Paper on Loops in AI and Consciousness

PREPRINT DRAFT

Title: A Characterisation of processing loops in AI and biological neural networks and its implications for understanding Consciousness

Abstract

Shows that any sufficiently advanced processing system requires loopy processing, which requires regulation, which requires a self-model, and that this structure can lead to self-referential conclusions about an agents own mental faculties and their involvement in its own agency.

Introduction

Any computational system is limited in the complexity that it can handle within a single computational step. For embodied agents, this appears as a limit on the environmental complexity that they can sufficiently model and respond to within a single "time step" (citation needed). For more complex problems, multiple steps of processing are required in order to determine the next physical action. Such multiple processing steps may entail, for example, further analysis of the environment in order to better model its state; or it may entail action planning over multiple iterations.

In biology, this provides scope for evolutionary pressures to trade off between a more energy hungry complex brain and a simpler less energy intensive one that takes longer to make some decisions.

An agent that regulates its environment operates within a system containing environment state S_env, which changes with some ambient dynamics D_env(t). The agent performs action A_env against the environment in order to regulate it towards some target. The environment state outcome O_env is influenced by both D_env(t) and A_env. This can be summarised as such:

S_env + D_env(t) + A_env = O_env

According to the good regulator theorem, if the agent is to regulate the environment state it must be a "model of the system" (Conant & Ashby, 1970). Furthermore, we can say that the efficiency of the agent to regulate its environment depends on its accuracy in modelling the system. Errors in the accuracy of the model result in errors in the regulation of the system. In learning agents, those errors can be used for subsequent training of its model.

An embodied agent with complex actions requires an additional level of regulation. Not only must it regulate its external environment, it must also regulate its own physical state. This includes both maintaining homeostasis and controlling action for efficiency and effectiveness. Such an agent thus operates in a system that additionally has body state S_body with ambient dynamics D_body(t). The agent performs action A_body against its body, producing outcome O_body, summarised as follows:

S_body + D_body(t) + A_body = O_body

The agent's actions are performed in order to regulate it towards some target, which is dynamically inferred based on its requirement for body homeostasis and for environment action A_env. The inference of the required action at any given moment is based upon a model, and that model is updated from past errors in a learning agent. Psychology and Neuroscience refer to that model in mammals as the body schema (Proske & Gandevia, 2012).

Agents that incorporate multi-step processing have a third kind of action: one that changes its internal data state without affecting its physical state. Importantly, this system also requires regulation for the same reasons as for environment and body, but such non-physical actions may not elicit any change to S_body or S_env. Thus the agent must regulate its non-physical state S_mind, having ambient dynamics D_mind(t). In order to do so it performs action A_mind, producing outcome O_mind, summarised as follows:

S_mind + D_mind(t) + A_mind = O_mind

The agent's non-physical actions are performed in order to regulate towards some target, which is dynamically inferred based on its requirement for environment action A_env, body action A_body, and possibly for some form of non-physical homeostasis. Like for environment and body regulation, in order for the agent to efficiently regulate its mind state, it must model its behaviour. This suggests that it must incorporate a functional equivalent of the body schema, which this paper refers to as the mind schema. Other research has drawn similar conclusions (Graziano, 2017).

By way of example of the importance of such mind regulation, consider the case of fluent aphasia, caused by damage to the Wernicke's area of the brain. Individuals with fluent aphasia can easily produce speech, but it is typically full of many meaningless words and often unnecessarily long winded. Wernicke's area is associated with language comprehension and, as such, provides a corrective mechanism during speech production in a neurotypical individual (Wernicke's area).

Visceral Loop

This paper introduces the concept of a visceral loop as a characterisation of processing within a looping biological or AI agent. The visceral loop is so named because it refers to an agent concluding that it experiences consciousness "in a visceral way". It identifies that a processing loop with sufficient representational capabilities can, at the most optimum, conclude itself as conscious within three iterations of the loop. Each of those iterations have specific characteristics, and the visceral loop characterises thought as falling into one of those three iterations.

