist or engineer of the past 200 years will tell you, understanding these patterns is the first step toward being able to exploit them.” (Hannah Fry, The Mathematics of Love)
If we were going to examine the fundamentals of Artificial Intelligence (AI), we might start by examining human cognition. This was the challenge posed in Alan Turing’s famous 1951 paper, “Intelligent Machinery, A Heretical Theory”, in which he stated:
“‘You cannot make a machine to think for you. This is a commonplace that is usually accepted without question. It will be the purpose of this paper to question it.”
Five years later in 1956, in a grant submission to the Rockefeller Foundation, an esteemed group of scholars, including John McCarthy, Marvin Minsky, Claude Shannon, John Nash, John Holland and Herbert Simon, proposed:
“..that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.”[1]
With what became known as the Dartmouth Conference, AI was launched. From its earliest days, one objective for AI was to “form abstractions and concepts”. In order to pursue this goal, from the beginning AI has been modeled after human cognition. Santa Fe Institute scholar Melanie Mitchell and her thesis advisor Richard Hofstadter make the point well:
“Understanding the cognitive process of analogy — how human beings make abstract connections between similar ideas, perceptions and experiences — would be crucial to unlocking humanlike artificial intelligence.”[2]
I sometimes feel that the original premise for AI development has been lost in the recent hype and excess surrounding ChatGPT. The purpose of this article is to show the relationship between human cognition, AI and how recent developments in AI outside Chat are bringing us closer to the original plan.
Cognition
“Information is the structure of reality — the fabric of the universe/nature for a cognitive agent….In other words, information is observer-relative. It does not mean it is subjective. It is the same kind of observer-relativity as in the theory of relativity in physics. Information and reality are seen as one by living organisms as cognizing agents.”[3] — Gordana Dodig-Crnkovic and Marcin Miłkowski
Cognition can be understood as:
“…concept formation, abstraction, and adaptation …undergirded by processes of analogy — the act of perceiving essential similarities between entities or situations.”[4]
In simpler terms, cognition is a two-part process of pattern recognition (concept formation, abstraction) and probability (adaption)[5]. Humans are constantly acquiring sensory data and deciding what “information” to retain. Such an approach reduces the demand for energy,[6] which is particularly relevant given that the brain is the largest user of energy in the human body. When the decision is made to retain, explicitly the brain assigns the information to one or more models which reflect existing knowledge or patterns. Explicit in this assignment is a risk assessment using probability in order for the human to reduce uncertainty and “adapt”, if necessary. “It appears that such intrinsic random and probabilistic elements are needed in order for a comparatively small population of simple components (ants, cells, molecules) to explore an enormously larger space of possibilities, particularly when the information to be gained is statistical in nature and there is little a priori knowledge about what will be encountered.”[7] Over time the models or concepts become less and less specific and more and more precise, which increases their usefulness. Over time, adaptive behavior becomes more about risk management [probability] and less about the information.[8]
M. Mitchell Waldrop makes the point well about the constant process of model development and improved risk assessment. His point expands on John Holland’s “internal model”.
“Namely, that knowledge can be expressed in terms of mental structures that behave very much like rules; that these rules are in competition, so that experience causes useful rules to grow stronger and unhelpful rules to grow weaker; and that plausible new rules [insights] are generated from combinations of old rules.”[9]
Daphne Bavelier at the University of Geneva adds an important point to understanding this process.
“…that we do not actually “see” with our eyes but with our brain. And we have learned that in turn by becoming able to see how the brain operates. What we see with the eyes, it turns out, is less like a photograph than it is like a rapidly drawn sketch. Seeing the world is not about how we see but about what we make of what we see. We put together an understanding of the world that makes sense from what we already know or think we know.”[10]
The view described by Bavelier clearly takes us away from Plato’s epistemology of an objective reality and Descartes does not do much better. Descartes advocated for a mind-matter duality wherein reality is what the observer sees, which dramatically increased the use of empirical methods in science.
Today, the emerging view is that the mind is constantly determining the stochastic likelihood of your “reality”.[11][12] The neuroscientists would say that there is a reality. The quantum physicists would probably say there may be a reality, but what we know are probability statements about sub-atomic components (waves or particles), the features or variables of the components (mass, charge, and spin) and how they combine to build from the microscopic to the macroscopic level. There may be a reality, but we are limited to what we observe, when we observe, and what we know about any system is explained by probability statements.
