How do we know what we know?

This article considers:

(3) the implications for building artificial intelligence and robots based on the human operating system.

I recently came across a video (on the site http://www.theoryofknowledge.net) that identified the following ‘ways of knowing’:

• Sensory perception
• Memory
• Intuition
• Reason
• Emotion
• Imagination
• Faith
• Language

This list is mainly about mechanisms or processes by which an individual acquires knowledge. It could be supplemented by other processes, for example, ‘meditation’, ‘science’ or ‘history’, each of which provides its own set of approaches to generating new knowledge for both the individual and society as a whole. There are many difference ways in which we come to formulate beliefs and understand the world.

Youtube Video, TOK Ways of Knowing EXPLAINED | Theory of Knowledge Advice, Ivy Lilia, October 2018, 6:16 minutes

In the spirit of working towards a description of the ‘human operating system’, it is interesting to consider how a robot or other Artificial Intelligence (AI), that was ‘running’ the human operating system, would draw on its knowledge and beliefs in order to solve a problem (e.g. resolve some inconsistency in its beliefs). This forces us to operationalize the process and define the control mechanism more precisely. I will work through the above list of ‘ways of knowing’ and illustrate how each might be used.

Let’s say that the robot is about to go and do some work outside and, for a variety of reasons, needs to know what the weather is like (e.g. in deciding whether to wear protective clothing, or how suitable the ground is for sowing seeds or digging up for some construction work etc.) .

First it might consult its senses. It might attend to its visual input and note the patterns of light and dark, comparing this to known states and conclude that it was sunny. The absence of the familiar sound patterns (and smell) of rain might provide confirmation. The whole process of matching the pattern of data it is receiving through its multiple senses, with its store of known patterns, can be regarded as ‘intuitive’ because it is not a reasoning process as such. In the Khanemman sense of ‘system 1’ thinking, the robot just knows without having to perform any reasoning task.

Youtube Video, System 1 and System 2, Stoic Academy, February 2017, 1:26 minutes

The knowledge obtained from matching perception to memory can nevertheless be supplemented by reasoning, or other forms of knowledge that confirm or question the intuitively-reached conclusion. If we introduce some conflicting knowledge, e.g. that the robot thinks it’s the middle of the night in it’s current location, we then create a circumstance in which there is dissonance between two sources of knowledge – the perception of sunlight and the time of day. This assumes the robot has elaborated knowledge about where and when the sun is above the horizon and can potentially shine (e.g. through language – see below).

In people the dissonance triggers the emotional state of ‘surprise’ and the accompanying motivation to account for the contradiction.

Youtube Video, Cognitive Dissonance, B2Bwhiteboard, February 2012, 1:37 minutes

Likewise, we might label the process that causes the search for an explanation in the robot as ‘surprise’. An attempt may be made to resolve this dissonance through Kahneman’s slower, more reasoned, system 2 thinking. Either the perception is somehow faulty, or the knowledge about the time of day is inaccurate. Maybe the robot has mistaken the visual and audio input as coming from its local senses when in fact the input has originated from the other side of the world. (Fortunately, people do not have to confront the contradictions caused by having distributed sensory systems).

Probably in the course of reasoning about how to reconcile the conflicting inputs, the robot will have had to run through some alternative possible scenarios that could account for the discrepancy. These may have been generated by working through other memories associated with either the perceptual inputs or other factors that have frequently led to mis-interpretations in the past. Sometimes it may be necessary to construct unique possible explanations out of component part explanations. Sometimes an explanation may emerge through the effect of numerous ideas being ‘primed’ through the spreading activation of associated memories. Under these circumstances, you might easily say that the robot was using it’s imagination in searching for a solution that had not previously been encountered.

Youtube Video, TEDxCarletonU 2010 – Jim Davies – The Science of Imagination, TEDx Talks, September 2010, 12:56 minutes

Lastly, to faith and language as sources of knowledge. Faith is different because, unlike all the other sources, it does not rely on evidence or proof. If the robot believed, on faith, that the sun was shining, any contradictory evidence would be discounted, perhaps either as being in error or as being irrelevant. Faith is often maintained by others, and this could be regarded as a form of evidence, but in general if you have faith in or trust something, it is at least filling the gap between the belief and the direct evidence for it.

Here is a religious account of faith that identifies it with trust in the reliability of God to deliver, where the main delivery is eternal life.

