RE: teleological concerns:

Let me word my point differently: Evolution just does not care about what is optimal or best, what is fittest, what survives, what improves the chances of survival, what is preserved, or what solves which problems. Those involve anthropomorphizations of evolution via values, interests, and ideas of means-to-ends relationships (whether as end-product or byproduct). Sure, there are causal tendencies for the likelihood of certain outcomes. But a large meteor could nip any of those in the bud, no matter how optimal or problem-solving we think they might have been.

Yes, nothing particularly exciting. It is just something you might expect AI to be able to do, but it would not be emblematic or definitive of the AI's intelligence (it wouldn't be AI qua AI). Perhaps it should've been billed as an essential tool in an AI's toolbox?

Dear Thomas Neusius

In contrast to the view you present to us, true AGI (General Artificial Intelligence) people have to find all the reliable and important KEY species-typical behavior patterns and do as they do "mechanically", but appropriately, as specifically triggered in types of circumstances (by aspect of the environments) AND in different ways at different points in the ontogeny of the system [(this including developing true and hierarchical (hierarchically-related) patterns of responding/learning over time -- in some sort of NECESSARY sequence, with likely some contingent "trimming down" of earlier behavior patterns)]. This is not remotely just optimization, but you can see how it could surely involve some of that; but this is a "far cry" from just optimizing some "systems" of more questionable origin, which is what some of the people you speak of may be doing, using their own hypothetico-deductive systems, somehow reasoned out, from some "things" (perhaps, from poorly founded models and maybe, even worse, by being problematically compounded by being "by analogy").

Artificial Intelligence is an algorithm formulated by a human that tries to mimic human thinking and produces information and insights automatically for decision making.

The first type of artificial intelligence was called expert systems (+1950s) where businesses will question senior manages on how they make decisions so that when the senior manager dies or leaves the business, the business will still be productive and sustainable. If-then algorithms including the action that needs to be taken is coded in an algorithm with alerts in the system when an urgent task needs to be attended to.

Today, we have machine learning which is a subset of artificial intelligence and deep learning which is a subset of machine learning. Machine learning consists of unsupervised learning (cluster analysis, dimension reduction, etc), supervised learning (logistic regression modelling, decision tree modelling, neural networks, etc) and reinforcement learning.Machine Learning models solve classification or regression problems.

Deep learning problems usually have vast more data than machine learning models and usually solved using neural networks (CNN) with many more hidden layers than machine learning problems. Typical deep learning problems are natural language processing problems, audio to text problems, face/digit recognition, picture classification, etc

In a nutshell, artificial intelligence is a technology that integrates data, processes, and machine learning techniques for real-time decision making so that the business can be proactive and make decisions faster, smarter and cheaper, giving the business a competitive advantage over its competitors.

Artificial intelligence is innovatively changing the way we do business, focusing on the customer experience and real-time information for strategic decision making and competitive advantage.

Dear Aleš Kralj ,

thank you for your comment. I agree with you and Joachim Pimiskern

That may also be the reason why I felt, thinking of AI just as another optimization is missing central features of AI or of the kind of problems, to which AI is applied. E.g., in optimization problems mathematicians turn their attention to some of the following questions:

• Is there a global optimum?
• Are there local optima?
• Is the algorithm able to identify the global optimum?
• How quickly converges the algorithm?

I'm unsecure, that this perspective focuses on the most relevant aspects.

Best,

Thomas

Hi All,

I agree with the optimisation idea, algorithms are developed to bring something about and are altered in terms of making it quicker, easier, less expensive to reach this specified 'something'

Of course there is a thought to be spent on the human, biologically implied feature of aspired self-reproduction, or shall we say - more generally - reproduction (as it is not me that I reproduce but another, individual human being).

However - and that seems to be an important point: This self-reproductive feature only gains important in the context of hardware solutions (human bodies) that come with a limited life-time. Away from that, the ideas (how to develop an algorithm or how to solve an equation) run on these biological systems are not really limited to these systems but passed on in what is known as education.

So there seems to be some system-development in humans which is not strictly linked to a specific human life-time?

But there is another aspect. Humans are educated, but so are - at least some - AI systems who have to develop their algorithms in various test-runs to adopt their actual performance to the task at hand. These AI systems are thus trained by humans, or even by other AI-systems (which again were trained by humans).