Consider this edited quote:
“This family of research approaches is deeply rooted in various forms of predictive coding in which prediction signals, as representations constructed from past experiences, are compared with incoming information to form prediction errors; prediction errors can be encoded and learned to update stored experience, which is then available for use in future predictions.”
To me this quote looks to be a description of basic algorithm training in machine learning (ML). In fact, the quote is from an excellent recent article, “The power of predictions: An emerging paradigm for psychological research”.[13] My point is that we are now at a point in intellectual history where neuroscience, quantum physics and AI are all using stochastic determinations to derive knowledge. The founders of Santa Fe Institute make the point well.
“But the way in which complex phenomena are hidden, beyond masking by space and time, is through nonlinearity, randomness, collective dynamics, hierarchy, and emergence — a deck of attributes that have proved ill-suited to our intuitive and augmented abilities to grasp and to comprehend.”[14]
What this stochastic method does is it frees us from the limitations of Descartes’ perceived reality and allows us to pursue knowledge at a more foundational level. For example, we might no longer need to treat symptoms in medicine and could rely instead on diagnostics through precision medicine based on the patient’s genetics. Such approaches are increasingly made possible by AI.
Artificial Intelligence
“To beget complex forms of information, such as those that populate our modern society, we need to evolve complex forms of computation that involve networks of humans. Our society and economy, therefore, act as a distributed computer that accumulates the knowledge and knowhow needed to produce the info.”[15] — Cesar Hidalgo
This graphic from the Economist does a nice job of giving an overview of AI.
The excellent ivp “AI Primer”[16] explains AI as:
“Artificial Intelligence (AI) is a scientific technique that enables machines to mimic intelligent human behavior. Artificial General Intelligence (AGI) refers to the ability of a machine to perform any task a human being is capable of.”
The researchers at Caltech explain machine learning simply as:
“At its core, machine learning is the process of computers using algorithms to make intelligent decision based on data.”[17]
Machine learning and all AI is currently built with three components: 1) datasets, 2) algorithms, 3) computing[18] in order to identify patterns that form the basis for prediction.
In the early days of AI, the thinking was that AI needed better computers to reach its lofty goals. Machine Learning (ML), “a way to teach computers how to learn from data and make decisions and predictions without being explicitly programmed”,[19] was the common form of AI. During this period AI might be described as automated statistics; for example, using linear regression and logistical regression algorithms. From its earliest days, ML (and AI) has had the same objective as the cognitive neuroscientists — to determine the variables (pattern recognition) and build the statistical model for prediction. In statistics we use the term coefficient and in ML we use the term weighting to understand the importance of the variables. As we review the history of ML below, we should note that the computation is adding more and more variables to increase the certainty of prediction for a wider range of problems. This is very similar to Piaget’s[20] explanation of the behavior of a child exploring the world and building a robust framework for risk assessment.
In about 2010 neural networks and deep learning algorithms emerged, supported by much improved computers and massive datasets, frequently housed in the cloud infrastructure of Google, Microsoft and AWS. What these new algorithms helped to show was that the limitation in AI was not the computing power but rather the available data. In about 2017 new algorithms for natural language processing appeared, which eventually led to the frenzy around generative AI and ChatGPT. ChatGPT-4 is the latest version (2023) of the commercially available “text editor” that is automating student and office tasks in a way that alarms some people. I see ChatGPT more like a calculator or Excel, a tool to be learned and no real cause for alarm. In fact, as one observes the early success of ChatGPT-4, one could be more encouraged about the future contribution of AI. For those with ethical concerns, I would point out that every technology paradigm from mechanization to electricity and computing has spawned social and ethical concerns in its early days. These issues may be more concerning today because it is so much harder for policymakers to understand the issues (even with help from the Twitter pundits), but I believe we will work it out satisfactorily and hopefully without destroying democracy or capitalism.
I believe that the risks of further advances in AI are worth considering. Capra and Luisi frame the future[21] of AI well:
“Phenomena is reached only when we approach it through the interplay of three different levels of description — the biology of the observed phenomena, the laws of physics and biochemistry, and the nonlinear dynamics of complex systems.”