Youtube video, What is Faith – Matt Morton – The Essence of Faith – Grace 360 conference 2015,Grace Bible Church, September 2015, 12:15 minutes

Language as a source of evidence is a catch-all for the knowledge that comes second hand from the teachings and reports of others. This is indirect knowledge, much of which we take on trust (i.e. faith), and some of which is validated by direct evidence or other indirect evidence. Most of us take on trust that the solar system exists, that the sun is at the centre, and that earth is in the third orbit. We have gained this knowledge through teachers, friends, family, tv, radio, books and other sources that in their turn may have relied on astronomers and other scientist who have arrived at these conclusions through observation and reason. Few of us have made the necessary direct observations and reasoned inferences to have arrived at the conclusion directly. If our robot were to consult databases of known ‘facts’, put together by people and other robots, then it would be relying on knowledge through this source.

Pitfalls

People like to think that their own beliefs are ‘true’ and that these beliefs provide a solid basis for their behaviour. However, the more we find out about the psychology of human belief systems the more we discover the difficulties in constructing consistent and coherent beliefs, and the shortcomings in our abilities to construct accurate models of ‘reality’. This creates all kinds of difficulties amongst people in their agreements about what beliefs are true and therefore how we should relate to each other in peaceful and productive ways.

If we are now going on to construct artificial intelligences and robots that we interact with and have behaviours that impact the world, we want to be pretty sure that the beliefs a robot develops still provide a basis for understanding their behaviour.

Unfortunately, every one of the ‘ways of knowing’ is subject to error. We can again go through them one by one and look at the pitfalls.

Sensory perception: We only have to look at the vast body of research on visual illusion (e.g. see ‘Representations of Reality – Part 1’) to appreciate that our senses are often fooled. Here are some examples related to colour vision:

Youtube Video, Optical illusions show how we see | Beau Lotto,TED, October 2009, 18:59 minutes

Furthermore, our perceptions are heavily guided by what we pay attention to, meaning that we can miss all sorts of significant and even life-threatening information in our environment. Would a robot be similarly misled by its sensory inputs? It’s difficult to predict whether a robot would be subject to sensory illusions, and this might depend on the precise engineering of the input devices, but almost certainly a robot would have to be selective in what input it attended to. Like people, there could be a massive volume of raw sensory input and every stage of processing from there on would contain an element of selection and interpretation. Even differences in what input devices are available (for vision, sound, touch or even super-human senses like perception of non-visual parts of the electromagnetic spectrum), will create a sensory environment (referred to as the ‘umwelt’ or ‘merkwelt’in ethology) that could be quite at variance with human perceptions of the world.

YouTube Video, What is MERKWELT? What does MERKWELT mean? MERKWELT meaning, definition & explanation, The Audiopedia, July 2017, 1:38 minutes

Memory: The fallibility of human memory is well documented. See, for example, ‘The Story of Your Life’, especially the work done by Elizabeth Loftus on the reliability of memory. A robot, however, could in principle, given sufficient storage capacity, maintain a perfect and stable record of all its inputs. This is at variance with the human experience but could potentially mean that memory per se was more accurate, albeit that it would be subject to variance in what input was stored and the mechanisms of retrieval and processing.

Intuition and reason: This is the area where some of the greatest gains (and surprises) in understanding have been made in recent years. Much of this progress is reported in the work of Daniel Kahneman that is cited many times in these writings. Errors and biases in both intuition (system 1 thinking) and reason (system 2 thinking) are now very well documented. A long list of cognitive biases can be found at:

https://en.wikipedia.org/wiki/List_of_cognitive_biases

Would a robot be subject to the same type of biases? It is already established that many algorithms, used in business and political campaigning, routinely build in the biases, either deliberately or inadvertently. If a robot’s processes of recognition and pattern matching are based on machine learning algorithms that have been trained on large historical datasets, then bias is virtually guaranteed to be built into its most basic operations. We need to treat with great caution any decision-making based on machine learning and pattern matching.

Youtube Vide, Cathy O’Neil | Weapons of Math Destruction, PdF YouTube, June 2015, 12:15 minutes

As for reasoning, there is some hope that the robustness of proofs that can be achieved computationally may save the artificial intelligence or robot from at least some of the biases of system 2 thinking.

