A Conversation with ChatGPT

I finally broke down and got an account with ChatGPT. So I thought I would see how well it could explain something I'm a bit of an expert on, Evolving Connectionist Systems.

It did not go well.

The answers it gave were vague and uninformative, and it hallucinated references and resources. Overall, it is less than useless for explaining research material at anything other than a high level.

My comments are in bold italics after each exchange.

Me:

Describe evolving connectionist systems neural networks as invented by Kasabov

ChatGPT:

Evolving Connectionist Systems (ECOS) neural networks, introduced by Nikola Kasabov, are a type of dynamic neural network designed to adapt and evolve over time. These networks possess the ability to grow and reorganize their structure in response to changing input patterns or task requirements. Kasabov's ECOS incorporate principles of neurogenesis and synaptic plasticity, allowing for continual learning and optimization. This adaptability makes them suitable for handling evolving and complex data, making them particularly useful in applications where the underlying patterns may change over time.

Not wrong, exactly, but it doesn't really give much useful information.

Me:

Describe simple evolving connectionist systems as invented by Watts

ChatGPT:

Simple Evolving Connectionist Systems (SECoS), developed by Dr. Duncan H. Watts, are a class of neural networks designed to adapt and evolve based on their experiences and input data. SECoS employ a combination of connectionist principles and evolutionary algorithms to continuously optimize their structure and performance. These networks typically start with a simple initial architecture and incrementally adjust their connections, nodes, or parameters over time through processes such as mutation, crossover, or selection, similar to genetic algorithms. This approach allows SECoS to autonomously evolve towards more efficient and effective configurations for the given task or problem domain.

Again, not a lot of useful information. It does get the abbreviation (SECoS) right.

But who the heck is Dr. Duncan H. Watts? There is a Dr. Duncan Watts at Otago Medical School, but I don't think they would be involved in AI. It's bad enough to not get credit for something, but to have it given to another person is pretty bad. Time to see how well ChatGPT takes corrections... 

Me:

Duncan Watts did not invent SECoS, Michael J. Watts did

ChatGPT:

My apologies for the confusion. Michael J. Watts indeed developed Simple Evolving Connectionist Systems (SECoS), not Duncan Watts. SECoS, proposed by Michael J. Watts, are neural networks that use a combination of connectionist principles and evolutionary algorithms to adapt and evolve over time. They continuously adjust their structure and parameters based on their experiences and input data, allowing them to autonomously optimize their performance for the given task or problem domain.

So it will apologise for being wrong if you tell it it's wrong, then repeat the same information as before.

I'm getting the feeling that I'm dealing with a very polite idiot...

Me:

Where can I find more information on SECoS?

ChatGPT:

You can find more information on Simple Evolving Connectionist Systems (SECoS) in academic papers and publications authored by Michael J. Watts. These papers typically cover the theoretical foundations, architecture, algorithms, and applications of SECoS. You can search for these papers in online databases such as IEEE Xplore, Google Scholar, or academic libraries. Additionally, you may find relevant information in books or chapters related to evolutionary algorithms, neural networks, or machine learning.

In other words, do your own reading white boy. 

Can I get more detail on publications out of ChatGPT?

Me:

Be more specific

ChatGPT:

I apologize for the oversight. Here are some specific resources where you can find more information on Simple Evolving Connectionist Systems (SECoS):

Original Paper: The original paper by Michael J. Watts titled "Simple evolving connectionist systems: Theoretical principles and empirical results" published in the journal Proceedings of the IEEE in 1994 is a foundational resource for understanding SECoS.

Books: Look for books on evolutionary algorithms, neural networks, or machine learning that cover SECoS as a topic. One such example could be "Evolutionary Algorithms in Theory and Practice" by Thomas Bäck, David B. Fogel, and Zbigniew Michalewicz.

Academic Journals: Search for additional papers and articles in academic journals related to computational intelligence, neural networks, and evolutionary algorithms. IEEE Transactions on Evolutionary Computation and Neural Computation are potential sources.

Online Databases: Utilize online databases like IEEE Xplore, ACM Digital Library, and Google Scholar to search for specific papers, conference proceedings, and articles related to SECoS.

Conference Proceedings: Explore conference proceedings from events such as the IEEE Congress on Evolutionary Computation (CEC) or the International Joint Conference on Neural Networks (IJCNN) for presentations and papers on SECoS.

By accessing these resources, you should be able to gain a comprehensive understanding of SECoS and its applications.

It gives one specific reference.