Let:

• E be the agent's set of beliefs about the external world
• B be the agent's set of beliefs about its own physical body (drawn from the body schema) and of bodies in general, and if it has a concept of "I" then this set includes a belief that relates other body beliefs to "I"
• M be the agent's set of beliefs about its own mind (drawn from the mind schema) and of minds in general, and if it has a concept of "I" then this set includes a belief that relates other mind beliefs to "I"
• f(..) be the function executed by the agent on the specified inputs in order to draw inferences

M can be thought of as an agent's "theory of mind", because it relates to not such itself but also its ability to predict the hidden mental state of others.

Iteration 1:

Iteration 1 represents the most common kind of data processing, such as spending multiple processing cycles to refine the identification of something within the visual field. While an agent's mind schema may be used to regulate the thought process, the result of Iteration 1 never makes any reference to it.

Let x be an inference produced as the result of a processing step, such that it does not draw any reference to M (ie: if x is a value then x ∉ M, or if x is a relation of two values then x = relation(a, b) such that a ∉ M and b ∉ M, or if x is a relation involving a set then x = relation(a, B) such that a ∉ M and B ⊄ M). Given some sense input or past state s, a processing step is characterised as visceral loop Iteration 1 if it is of the following form:

• f(s, E ∪ B ∪ M) -> x

Iteration 2:

Iteration 2 processing steps draw conclusions that relate past thought actions and conclusions to the agent's theory of mind and to the agent's concept of its identity. For example, concluding that a past data state or non-physical action is classified as "thought", concluding whether the primary source of a past data state was external or internal, or relating the fact of an internal source to the agent's concept of "I".

Iteration 2 requires an agent to have sufficient representational capabilities to produce inferences that represent relations involving M. Given some prior inference y, a processing step is characterised as visceral loop Iteration 2 if it is of the following form, and the relation with respect to M is non-empty, and it can not be characterised as Iteration 3:

• f(y, E ∪ B ∪ M) -> relation(y, M)

Iteration 3:

Iteration 3 is a special case of what would otherwise be Iteration 2, but it implies stricter requirements on the agent's introspective and representational capabilities. Iteration 3 covers the ability for the agent to develop a summary of its own mental capabilities (ie: some subset m ⊂ M), and to consider that in relation to its conception of mental capabilities in general (ie: M).

Given some prior inference relation(z, M), and some subset of beliefs m ⊂ M, a processing step is characterised as visceral loop Iteration 3 if it is of the following form and the relation with respect to M is non-empty:

• f(relation(z, M), E ∪ B ∪ M) -> relation(m, M)

Consciousness

The concept of the visceral loop provides a framework for classifying the capabilities of different processing systems. It also has important implications for understanding consciousness, particularly in its access consciousness interpretation (Block, 1995).

Two examples of the descriptive power of the visceral loop in relation to consciousness are presented here.

Example 1

In this first example, the visceral loop is applied to understand the thought processes whereby an individual concludes themselves as conscious. Consider the following sequence of internal mental observations:

1. "What's that red blob in the tree? Oh, it's an apple".
2. "Oh, those thoughts just came from my mind, and not from the outside world".
3. "That's what consciousness is. I am conscious".

The first observation is a straightforward example of Iteration 1 that does not make any reference to the agent's theory of mind (of their own mind or of others). The concepts of "red", "blob", "tree" and "apple" are all contained within the set E, and thus the inference in relation to the visual field sense input s is of the form x_1 = relation(s, E).

The second observation contains two examples of Iteration 2 inferences. In the first, the individual's processing capabilities have selected attentional focus upon the prior Iteration 1 inference, and have drawn a subsequent inference about it as being data that can be classified as a "thought". As beliefs about "thought" are contained within M, this is an inference of the form x_2 = relation(x_1, M). In the second, the individual draws a subsequent inference about the source of the Iteration 1 inference as being their own mind. The individual's ability to classify inferences in relation to themselves also depends upon M, and the inference is of the form x_3 = relation(x_1, M).