Emotion: Biases in people due to emotional reactions are commonplace. See, for example:

Youtube Video, Unconscious Emotional Influences on Decision Making, The Rational Channel, February 2017, 8:56 minutes

However, it is also the case that emotions are crucial in decision–making. Emotions often provide the criteria and motivation on which decisions are made and without them, people can be severely impaired in effective decision-making. Also, emotions provide at least one mechanism for approaching the subject of ethics in decision-making.

Youtube Video, When Emotions Make Better Decisions – Antonio Damasio, FORA.tv, August 2009, 3:22 minutes

Can robots have emotions? Will robots need emotions to make effective decisions? Will emotions bias or impair a robot’s decision-making. These are big questions and are only touched on here, but briefly, there is no reason why emotions cannot be simulated computationally although we can never know if an artificial computational device will have the subjective experience of emotion (or thought). Probably some simulation of emotion will be necessary for robot decision-making to align with human values (e.g. empathy) and, yes, a side-effect of this may well be to introduce bias into decision-making.

For a selection of BBC programmes on emotions see:
http://www.bbc.co.uk/programmes/topics/Emotions?page=1

Imagination: While it doesn’t make much sense to talk about ‘error’ when it comes to imagination, we might easily make value-judgments about what types of imagination might be encouraged and what might be discouraged. Leaving aside debates about how, say excessive experience of violent video games, might effect imagination in people, we can at least speculate as to what might or should go on in the imagination of a robot as it searches through or creates new models to help predict the impacts of its own and others behaviours.

A big issue has arisen as to how an artificial intelligence can explain its decision-making to people. While AI based on symbolic reasoning can potentially offer a trace describing the steps it took to arrice at a conclusion, AIs based on machine learning would be able to say little more than ‘I recognized the pattern as corresponding to so and so’, which to a person is not very explanatory. It turns out that even human experts are often unable to provide coherent accounts of their decision-making, even when they are accurate.

Having an AI or robot account for its decision-making in a way understandable to people is a problem that I will address in later analysis of the human operating system and, I hope, provide a mechanism that bridges between machine learning and more symbolic approaches.

Faith: It is often said that discussing faith and religion is one of the easiest ways to lose friends. Any belief based on faith is regarded as true by definition, and any attempt to bring evidence to refute it, stands a good chance of being regarded as an insult. Yet people have different beliefs based on faith and they cannot all be right. This not only creates a problem for people, who will fight wars over it, but it is also a significant problem for the design of AIs and robots. Do we plug in the Muslim or the Christian ethics module, or leave it out altogether? How do we build values and ethical principles into robots anyway, or will they be an emergent property of its deep learning algorithms. Whatever the answer, it is apparent that quite a lot can go badly wrong if we do not understand how to endow computational devices with this ‘way of knowing’.

Language: As observed above, this is a catch-all for all indirect ‘ways of knowing’ communicated to people through media, teaching, books or any other form of communication. We only have to consider world wars and other genocides to appreciate that not everything communicated by other people is believable or ethical. People (and organizations) communicate erroneous information and can deliberately lie, mislead and deceive.

We strongly tend to believe information that comes from the people around us, our friends and associates, those people that form part of our sub-culture or in-group. We trust these sources for no other reason than we are familiar with them. These social systems often form a mutually supporting belief system, whether or not it is grounded in any direct evidence.

Youtube Video, The Psychology of Facts: How Do Humans (mis)Trust Information?, YaleCampus, January 2017

Taking on trust the beliefs of others that form part of our mutually supporting social bubble is a ‘way of knowing’ that is highly error prone. This is especially the case when combined with other ‘ways of knowing’, such as faith, that in their nature cannot be validated. Will robot communities develop, who can talk to each other instantaneously and ‘telepathically’ over wireless connections, also be prone to the bias of groupthink?

The validation of beliefs

So, there are multiple ways in which we come to know or believe things. As Descartes argued, no knowledge is certain (see ‘It’s Like This’). There are only beliefs, albeit that we can be more sure of some that others, normally by virtue of their consistency with other beliefs. Also, we note that our beliefs are highly vulnerable to error. Any robot operating system that mimics humans will also need to draw on the many different ‘ways of knowing’ including a basic set of assumptions that it takes to be true without necessarily any supporting evidence (it’s ‘faith’ if you like). There will also need to be many precautions against AIs and robots developing erroneous or otherwise unacceptable beliefs and basing their behaviours on these.

There is a mechanism by which we try to reconcile differences between knowledge coming from different sources, or contradictory knowledge coming from the same source. Most people seem to be able to tolerate a fair degree of contradiction or ambiguity about all sorts of things, including the fundamental questions of life.