What these physicists are saying is that Descartes’ reductionist, mechanistic science is gone, replaced by a modern science based in quantum physics and the non-linear dynamics of adaptive systems (complexity theory). The tools required to pursue this new direction (the problem) are here, working and advancing to paradigm status, as W. Bryan Arthur predicted in his 2009 book, The Nature of Technology: What it is and How it Evolves. That tool of course is AI and ML, and the tools have consistently appeared throughout history to solve the then current problems. Today the AI tools provide more advanced computation, models and understanding.
As I look at the future of AI, what encourages me is the evolution of data to synthetic data. Synthetic data[22], by its nature original and “manmade”, is produced by generative algorithms and provides the opportunity to find new patterns that might lead to new science. Research should happen at the frontier of a discipline where new theory is discovered. Synthetic data might help us to redefine the frontiers in biology and medicine by showing us previously unseen components that become patterns. The success of transformer algorithms, which offer better understanding of time series data, also provide the opportunity for earlier determinations and findings in multiple disciplines.
Multimodal AI permits a whole new type of data analysis. “Multimodal AI is a new AI paradigm, in which various data types (image, text, speech, numerical data) are combined with multiple intelligence processing algorithms to achieve higher performances.[23] I think this ML might finally help us to tackle the multivariable, “wicked problems” that have challenged social science for so long. Another application of this multimodal logic is multimodal networks (MMN), “a novel graph-theoretic formalism designed to capture the structure of biological networks and to represent relationships derived from multiple biological databases.”[24] Capturing complex network structure, relying on multiple databases, is a promising area in AI with applications in biology, medicine, social science, information theory and many other areas. (A CMU tutorial on Multimodal Machine Learning is here.)
The increasing use of graph theory in AI will also help researchers to better realize the benefits of increasingly larger datasets and networks. As described in a recent Quanta Magazine article, graph theory may provide an entirely new future direction in computation.
“A hyperdimensional vector, or hypervector, could be an array of 10,000 numbers, say, representing a point in 10,000-dimensional space. These mathematical objects and the algebra to manipulate them are flexible and powerful enough to take modern computing beyond some of its current limitations and foster a new approach to artificial intelligence.
The SuperMind project at MIT might explain the point as follows:
“…the current discussion about AI and other advanced technologies largely misses an important point: that the locus of intelligence isn’t in the individual, but rather in the emerging collective intelligence that stems from the interoperation of large networks composed of both human and machine nodes.” [25]
This SuperMind graphic[26] illustrates the point and the evolution of our use of networks.
Conclusion
“The view of living systems as networks provides a novel perspective on the so-called “hierarchies” of nature. Since living systems at all levels are networks, we must visualize the web of life as living systems (networks) interacting in network fashion with other systems (networks).”[27] — Fritjof Capra, Pier Luigi Luisi
The SuperMind graphic shows me that the future of the organization (public, private or non-profit) will be shaped by large networks, the development of large network-based datasets and the AI algorithms that will be required for complex computation. Traditional ways of working and the private sector “business model” will need to be dramatically updated. Such networks foster self-organizing, bottom-up hierarchies and may lead to a redefinition of national and state government’s role and the continued emergence of cities as the principal governmental players.
Microsoft describes the current state of AI as a “phase change”[28], a transition from one state of matter to another. Perhaps the phase change is the 10,000 dimension data space, perhaps it is the role of the network or perhaps it is just “synthetic” data. However, you wish to identify the phase change, the changes coming from AI are under appreciated. Buckle Up!!
