Dear Simon,

Good to hear from you. Your comment made my day, as you indeed captured the essence of my essay. The animats are such a great model system as they force one to consider the implementation of suggested potential solutions to intrinsic meaning, based on "information processing", "models about the environment", etc. Most of the time these ideas are presented abstractly, sound...

view entire post

Dear Simon,

Good to hear from you. Your comment made my day, as you indeed captured the essence of my essay. The animats are such a great model system as they force one to consider the implementation of suggested potential solutions to intrinsic meaning, based on "information processing", "models about the environment", etc. Most of the time these ideas are presented abstractly, sound really great, and resonate with many people, but on closer consideration fail to pass the implementation test.

With respect to the question of dynamics vs. computation, and whether the gas in the seminar room performs a computation, David Chalmers addressed a similar point here: Chalmers, D.J. (1996). Does a rock implement every finite-state automaton? Synthese 108, 309–333. It's about mapping any kind of computation onto a system that can assume various states. I think the conclusion is that in order to say that two systems perform the same computation, it is not sufficient for them to have a dynamical sequence of states that can be mapped onto each other. Instead, there has to be a mapping of all possible state transitions, which basically means the same causal structure, i.e. a mapping of the causal implementation.

Along these lines, computation, in my view, requires knowing all counterfactuals. I.e. to know that an AND gate is an AND gate and performs the AND computation, it is not sufficient to know that it transitions from 11 -> 1. One needs to know all possible input states (all possible counterfactuals) and the resulting output state.

With respect to game theory, I know that Chris Adami and Arend Hintze successfully applied the animats to solve games such as the prisoner's dilemma, but we haven't measured their integrated information in such environments yet. Memory does play a crucial role for evolving integration. Games that can be solved merely by "reflexes" based on current sensory inputs will produce mostly feedforward systems. Evaluating the animats on multiple-game versions with different pay-off matrices should indeed be interesting. Thank you for bringing that up! Relatedly, we are currently evaluating "social" animats that can sense other agents and mostly replicated the past results.

Best regards,

Larissa

view post as summary

Dear Tommaso,

Thank you very much for your comment and insightful questions. By contrast to something like measures of local entropy decreases, the IIT formalism does not just yield a quantity (integrated information) but also a characterization of the system, its cause-effect structure, which is the set of all system mechanisms that constrain the past and future states of the system...

view entire post

Dear Tommaso,

Thank you very much for your comment and insightful questions. By contrast to something like measures of local entropy decreases, the IIT formalism does not just yield a quantity (integrated information) but also a characterization of the system, its cause-effect structure, which is the set of all system mechanisms that constrain the past and future states of the system itself. The mechanisms specify everything within the system that makes a difference to the system itself. In this way I don't just find out whether there is an intrinsic system in the first place, but also get a picture of its capacity for 'understanding', of what matters to the system and what cannot possibly matter to the system because it doesn't have the right mechanism to pick it up. I hope this helped. Characterizing more precisely how intrinsic meaning could arise from the cause-effect structure is work in progress in our lab.

I completely agree on your point regarding 'acting' vs. 'reacting'. In fact, this is basically the topic of my fellowship project for the next 3 years. Our goal is to quantify when and how strongly an action was caused by the system as opposed to the environment. Autonomous action here means that the system's action is not entirely driven by its current sensory inputs from the environment. Making a choice based on memory, however, would count as autonomous. If you look at the website (ref [8]) under the 'task simulation' tab and set it to trial 55, for example, you can see that the animat already shows a little bit of autonomous behavior in that sense. It first follows the block, then goes in the other direction, then follows again. This means that its behavior didn't just depend on the sensory inputs, but is context-dependent on its own internal state. This is a little different than your definition of autonomy ('doing something for no reason'). That could be achieved with just a little noise inside the system.

As for your last question: the issue with a trial-by-trial correlation of Phi and fitness is that an animat can always have more Phi than is necessary, as there is no real cost on being more integrated than needed the way the simulation is set up. Moreover, fitness can also increase due to e.g. a connection from a sensor to a motor (a reflex), which would not increase the integration. In practice, for complex tasks, there should be a lower limit on the amount of integration that is required for a given task given constraints on the number of elements, connections, and time available to perform the computations as integrated systems are more economical than feedforward systems.

Best regards,

Larissa

view post as summary

Dear Rajiv,

Thank you very much for your thorough reading of my essay!

Let me try to address some of the issues you raised. Many have to do with (1) the difference between the ‘intrinsic information’ that I outlined, which is a measure of causal constraint, and the classic notion of information: Shannon information, (2) a distinction between causal constraint and predictability,...

view entire post

Dear Rajiv,

Thank you very much for your thorough reading of my essay!

Let me try to address some of the issues you raised. Many have to do with (1) the difference between the ‘intrinsic information’ that I outlined, which is a measure of causal constraint, and the classic notion of information: Shannon information, (2) a distinction between causal constraint and predictability, and (3) taking the intrinsic perspective of the system itself vs. the typical extrinsic perspective of an observer of the system (as is the common perspective in physics).

Causal autonomy and (in)determinism:

Of course, in a deterministic system, given the full state of the system at some point in time, the system’s future evolution is entirely predictable. What happens happens. However, performing this prediction takes a “god’s” perspective of the entire system.

