I ask silly questions and my silly questions have lead me to asking even sillier questions. I've asked these questions as part of my work in artificial intelligence (AI). Fortunately, other people have asked similar questions and they've also provided compelling answers. My silly questions have lead me to fields both inside and outside of artificial intelligence and to fields that I didn't even know existed.
Question 1: Why can't AI systems keep learning?
There were many answers. We don't have powerful enough computers. We don't know enough about artificial and natural intelligence. AI systems have common sense problems, complexity problems that hinder extension, frame problems, symbol-grounding problems, context problems, and relevance problems.
As part of exploring this question I was asked to define what I was after but I had a lot of trouble trying to do this. Applications that seemed achievable could be done with shortcuts. Applications that could not be done with shortcuts seemed unachievable. The area that best represented what I was after was natural language processing (NLP).
I thought that a system that could use language could then use language to learn. I thought that knowledge bases could help NLP systems with the common sense problem and cognitive architectures could help with learning. However, determining relevance, context, and frames was still difficult. Getting the system to learn the way I wanted was hard. On the other hand, getting the system to do achievable things resulted in a system that could not keep developing.
The root problem was context, which in turn was due to agents not being situated and not having needs. In other words, problems were due to trying to formalize knowledge rather than getting agents to cope skilfully with a body. I decided to focus on getting AI systems to benefit from having bodies, needs, and from being situated.
I looked at embodiment artificial intelligence (EAI) and artificial life (AL). I came across the idea of scaffolding in EAI, which is the idea of agents modifying the environment to help them perform their tasks. Modifications could be changes to the physical environment or even the development of language. However, I saw that the majority EAI researchers were focusing mainly on low-level intelligence such as walking. At this point, I was still thinking about language.
In a discussion about changing from NLP systems to EAI systems, I was asked about the difference between my agents and animals. What I had didn't seem to be enough. My creatures had bodies, were engaged in form of life, and were situated in an environment. So are animals. Animals face similar challenges including having to find food and evade predators. They have similar physical needs and capabilities including needing food and being able to move. However, animals have not developed the type of intelligence exhibited by humans. Why then, would my system evolve any further?
I was still thinking about language. I looked at language evolution and emergence. I found that researchers had gotten agents to develop basic languages but I wasn't satisfied. I came across the idea of cultural evolution (Plotkin, 1997). This idea holds that life and evolution is a knowledge acquiring process that occurs through adaptations at the genetic, personal, and cultural levels.
When cultural evolution is present, there is a new environment where the ability to acquire, use and share concepts which underpins intelligence, is favored. Animals have basic culture in the form of their communications but only human beings have cumulative culture evolutions. This lead to my second question.
Question 2: Why are humans more intelligent than animals?
Even with the idea of cultural evolution, I was left with the question of why are humans more intelligent than animals. I realize that there are arguments about whether humans are more intelligent than animals. The following are some reasons for thinking that humans are more intelligent than animals. Humans can communicate, build, cooperate, learn, and acquire concepts by themselves and from others.
The following are some ideas that first come to mind to answer the question of why humans are more intelligent than animals. Humans had to acquire abilities described above because humans find it harder to get food, and harder to avoid danger than animals. Humans could acquire abilities described above because of they had bigger brains, better vocal control and had toes and thumbs so there was more we could learn.
Up to now, I've haven't described what I'm after. I'm after a system that can acquire and use concepts through personal learning and social learning. As part of that, I'm after agents that can learn to communicate.
We could make changes to the environment to increase pressure. If the pressure is great enough then agents would be forced to adapt anyway they could to survive. If the environment was challenging enough then genetic and personal adaptation would not be enough and so cultural evolution must occur.
There are many choices. However, the choices are unclear. Do I make it harder to get food? Do I make it harder to avoid danger? Do I make agents have bigger brains? Do I make agents have better vocal control? Do I make agents have toes and thumbs?
Increasing the challenge from the environment will select better accumulators and select better personal learners. In regards to social learning and communication, there is the idea that that pressure drives adaptation at the genetic level, then the personal level and then the cultural level.
We know that when cultural adaptation is present there is a pressure to select better accumulators, select better learners, and select better communicators. This is related to the ratchet effect where having a little language creates an environment that encourages further language development and consequently language development ratchets up.