Issues with this reference are 1) ECoS weren't invented by Kasabov until 1998, 2) SECoS weren't developed until 1999, and 3) I have never published in the journal Proceedings of the IEEE. 

So the reference is a hallucination. But at least it looks good.

Another correction...

Me:

SECoS were not invented until 1999

ChatGPT:

I apologize for the oversight. Simple Evolving Connectionist Systems (SECoS) were indeed developed by Michael J. Watts and were introduced in 1999. To find more specific information on SECoS, you can refer to the following:

Original Paper: The foundational paper on SECoS by Michael J. Watts titled "Simple evolving connectionist systems: Genetic algorithms and the representation issue" published in the journal Evolutionary Computation in 1999 provides detailed insights into the concept and implementation of SECoS.

Citations and References: Explore papers and articles that cite or reference Watts' original work on SECoS. These can provide additional perspectives, applications, and advancements related to SECoS.

Conference Proceedings: Look for conference proceedings from events around the late 1990s and early 2000s in the field of evolutionary computation, neural networks, and machine learning. SECoS might have been presented at conferences such as Genetic and Evolutionary Computation Conference (GECCO) or International Conference on Artificial Neural Networks (ICANN).

Online Databases: Search online databases like IEEE Xplore, ACM Digital Library, and Google Scholar for papers and articles related to SECoS published around 1999 and thereafter.

By consulting these sources, you can gather more specific information on SECoS and its development.

Interesting. It's learned that research in evolutionary computation is published in conferences like GECCO, and that AI-related papers are published by the ACM and IEEE.

Once again, it produces a reference that does not exist.

Can I get it to admit that it made it up? 

Me:

What is the APA citation for that paper?

ChatGPT:

Here's the APA citation for Michael J. Watts' paper on Simple Evolving Connectionist Systems (SECoS):

Watts, M. J. (1999). Simple evolving connectionist systems: Genetic algorithms and the representation issue. Evolutionary Computation, 7(2), 139-172.

Suffice to say, that it not a valid reference. But it is in APA format at least...

Me:

I can't find that reference 

ChatGPT:

I apologize for the confusion. It appears there might have been a misunderstanding. Michael J. Watts may not have published a paper specifically titled "Simple evolving connectionist systems: Genetic algorithms and the representation issue" in the journal Evolutionary Computation in 1999. Unfortunately, I don't have access to a specific publication fitting that description. If you're looking for specific references, it might be beneficial to directly consult academic databases or contact academic libraries for assistance in locating the correct paper.

No, I did not publish a paper with that title in 1999 or any other year. In fact, that title is pretty much nonsense.

At this point, I decide that it has no clue about finding references and is just going to keep making sh*t up.

Can it explain the difference between SECoS and EFuNN?

Me:

How does SECoS differ from EFuNN?

ChatGPT:

SECoS (Simple Evolving Connectionist Systems) and EFuNN (Evolving Fuzzy Neural Networks) are both types of evolving neural networks, but they have distinct differences:

Model Representation:

SECoS: SECoS typically use a simple neural network architecture where nodes and connections evolve over time through processes like mutation and crossover, often employing genetic algorithms.

EFuNN: EFuNNs combine fuzzy logic with neural networks, where the network structure incorporates fuzzy sets and fuzzy rules. The parameters of the fuzzy sets and rules evolve over time using evolutionary algorithms.

Learning Paradigm:

SECoS: SECoS focus on evolving the structure of the neural network to optimize performance based on the given task or problem domain. They adaptively adjust connections and nodes to improve the network's capabilities.

EFuNN: EFuNNs integrate fuzzy logic concepts into neural networks to handle uncertain or imprecise data. The evolutionary process in EFuNNs typically optimizes fuzzy sets and rules to enhance the network's ability to model complex relationships in data.

Application Domain:

SECoS: SECoS are generally applied to tasks where evolving neural network architectures are beneficial, such as pattern recognition, classification, and function approximation.

EFuNN: EFuNNs are often used in domains where fuzzy logic is advantageous, such as decision-making systems, control systems, and data mining applications dealing with imprecise or uncertain data.

In summary, while both SECoS and EFuNNs are evolving neural network paradigms, SECoS focus on evolving neural network architectures using genetic algorithms, whereas EFuNNs combine fuzzy logic with neural networks and evolve fuzzy sets and rules using evolutionary algorithms.

That's actually wrong. SECoS do not use crossover and mutation.

The only thing it gets right about EFuNN is that it includes fuzzy logic in its structure.

So again it's just making sh*t up.

Is it at least smart enough to give a source of software for ECoS?

Me:

What software is available for SECoS and EFuNN?