Youtube Video, Defining Ambiguity, Corey Anton, October 2009, 9:52 minutes

We can hold and work with knowledge that is inconsistent for long periods of time, but nevertheless there is a drive to seek consistency.

In the description of the human operating system, it would seem that there are many ways in which we establish what we believe and what beliefs we will recruit to the solving of any particular problem. Also, the many sources of knowledge may be inconsistent or contradictory. When we see inconsistencies in others we take this as evidence that we should doubt them and trust them less.

Youtube Video, Why Everyone (Else) is a Hypocrite, The RSA, April 2011, 17:13 minutes

However, there is, at least, a strong tendency in most people, to establish consistency between beliefs (or between beliefs and behaviours), and to account for inconsistencies. The only problem is that we are often prone to achieve consistency by changing sound evidence-based beliefs in preference to the strongly held beliefs based on faith or our need to protect our sense of self-worth.

Youtube Video, Cognitive dissonance (Dissonant & Justified), Brad Wray, April 2011. 4:31 minutes

From this analysis we can see that building AIs and robots is fraught with problems. The human operating system has evolved to survive, not to be rational or hold high ethical values. If we just blunder into building AIs and robots based on the human operating system we can potentially make all sorts of mistakes and give artificial agents power and autonomy without understanding how their beliefs will develop and the consequences that might have for people.

Fortunately there are some precautions we can take. There are ways of thinking that have been developed to counter the many biases that people have by default. Science is one method that aims to establish the best explanations based on current knowledge and the principle of simplicity. Also, critical thinking has been taught since Aristotle and fortunately many courses have been developed to spread knowledge about how to assess claims and their supporting arguments.

Youtube Video, Critical Thinking: Issues, Claims, Arguments, fayettevillestatenc, January 2011

Implications

To summarise:

Sensory perception – The robot’s ‘umwelt’ (what it can sense) may well differ from that of people, even to the extent that the robot can have super-human senses such as infra-red / x-ray vision, super-sensitive hearing and smell etc. We may not even know what it’s perceptual world is like. It may perceive things we cannot and miss things we find obvious.

Memory – human memory is remarkably fallible. It is not so much a recording, as a reconstruction based on clues, and influenced by previously encountered patterns and current intentions. Given sufficient storage capacity, robots may be able to maintain memories as accurate recording of the states of their sensory inputs. However, they may be subject to similar constraints and biases as people in the way that memories are retrieved and used to drive decision-making and behaviour.

Intuition – if the robot’s pattern-matching capabilities are based on the machine learning of historical training sets then bias will be built into its basic processes. Alternatively, if the robot is left to develop from it’s own experience then, as with people, great care has to be taken to ensure it’s early experience will not lead to maladaptive behaviours (i.e. behaviours not acceptable to the people around it).

Reason – through the use of mathematical and logical proofs, robots may well have the capacity to reason with far greater ability than people. They can potentially spot (and resolve) inconsistencies arising out of different ‘ways of knowing’ with far greater adeptness than people. This may create a quite different balance between how robots make decisions and how people do using emotion and reason in tandem.

Emotion – human emotion are general states that arise in response to both internal and external events and provide both the motivation and the criteria on which decisions are made. In a robot, emerging global states could also potentially act to control decision-making. Both people, and potentially robots, can develop the capacity to explicitly recognize and control these global states (e.g. as when suppressing anger). This ability to reflect, and to cause changes in perspective and behaviour, is a kind of feedback loop that is inherently unpredictable. Not having sufficient understanding to predict how either people or robots will react under particular circumstances, creates significant uncertainty.

Imagination – much the same argument about predictability can be made about imagination. Who knows where either a person’s or a robot’s imagination may take them? Chess computers out-performed human players because of their capacity to reason in depth about the outcomes of every move, not because they used pattern-matching based on machine learning (although it seems likely that this approach will have been tried and succeeded by now). Robots can far exceed human capacities to reason through and model future states. A combination of brute force computing and heuristics to guide search, may have far-reaching consequences for a robot’s ability to model the world and predict future outcomes, and may far exceed that of people.