Remember that I am asking these questions from the context of deciding what changes to make in my simulation. Making it harder to get food, harder to avoid danger, having larger brains, and having simple vocal control is relatively easy. Making agents have greater vocal control and making agents have toes and thumbs is relatively harder. In addition, I can't directly program behavior because I'm using neural networks. I also can't specify selection criteria because I'm not using a fitness function.
I came across Dawkins (1978) and Humphrey (1984). Dawkins talked about how adaptations occur to preserve genetic lines. Humphrey talked about how social groups are what created the pressure for the development of intelligence rather than challenges from the physical environment. That got me thinking about intrinsic fitness functions, social groups, gender, and maturation.
I could introduce gender and maturation to make things harder and to encourage the formation of social groups. There were already plenty of projects on physical environmental changes. I wasn't satisfied with language evolution research because they used predefined communication protocols and explicit fitness functions.
First, what if there were no changes? We can let pressure come from population growth. Consider a situation where the amount of resources such as food and dangers such predators in the environment are relatively stable. In addition, factors such as the speed, stamina, and strength of agent embodiment are also relatively stable. However, in this situation neural networks are allowed to evolve at genetic and personal levels.
Theoretically, intelligence will keep increasing. Agents that are more intelligent will grow in number. Eventually, agents will compete among themselves. Further, if communication, cooperation, and social learning skills are beneficial then agents will develop them.
Therefore, I don't have to change anything. However, I don't know for sure whether intelligence will evolve. It could take a long time. Depending on a large number of simulation parameters, agents could find relatively stable strategies such as reproducing less that temporarily reduce pressures for intelligence to evolve.
My next option was to let pressure come from changes in the physical world that is external to agents. Intelligence arises in response to pressures. I'm not talking about for example, the type of intelligence scientists use to notice natural phenomena. The motivation for that is subtle and complex. I'm talking about a minimum threshold intelligence that is foundational for further development of intelligence.
We could change the environment within a certain range to increase pressures. Increasing challenge from the environment means that agents have more to learn. Changing the challenge from the environment means that agents have to become better learners. However, going outside of the range that we know human intelligence developed in, we'll be entering the unknown. We don't know what will happen.
There is value in exploring ranges outside of which we know human intelligence developed. However, I am currently after human like rather than human level intelligence. Human like intelligence is about being aware of human contexts. For example, when humans point at their mouths and stomachs, they are probably referring to food and hunger. Human level intelligence is more about human competence. For example, noticing natural phenomena and expanding models of the world.
Let us return to the known range of changes in the environment. There is research that argues that pressures from the environment alone were not sufficient to drive the development of intelligence, cultural adaptation, social learning, or communication.
Please have a look in the notes at Humphrey, 1976, paragraph 10 and 12. Humphrey described an environment that was relatively plentiful but there was still a need to socialize to compete for mates, to learn about the environment and to raise children.
Cultural adaptations do occur in nature but cumulative cultural evolution is rare. Cumulative cultural evolution requires the ability to communicate but selection will only favor communication if the net benefit of social learning is significantly greater than the net benefit of personal learning (Boyd and Richerson). In addition, that only occurs when there is already cumulative cultural evolution in place.
Therefore, social learning is likely to have evolved because of another adaptation. The ability to understand that different agents have different beliefs and goals is critical for social learning. Without this ability, agents cannot connect their goals with the actions of other agents. It is likely that this ability evolved because agents needed to predict the actions of others in their social group (Boyd and Richerson, 1995).
Whether the environment needs to be challenging we do know that at the same time pressures came from the social environment played a big role. This lead me to my third question.
Question 3: Why are humans social?
The interplay between agent embodiment and the environment determines what agents can learn and what strategies are effective. If the environment is complex but agents lack the embodiment to exploit the environment then that limits what the agent can learn. Similarly, if agent embodiment is complex but the environment is simple, that also limits what agents can learn.
Like with intelligence, if we did nothing, social groups could theoretically form. However, we don't know for sure. It could take a long time and agents could find strategies that reduce the pressures to evolve.
There are a large number of factors that we can introduce to encourage to forming social groups.