“Mathematics is about abstracting away from reality, not about replicating it. And it offers real value in the process. By allowing yourself to view the world from an abstract perspective, you create a language that is uniquely able to capture and describe the patterns and mechanisms that would otherwise remain hidden. And, as any scientist or engineer of the past 200 years will tell you, understanding these patterns is the first step toward being able to exploit them.” — Hannah Fry, The Mathematics of Love
FOOTNOTES:
[1] https://www.cantorsparadise.com/the-birthplace-of-ai-9ab7d4e5fb00
[2] https://www.quantamagazine.org/melanie-mitchell-trains-ai-to-think-with-analogies-20210714/
[3] https://pdfs.semanticscholar.org/cf07/2b24cece6b7fa1e16993a4ab09b92b7adfb9.pdf?_gl=1*14kdo4r*_ga*MTY4MzMxMTcyMy4xNjgwMzYzODE1*_ga_H7P4ZT52H5*MTY4MDM2MzgxNC4xLjEuMTY4MDM2MzgzMS4wLjAuMA
[4] https://arxiv.org/pdf/2102.10717.pdf
[5] https://nautil.us/that-is-not-how-your-brain-works-238138/?utm_source=nautilus-newsletter&utm_medium=email&he=a72a66be0e8938958a83906d0e7aaee1
[6] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6867616/
[7] https://sfi-edu.s3.amazonaws.com/sfi-edu/production/uploads/sfi-com/dev/uploads/filer/67/c4/67c4bcce-7e58-4819-8aed-9a67163dd441/06-10-036.pdf
[8] This logic underlies the fundamental behavior provided through evolution called “explore — exploit”. Any living thing is programmed to constantly explore until they have a sufficient certainty (probability) and then they make the decision to exploit (act). For example, the ant explores every morning looking for a food source sufficient to feed the colony. When it finds such a source, the ant leaves a chemical trail for the other ants to follow which reduces the energy consumption and risk to the colony.
[9] https://www.google.com/books/edition/Complexity/m0yqDwAAQBAJ?hl=en&gbpv=1&dq=%22namely,+that+knowledge+can+be+expressed+in+terms+of+mental+structures+that+behave+very+much+like+rules%3B%22&pg=PT238&printsec=frontcover
[10] Nicholas Mirzoeff, How to See the World
[11] https://nautil.us/that-is-not-how-your-brain-works-238138/?utm_source=nautilus-newsletter&utm_medium=email&he=a72a66be0e8938958a83906d0e7aaee1
[12] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6867616/
[13] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6867616/
[14] David Krakauer, Murray Gell-Mann et al. Worlds Hidden in Plain Sight: The Evolving Idea of Complexity at the Santa Fe Institute 1984–2019
[15] Cesar Hidalgo. Why Information Grows: The Evolution of Order, from Atoms to Economies
[16] https://assets.ivp.com/system/uploads/fae/file/asset/232/AI_Primer_vF2.pdf
[17] https://www.caltech.edu/about/news/improving-aircraft-design-with-machine-learning-and-a-more-efficient-model-of-turbulent-airflows
[18] https://assets.ivp.com/system/uploads/fae/file/asset/232/AI_Primer_vF2.pdf
[19] Ibid.
[20] https://link.springer.com/article/10.1007/BF02686930
[21] Fritjof Capra, Pier Luigi Luisi, The Systems View of Life
[22] “Synthetic data supplements real-world observations with computer-generated outputs programmed to normalize the distribution of datasets. Methods to generate synthetic data include statistical sampling, simulation scenarios, or generative adversarial networks, and can retain multivariate characteristics similar to the source dataset while offering richer metadata and a larger sample size.” — PitchBook
[23] https://www.aimesoft.com/multimodalai.html
[24] https://pubmed.ncbi.nlm.nih.gov/19407355/#:~:text=A%20multimodal%20network%20(MMN)%20is,derived%20from%20multiple%20biological%20databases.
[25] https://www.supermind.design/
[26] https://static1.squarespace.com/static/5e95059565bf963c169f906a/t/6305df80cf3e402080d8d663/1661329299417/Augmented+Collective+Intelligence+-+PRACTITIONER.pdf
[27] Fritjof Capra, Pier Luigi Luisi, The Systems View of Life: A Unifying Vision
[28] https://www.microsoft.com/en-us/research/podcast/ai-frontiers-ai-for-health-and-the-future-of-research-with-peter-lee/?ocid=eml_pg397252_gdc_comm_mw&mkt_tok=MTU3LUdRRS0zODIAAAGLV8yQ89HTlRR3CUEeWhj0lzBGXL0lmyIOnOrrsXaJNN_QoTkMbgjehvZXGL2lEUoq2go0Acqnb5sbU1Lki0z7x_Fu_mWhCYOsmr762kLuSvgqTqEowfPQGNg