Faith – faith is axiomatic for people and might also be for robots. People can change their faith (especially in a religious, political or ethical sense) but more likely, when confronted with contradictory evidence or sufficient need (i.e. to align with a partner’s faith) people with either ignore the evidence or find reasons to discount it. This way can lead to multiple interpretations of the same basic axioms, in the same way as there are many religious denominations and many interpretations of key texts within these. In robots, Asimov’s three laws of robotics would equate to their faith. However, if robots used similar mechanisms as people (e.g. cognitive dissonance) to resolve conflicting beliefs, then in the same way as God’s will can be used to justify any behaviour, a robot may be able to construct a rationale for any behaviour whatever its axioms. There would be no guarantee that a robot would obey its own axiomatic laws.

Communication – The term language is better labeled ‘communication’ in order to make it more apparent that it extends to all methods by which we ‘come to know’ from sources outside ourselves. Since communication of knowledge from others is not direct experience, it is effectively taken on trust. In one sense it is a matter of faith. However, the degree of consistency across external sources and between what is communicated (i.e. that a teacher or TV will re-enforce what a parent has said etc.) and between what is communicated and what is directly observed (for example, that a person does what he says he will do) will reveal some sources as more believable than others. Also we appeal to motive as a method of assessing degree of trust. People are notoriously influenced by the norms, opinions and behaviours of their own reference groups. Robots with their potential for high bandwidth communication could, in principle, behave with the same psychology of the crowd as humans, only much more rapidly and ‘single-mindedly’. It is not difficult to see how the Dr Who image of the Borg, acting a one consciousness, could come about.

Other Ways of Knowing

It is worth considering just a few of the many other ‘ways’ of knowing’ not considered above, partly because some of these might help mitigate some of the risks of human ‘ways of knowing’ .

Science – Science has evolved methods that are deliberately designed to create impartial, robust and consistent models and explanations of the world. If we want robots to create accurate models, then an appeal to scientific method is one approach. In science, patterns are observed, hypotheses are formulated to account for these patterns, and the hypotheses are then tested as impartially as possible. Science also seeks consistency by reconciling disparate findings into coherent overall theories. While we may want robots to use scientific methods in their reasoning, we may want to ensure that robots do not perform experiments in the real world simply for the sake of making their own discoveries. An image of concentration camp scientists comes to mind. Nevertheless, in many small ways robots will need to be empirical rather than theoretical in order to operate at all.

Argument – Just like people, robots of any complexity will encounter ambiguity and inconsistencies. These will be inconsistencies between expectation and actuality, between data from one way of knowing and another (e.g. between reason and faith, or between perception and imagination etc.), or between a current state and a goal state. The mechanisms by which these inconsistencies are resolved will be crucial. The formulation of claims; the identification, gathering and marshalling of evidence; the assessment of the relevance of evidence; and the weighing of the evidence, are all processes akin to science but can cut across many ‘ways of knowing’ as an aid to decision making. Also, this approach may help provide explanations of a robot’s behaviour that would be understandable to people and thereby help bridge the gap between opaque mechanisms, such as pattern matching, and what people will accept as valid explanations.

Meditation – Meditation is a place-holder for the many ways in which altered states of consciousness can lead to new knowledge. Dreaming, for example, is another altered state that may lead to new hypotheses and models based on novel combination of elements that would not otherwise have been brought together. People certainly have these altered states of consciousness. Could there be an equivalent in the robot, and would we want robots to indulge in such extreme imaginative states where we would have no idea what they might consist of? This is not to necessarily attribute consciousness to robots, which is a separate, and probably meta-physical question.

Theory of mind – For any autonomous agent with its own beliefs and intentions, including a robot, it is crucial to its survival to have some notion of the intentions of other autonomous agents, especially when they might be a direct threat to survival. People have sophisticated but highly biased and error-prone mechanisms for modelling the intentions of others. These mechanisms are particularly alert for any sign of threat and, as a proven mechanism, tend to assume threat even when none is present. The people that did not do this, died out. Work in robotics already recognizes that, to be useful, robots have to cooperate with people and this requires some modelling of their intentions. As this last video illustrates, the modelling of others intentions is inherently complex because it is recursive.

YouTube Video, Comprehending Orders of Intentionality (for R. D. Laing), Corey Anton, September 2014, 31:31 minutes

If there is a conclusion to this analysis of ‘ways of knowing’ it is that creating intelligent, autonomous mechanisms, such as robots and AIs, will have inherently unpredictable consequences, and that, because the human operating system is so highly error-prone and subject to bias, we should not necessarily build them in our own image.