We can develop an environment where cooperation brings greater gain than conflict. This is why abstract environments need to become more detailed. We need an environment that rewards intelligence and cooperation.
McNamara (2004) references Pfeiffer (1982) who found out that if humans ``were still hunter gatherers today, using the tools, knowledge, and social structures of that period, the entire land mass of the United States would only be able to support about 60,000 people instead of the 280 million that it actually does now.''
It is likely that a combination of scarcity, and an environment that rewarded cultivating, and cooperating helped social groups to form.
Humphrey (1976) described a different environment where scarcity was less of an issue but where social groups were important. In this environment, learning about the environment was difficult, there was a need to compete for partners, and cooperation benefited individuals.
Therefore, even if food was abundant and danger scarce, having to socialize still drove development. The following are factors that drove the formation of social groups.
Sharing the results of a hunt if prey are large enough (Campbell, 1999).
Having an embodiment that allows bringing results of a hunt back to a base share (Campbell, 1999).
Along the way, I came across the ideas that gender and maturation play a role in encouraging social groups to form but I struggled to find systematic explanations. That lead me to my fourth and fifth questions.
Question 4: Why does gender affect social groups?
Having to reproduce creates some pressure to socialize.
However, having gender creates additional pressure because females become increasingly scarce meaning that males have to compete and females could select males that are more capable. Gender evolved because when males have more eggs and when females have fewer eggs it is an optimum strategy for both. Loss of estrus, which is the female not showing signs when fertile, causes the male to stick around and to compete for her attention. The male then contributes to the chance of the female and her offspring surviving (Barash & Barash, 2001).
Instinctual incest avoidance helped stabilized inter-group relations (Chapais, 2008). Inter-group relations had to be stable in order for males or females to transition between groups to ensure genetic variation.
Introducing gender might sound weird. However, introducing food probably also once sounded weird.
Question 5: Why does maturation affect social groups?
Having children increases the cost of conflict. Conflicts become less favorable because the costs become too high. Raising children requires stability.
Raising children is difficult. It created pressure for cooperation and for role specialization such as provider, carer, and teacher (Baker, 2009).
Having children being less physically capable gives them more opportunities to learn because they can't be forced to work right away.
In addition, they are easier to train because good behavior can be rewarded because they are dependent and bad behavior can be punished because they can be overpowered.
Weaknesses of our approach include there are other ways that seem easier, it is hard to balance gender and there is a need for better kin recognition.
I want to stimulate my agents to evolve higher intelligence. Evolving higher intelligence requires agents to form social groups. Scarcity creates competition. Conflict and cooperation are strategies.
Effective strategies are dependent upon the interplay between the environment and embodiment. The environment must have options that reward agents for learning, sharing knowledge and cooperating.
Gender and maturation discourage conflict and encourage cooperation. There's been research in environment and embodiment although not exactly as I'd like and I will have to incorporate that work but gender and maturation have not been looked at as much. So I'm looking at how do agents cope with gender and maturation.
Silly questions aren't.