So even when we are talking about opimising processes in relation to human-set tasks and human observed acomplishments, we do not seem to have the ability to 'step out' of the human-shadow in which all these projects are - invariably - cloaked.

best

tom

Mathematical Modelling , simulation and Optimization and deep learning covering

most of the portion of artificial intelligence . ANN is also a part .

All this talk of evolutionary optimization is too teleological for my taste. Evolution is not goal-directed.

Dear Karl Pfeifer,

I do share your teleological concerns. However the optimisation, supposedly driven by evolution is not the endgoal of the process, but rather the byproduct of the weeding out of less then optimal solutions. To avoid the notion of an 'outside' goal being grafted upon evolutionary processes the term telenomy is nowadays used. I think that should somehow address your valid concerns?

Dear Karl Pfeifer , dear Tom Feldges ,

of course, biological evolution is not goal directed, but everything which under the present circumstances improves the chances of self-reproduction is strengthened. This is not just maximizing the number of offspring.

As said, I consider self-reproduction as central feature of evolutionary processes. Tom Feldges pointed out, this is self-reproduction on the level of species. It leads to an implicit objective in evolutionary systems: adapt to the environment as good as possible to secure successful reproduction.

However, the analogy of evolution and AI is far from being perfect.

I think AI is not reduced to an optimization problem. optimum of neural network is Local but for SVM for example the optimum is global for binary classification. On the other hand, The unsupervised Learning doesn't require any inputs. The machine learn and classify alone. The challenge of AI algorithmes i think is to underfitt models rather than find optimums.

RE: teleological concerns:

Let me word my point differently: Evolution just does not care about what is optimal or best, what is fittest, what survives, what improves the chances of survival, what is preserved, or what solves which problems. Those involve anthropomorphizations of evolution via values, interests, and ideas of means-to-ends relationships (whether as end-product or byproduct). Sure, there are causal tendencies for the likelihood of certain outcomes. But a large meteor could nip any of those in the bud, no matter how optimal or problem-solving we think they might have been.

Dear Karl Pfeifer ,

catastrophic events in the surroundings may extinguish the whole process - that applies to the biological evolution as well as to me running to catch a train. While evolution is not goal-oriented, me running certainly is. One may say, the meteor doesn't care.

Evolution as a process does not care, I agree. But that's a remark on precise phrasing rather than on the optimization involved. However, the process of evolution favors certain outcomes in the sense of fitness. Clearly, fitness is a vague term. Under constant environmental conditions, low variability may be beneficial ; under quickly changing conditions, once advantageous features may turn out to be detrimental. And abrupt changes like the meteor may even wipe out everything. But that is not in conflict with the observation, that evolutionary dynamics exhibit astonishing solutions we may try to imitate - eventually, human intelligence being among them.

Best,

Thomas Neusius

Dear Karl,

I am not sure I want to buy into your verdict that optimising evolution can not apply because a meteor may ruin it. That does not sound like convincing reasoning to me? I do agree that there is a certain tendency to anthropomorphise the whole evolutionary story and I do agree that if we talk about value-concepts like fitness, it is always important to ask who defines the parameters of what is to be classed as fit or unfit. Economics has tried to define 'good', i.e. a massively value-soaked concept by stating that 'good' means 'fit for purpose'. If you then - even in changing environments ascribe 'goodness' to some evolutionary developments than we're actually no longer taking about an anthropocentric view, but on which is merely orientated on the 'fit' of the radiative adaptation into or along an ecological niche. Not much Darwin here any more, and rather un-anthropocentric? In that respect the meteor could be forgotten as well, as the changes brought about by its impact would merely alter the ecological requirements.

Best

Tom

Tom Feldges Evolution is the history of the succession of the pro tem survivors, and you can explain why they survived because of whatever genetic properties they happened to have that interacted with their environment. Why call that "fit for purpose", which already brings in a teleological notion? Talking about optimization is an after-the-fact rationalization. Evolution doesn't pursue strategies for optimization or employ algorithms (or whatnot), although it can be described as if it did (i.e. we can map algorithms onto the causal processes). The microbes and small animals that survive a large-scale meteoric collision aren't solving a problem in a more optimal way than the larger animals reduced to extinction levels. This is not to deny that their survival might teach us important lessons that we might make use of for some of our purposes and problems, which may include making ourselves more "fit for [some] purpose".