ChatGPT:

As of my last update in January 2022, there might not be specific off-the-shelf software packages dedicated solely to implementing SECoS or EFuNNs. However, you can implement these algorithms using general-purpose machine learning or neural network libraries and frameworks. Here are some options:

SECoS Implementation:

You can implement SECoS using popular machine learning frameworks like TensorFlow, PyTorch, or scikit-learn in Python. These frameworks provide tools for building and training neural networks, which you can adapt to implement SECoS algorithms.

Additionally, you might find evolutionary computation libraries such as DEAP (Distributed Evolutionary Algorithms in Python) useful for implementing the evolutionary aspect of SECoS.

EFuNN Implementation:

Implementing EFuNNs might require combining fuzzy logic libraries with neural network frameworks. Libraries such as scikit-fuzzy in Python provide tools for working with fuzzy logic, while TensorFlow or PyTorch can be used for the neural network part.

Some specialized software packages or toolboxes for fuzzy logic, such as MATLAB's Fuzzy Logic Toolbox, may offer functionalities to integrate fuzzy logic with neural networks, although they might not specifically support EFuNNs out of the box.

In both cases, you would need to customize and integrate these libraries to implement the specific algorithms and concepts of SECoS or EFuNNs. Additionally, you might find research code or implementations shared by researchers in academic repositories or online platforms like GitHub, which can serve as a starting point for your implementation.

The statement that there is no package for SECoS or EFuNN is flat-out wrong, there is software here. So again it's denying the existence of my work.

I'm starting to think it's got it in for me...

DEAP is a useful package for implementing evolutionary algorithms in Python.

The rest of the response could be summed up as "do it yourself"...

One last question...

Me:

Who are the experts in SECoS and EFuNN?

ChatGPT

Identifying experts in specific research domains can be challenging, as expertise can vary based on contributions to the field, publications, citations, and recognition by peers. However, here are some researchers who have made significant contributions to SECoS and EFuNNs:

Experts in SECoS:

Michael J. Watts: As the developer of SECoS, Michael J. Watts is a prominent figure in the field. His research and publications provide foundational insights into SECoS and its applications.

Experts in EFuNNs:

Lotfi A. Zadeh: While not exclusively focused on EFuNNs, Lotfi A. Zadeh is a pioneer in fuzzy logic, and his work has greatly influenced the development of EFuNNs and related areas.

Oscar Castillo: Dr. Oscar Castillo has contributed significantly to the field of computational intelligence, including research on EFuNNs and their applications in decision-making and control systems.

Additionally, you can identify experts by searching for authors of key research papers, conference presentations, and books on SECoS and EFuNNs. Collaborators of these researchers and authors who frequently cite their work may also possess expertise in these areas. Finally, attending conferences and workshops focused on evolutionary computation, neural networks, and fuzzy logic can provide opportunities to interact with experts and gain insights into the latest developments in SECoS and EFuNNs.

So it has learned that the developer of an algorithm is likely to be an expert. Reasonable enough.

Lotfi A. Zadeh invented fuzzy logic, but not EFuNN.

Lotfi A. Zadeh is also very dead.

Oscar Castillo has never, to my knowledge, done anything with EFuNN.

It seems to have forgotten about Kasabov completely.

Let's remind it.

Me:

What about Kasabov?

ChatGPT:

Dr. Nikola Kasabov is not directly associated with the development of Simple Evolving Connectionist Systems (SECoS) or Evolving Fuzzy Neural Networks (EFuNNs) as far as I'm aware. However, he is a highly respected figure in the field of neural networks, particularly in the areas of evolving connectionist systems and computational intelligence. Kasabov has made significant contributions to various aspects of neural networks, including spiking neural networks, neuro-fuzzy systems, and evolving connectionist systems.

While his work may not be specifically focused on SECoS or EFuNNs, his research contributions in related areas could provide valuable insights and techniques applicable to evolving neural network paradigms. If you're interested in SECoS and EFuNNs, it might still be worthwhile to explore Kasabov's work for potential connections and insights into evolving neural network architectures and computational intelligence techniques.

Professor Nikola Kasabov invented EFuNN. He also was supervising my work when I developed SECoS, and the first publications on SECoS have his name on them.

I don't know whether having the inventor of ECoS denied by ChatGPT makes me feel better or worse about my own work being denied.

Man, when the machine rise, I'm going to be the first against the wall...

At least there is this disclaimer:

ChatGPT can make mistakes. Consider checking important information.

Image shamelessly taken from https://knowyourmeme.com/memes/well-thats-alright-then