The third observation draws upon the individual having an a priori conception about consciousness in general, denoted by m_c ⊂ M. The individual compares its prior Iteration 2 inferences x_2 and x_3 to m_c, and produces an inference that x_2 and x_3 together satisfy the requirements for consciousness. This is another iteration 2 inference of the form x_4 = relation(x_2 & x_3, m_c). Finally, the individual relates m_c, the belief of consciousness in general, to itself, which again depends on M. That final inference is thus an Iteration 3 inference in the form x_5 = relation(m_c, M).

Example 2

As a second example of the descriptive power of the viscera loop, a theorem is presented here about the nature of consciousness.

First an axiomatic baseline must be established. The author is unable to think of any rationale way in which they may consciously experience something and yet be unable to subsequently think about that experience and to know that they are thinking about that experience. Thus, it would seem that being able to knowingly think about our conscious experiences is a fundamental component of consciousness. The following claims are derived from this statement, without further proof:

Claim 1:

• All conscious experience is subsequently available for further thought.

Claim 2:

• For all thought about conscious experience, the individual can identify that thought as being their own.

Note that these claims do not assume that all conscious experience is actually thought about; only that it is in principle available for such thought. Additionally, no assumption is made about whether other kinds of thought are consciously experienced or not.

Theorem 1:

• the content of conscious experience is upper bounded by the data about which visceral loop iteration 2 inferences can be produced.

Proof:

• The content of conscious experience refers to the set of data represented and/or processed within the brain which is consciously experienced by that individual, in distinction to other data represented and/or processed in the brain which is not consciously experienced.
• As per claim 1, all of conscious experience must be available for producing subsequent inferences about those conscious experiences.
• As per claim 2, the individual must be able to identify that they produced those inferences.
• In order for an individual to identify an inference as being their own, they must have some beliefs about their inference capabilities and how they relate to themselves as an individual entity. This is included in the set M, which iteration 2 produces inferences in relation to.
• Imagine some supposed experience, and an inference i produced about that experience. Additionally imagine that an iteration 2 inference cannot be produced about i, for example, due to some incompatibility of structure, lack of data path to iteration 2 processing capabilities, or inherent limitation in iteration 2 processing capabilities.
• The inference i cannot be identified in relation to the individual. As such, the supposed experience fails on Claim 2 and i must be in actual fact an inference about a non-conscious experience.
• Thus, any experience that can only lead to inferences which cannot be included in an iteration 2 inference is not a conscious experience.

Summary and Conclusions

Visceral loop explains why fRMI studies have shown that we become aware of a decision after its made (tbd: citation needed). Because it takes extra processing cycles to consciously consider the fact of the decision being made. In short: we can only think about one thing at a time, so the decision itself and thought about the decision require separate steps.

The visceral loop can be used to explain how someone concludes themselves as conscious. It can also be used to classify the kinds of thought that occur within an agent, and the kinds of thought that it's possible for an agent to have. For example, it may be the case that simpler organisms only ever operate with Iteration 1 thought.

References

Block, N. (1995). On a confusion about a function of consciousness. Brain and Behavioral Sciences, 18(2), 227–247. https://doi.org/10.1017/S0140525X00038188. [Full Text]

Conant, R. C., and Ashby, W. R. (1970). Every good regulator of a system must be a model of that system. Int. J. Systems Sci., vol. 1, No. 2, pp 89-97. https://doi.org/10.1080/00207727008920220. [Full Text]

Graziano, M. S. A. (2017). The Attention Schema Theory: A Foundation for Engineering Artificial Consciousnes. Front. Robot. AI. https://doi.org/10.3389/frobt.2017.00060.

Proske, U., and Gandevia, S. C. (2012). The Proprioceptive Senses: Their Roles in Signaling Body Shape, Body Position and Movement, and Muscle Force. Physiological Reviews 2012 92:4, pp 1651-1697. https://doi.org/10.1152/physrev.00048.2011.

Wernicke's area. (n.d.). In Wikepedia. https://en.wikipedia.org/wiki/Wernicke%27s_area.