Karl, that clarifies it for me

best Tom

Ahoy Aleš Kralj .

RE: This would mean that the fate of life on Earth is sealed if we fail to become interstellar life within this time-frame.

That is certainly a possibility that I and many others would want to avoid; I'd make avoidance my goal if I had the means. (Of course not everyone wants that, e.g. the Voluntary Human Extinction Movement and David Benatar, who wrote Better Never to Have Been). I even agree that all right-minded people should see the 2nd law as a challenge that can give humanity a grand purpose (e.g. delaying the inevitable) in the hierarchy of purposes. So evolution has evolved some purposeful creatures. But the purposes are theirs, and not the purposes of evolution or of some intelligent designer. Evolution and the 2nd law provide the context for their purposeful pursuits, but evolution just chugs along in its uncaring way and would continue to do so were purposeful creatures to became extinct.

It seems as if we somehow get away from the initial topic of the question?

A few remarks, probably quite in line Ales’ comments, but with a different theoretical framework:

Living organisms are distinct from unanimated matter by what Varela and Maturana called an autopoiesis driven ‘urge’ to preserve their own existence in an inherently hostile environment (that's were Ales’ thermodynamics come in).

The world appears to be ordered in a certain way, whereby order does not imply or necessitate anyone doing the ordering, but it is a brute fact of life that humans cannot survive without oxygen and that certain objects drift on waters while others sink. I.e. there are cetain physical laws by which everything that is exposed to an environment governed by these laws adheres.

Evolution as a process allows for the already mentioned adaptive radiation of living organisms to fit in to ecological niches and thus maintain their existence according to their inherent autopoietic nature. It is here – and on this level – that the notion of ‘good-ness’ manifests itself. I.e. what kind of result of an otherwise un-interested evolutionary process is good for the individual life-form which either thrives or ceases to exist. Hence – we must be careful of not to conflate the levels of description here.

This brings me back to the actual question, which is a little bit smaller than our possible future as interstellar beings. Computer programs are designed to produce intelligent outcomes they do not (yet?) have this inherent autopoiesis which would provide them with in inner drive to secure their self-maintenance – they could not care less!

But that seems to imply:

• Evolution is a process beyond good and bad – it just is what it is.
• Human beings rate evolution-imposed results and adaptations as good or bad, hence it is the human who brings a teleological notion into the interpretation of merely rule-goverened, but otherwise coincidental processes.
• Humans program computers and design intelligent software. They do so with no intention of allowing the software to develop an individual point of view about their own existence of performance. They want to make the software achieve specific outcomes.
• In that respect it seems to follow that – at least – one of the main-difference in – let’s call it – intergenerational developmental processes in human beings and an optimization process of computer hard and software is the fact that human development is not goal-orientated as such. It follows the given laws of physics in a non-determinate way. While computer programs are designed by humans to bring about something and in that respect their evolutionary process follows a clear teleological purpose.

That of course would beg the question if computer based intelligence could ever step out of this overarching human shadow which appears to be woven into the creation of the algorithms and programmes (after all, this is all man-made with a specific purpose in mind)?

Hope that all makes sense

Best

Tom

Dear Tom Feldges ,

the topic I proposed initially for discussion was indeed much less philosophical than this thread turned out to be. Ironically, that is another example of unintended evolution. However, I found and still find a lot of interesting ideas and I thank you, Karl Pfeifer , Aleš Kralj and all others who shared their opinion with me and the followers of the discussion!

Best,

Thomas

Thank you Thomas, it’s been a pleasure although it might not quite have been what you asked for ...

Best

tom

The discussion has got a philosophical turn, as Tom Feldges and I pointed out.

However, I try to get back to my initial question. What I miss is the point of the underlying model. Typical optimization strategies involve mathematical models, which convey considerable information about the problem addressed. That has the advantage of a rather small set of parameters to be estimated on the basis of potentially few data points.

As I understood, AI is much more applied to problems, where a plethora of data is given, but little is known about the processes that brought these data about. AI then should help to identify patterns in those data without specifying explicitly a model.

So would you agree, that the models of AI approaches, like multi-layer neural networks, are different from classical statistical modelling in the sense that they are more flexible but also less specific, and that they require much larger samples?