The notion of causal autonomy defined in my essay applies from the intrinsic perspective of the system, which may be an open subsystem S of a larger system Z. What is measured is the mechanistic constraints of S on the direct past and future state of the subsystem S itself, using a counterfactual notion of causation. Roughly, this notion of causation is about what within the entire state of the system actually constraints what (not all mechanisms constrain each other all the time). So locally, right now, constraints on the system can come from within the system S and/or from outside the system. If all parts of the system S constrain each other mechanistically, above a background of constraints from outside of the system, S is causally autonomous to some degree (which can be measured).

Indeterminism within the system will actually only result in less intrinsic constraints.

Representation:

I think we can connect here in the sense that what I am arguing in my essay is that any emerged semantics have to come from the intrinsic cause-effect structure of the system itself. The crucial point is that, whatever intrinsic meaning there is does not mean what it means because of a correlation with the outside world. There may be (and should be) a correlation between the world and the intrinsic cause-effect structure of the system, however the intrinsic semantics (what it means for the system itself) must arise purely from the intrinsic structure, and cannot arise because of the correlation. Only external observers can make use of a correlation between something external and something internal to the system, not the system itself.

Mechanism and system states:

The way a mechanism is define here is specifically a set of elements within a physical system (with cause-effect power). A mechanism must have at least two states. So M is a physical entity (it can be observed and manipulated) and m is its current state. A neuron or a logic gate for example could be a mechanism.

Z is a system, let’s say 4 binary logic gates {ABCD}, which at any moment is in a state, let’s say ABCD = 1011 in its current state.

Intrinsic information and semantics:

You wrote:

“p(Zt-1 | mt) would be the probability of finding elements of Z in a

particular specification of states at a time t-1 given a specific

configuration of Logic Gates (LG). And p(Zt-1) would be the LG

independent probability of the same set of states irrespective.

Furthermore, as per the referred statement, the differences between

the two probabilities constitutes the information of the cause and

effect.”

This is correct. Let’s say mt is an AND gate ‘A’ in state ON with 2 inputs (‘B’ and ‘C’). p(BCt-1|A=1) is 1 for state BC = 11 and 0 for all other states (BC = 00, 01, 10). Now the crucial point is that this is what it means to be an AND gate in state ‘1’. So the shape of the distribution is the semantics of the mechanism in its state. If the AND gate is OFF (‘0’) it will specify a different distribution over BC. That means that the cause-effect structure of the system with the AND gate ON will be different from the cause-effect structure of the AND gate OFF. Of course this is a very impoverished notion of semantics and it remains to be shown that in a sufficiently complex system, ‘interesting’ semantics can be constructed from compositions of these probability distributions (cause-effect repertoires). What I’m arguing is that the composition of all cause-effect repertoires is the only kind of information that is available to the system itself, so if there is intrinsic meaning, it must come from this intrinsic information (the system’s cause-effect structure). It can’t come from anywhere else (like a correlation with the external world). Certainly, though, my essay does not give a satisfying answer as to where and how exactly the semantics can be found in the cause-effect structure.

Intrinsic goals:

You wrote:

“It is pertinent to note, all

physical entities, as they evolve, they follow the path of 'least

action', (usual units Energy x Time). So, if we ascribe 'goal of

performing least action, all mechanical entities naturally do that.”

Precisely. Something like “goal of least action” would not be an intrinsic goal. It’s like reaching perfect fitness in the task. The animat may do that without it possibly being a meaningful concept for the animat itself. Both the feedforward and the integrated animat evolved to perfect fitness. Yet, the feedforward one cannot even be seen as an intrinsic agent, a causally autonomous entity, in the first place.

“knowing” vs. “being”

You wrote:

“… Therefore, a state of an entity is not a correlating information in itself. …”

This is crucial and is what I intended to express in the last paragraph of section III. A system, from the intrinsic perspective, does not “have” information about itself, it does not “know” about itself. Instead it specifies information by being what it is, at the current moment: the system (with all its mechanisms) in a specific state. And this is all there is.

From the essay: “The cause-effect structure of a causally autonomous entity describes what it means to be that entity from its own intrinsic perspective.”

You wrote:

=> There is a big leap here. It is unclear 'what it means to be that

entity'. Though it may be so, but why and how it is so, is not worked

out in this essay. Considering all an element correlates with is the

states of other connected elements, therefore, the constrained element

may have state information about other elements. Furthermore, while

the process of unification is outlined by stating each element affects

every other in turn, but how does the semantic unification takes place

is left out.

I completely agree. My argument is merely that a) something like intrinsic meaning obviously exists for us humans, b) the only way it could possibly arise (in a non-dualist framework) is from the system’s cause-effect structure. And I hope that I have given convincing arguments that the only intrinsic information there is, is the system’s cause-effect structure, while something like correlations with the world can only be meaningful for an external observer.

References:

The word limit unfortunately didn’t allow for more detail about the evolutionary algorithm. However, in some sense this is irrelevant for the rest of the argument. That the animats in Fig. 2 are two solution that actually evolved via selection and adaptation merely makes it more convincing as a model of natural systems with seeming agency. While I’m very happy about your interest in animat evolution, references to the actual scientific papers should rather be seen as proof that I didn’t just make things up, but are not crucial for the argument made. Same for the R = 0.6 value. The only relevant point is that the R that was measured is too low to possible allow for an intrinsic representation of all task-relevant features (less than 1 bit) even in animats that perform the task well.

Thank you again for your insightful comments!

Best,

Larissa

view post as summary