The Frame Problem
- Cognitive Wheels: The Frame Problem of Artificial Intelligence (Dennett, 1984)
- The paper is in Thinking about Android Epistemology (Ford et al., 2006)
- The paper is also in The Philosophy of Artificial Intelligence (Boden, 1990)
The Symbol Grounding Problem
- The Symbol Grounding Problem (Harnad, 1990)
- What Computers Still Can't Do, Chapter Six, The Ontological Assumption (Dreyfus, 1992)
Natural Language Processing
- The Natural Language Toolkit
- Question Answering
- Information Extraction
- Information Retrieval
- Document Summarizing
- Cyc, a knowledge base project
- From 2001 to 2001: Common Sense and the Mind of HAL (Lenat, 1996)
- The paper is in Thinking about Android Epistemology (Ford et al., 2006)
- The paper is in HAL's Legacy: 2001's Computer as Dream and Reality (Stork, 1998)
- A Gentle Introduction to Soar (Lehman, Laird, and Rosenbloom, 2006)
The assumption that all knowledge can be formalised into rules
- What Computers Still Can't Do, Chapter Five, The Epistemological Assumption (Dreyfus, 1992)
- What Computers Still Can't Do, Chapter Six, The Ontological Assumption (Dreyfus, 1992)
- What Computers Still Can't Do, Chapter Seven, The Role of the Body in Intelligent Behavior (Dreyfus, 1992)
- What Computers Still Can't Do, Chapter Eight, The Situation: Orderly Behavior Without Recourse to Rules (Dreyfus, 1992)
- What Computers Still Can't Do, Chapter Nine, The Situation as a Function of Human Needs (Dreyfus, 1992)
- How the Body Shapes the Way We Think (Pfeifer & Bongard, 2007)
- Evolution of communication and language using signals, symbols and words, (Cangelosi, 2001)
- Progress in the Simulation of Emergent Communication and Language (Wagner et al., 2003)
- Darwin Machines and the Nature of Knowledge (Plotkin, 1997)
- Why Culture is Common but Cultural Evolution is Rare (Boyd & Richerson, 1995)
- The Selfish Gene (Dawkins, 1976)
How social pressures rather than physical pressures drove the development of intelligence
- The Social Function of Intellect (Humphrey, 1976)
- The paper is in Consciousness Regained: Chapters in Development of Mind (Humphrey, 1984)
- Evolution, Culture, and Consciousness (McNanamara, 2004)
- Human Evolution: An Introduction to Man's Adaptations (Campbell, 1999)
- The Mammal in the Mirror: Understanding Our Place in the Natural World (Barash & Barash, 2001)
- Primeval Kinship: How Pair Bonding Gave Birth to Human Society (Chapais, 2008)
- A review of Chapais (2008) by Linda Stone.
- BBC World Service, Discovery, Human Evolution Podcast by Andrew Luck Baker
"We humans have a very prolonged period of development and dependency before we become adult - not far off twice the length of time it takes a chimp to grow up.
Scientists are beginning to understand this and other oddities of human childhood by studying the remains of fossil children from the past few millions years."
- Dr Barry Bogin, Anthropologist, Loughborough University, UK
- Dr Holly Smith, Anthropologist, University of Michigan
- Dr Kristen Hawkes, Anthropologist, University of Utah
- Dr Tanya Smith, Anthropologist, Harvard University
- Dr Christopher Dean, University College of London
- Dr Jean-Jacques Hublin, Max Planck Institute for Evolutionary Anthropology
- Dr Louise Humphrey, Natural History Museum
Humphrey, 1976, paragraph 10:
``Paradoxically, I would suggest that subsistence technology, rather than requiring intelligence, may actually become a substitute for it. Provided the social structure of the species is such as to allow individuals to acquire subsistence techniques by simple associative learning, there is little need for individual creativity. Thus, the chimpanzees at Gombe, with their superior technological culture may in fact have less need than the neighboring baboons to be individually inventive. Indeed, there might seem on the face of it to be a negative correlation between the capacity of a species and the need to intellectual output. The great apes, demonstratably the most intellectually gifted of all non-human animals, seem on the whole to lead comparatively undemanding lives, less demanding than those not only of lower primates but also of many non-primate species. During two months I spent watching gorillas in the Virunga mountains of Rwanda I could not help being struck by the fact that of all the animals in the forest the gorillas seemed to lead much the simplest existence - food abundant and easy to harvest (provided they knew where to find it), few if any predators (provided they knew how to avoid them) ... little to do in fact (and little done) but eat, sleep and play. And the same is arguably true for natural man. Studies of contemporary Bushmen suggest that the life of hunting and gather, typical of early man, was probably, a remarkably easy one. The `affluent savage' seems to have established a modus vivendi in which, for periods of perhaps five million years, he could afford to be not only physically but intellectually lazy.''
Humphrey, 1976, paragraph 12:
``Why then do the higher primates need to be as clever as they are and, in particular, that much cleverer than other species? What - if it exists - is the natural equivalent of the laboratory tests of intelligence? The answer has, I believe, been ripening on the tree of the preceding discussion. I have suggested that the life of the great apes and man may not require much in the way of practical invention, but it does depend critically on the possession of wide factual knowledge of practical technique and the nature of the habitat. Such knowledge can only be acquired in the context of a social community - a community which provides both a medium for the cultural transmission of information and a protective environment in which individual learning can occur. I propose that the chief role of creative intellect is to hold society together.''