Hi! My name is Debbie Duong and I am very happy to work for Singularity Net! I am a computational social scientist and complex adaptive systems researcher. I wrote the first adaptive agent simulation in 1991 - a multi agent simulation in which each agent had an IAC neural net mind. Later my agents had other minds such as Evolutionary Computation, Boltzmann Machines, and Bayesian Optimization Algorithms. These are all social simulations that study Micro Macro Integration and the development of social institutions. These social institutions tie right into blockchain consensus, DAO governance, the new institutional economics, and the governance of the commons. I see no distinction between self organization, simulation, social science, AGI and governance enabled by blockchain. I think that simulation and social science are important missing links in AI complexification. Besides these first priorities I am also interested in teasing out cause in observational studies with our machine learning, making Bayesian networks and machine learning actually causal and thus actually semantic with the help of simulation. The sciences of Econometrics and Epidemiology have a lot to say about teasing cause from observational studies that Reinforcement Learning Scientists need to know for off-policy learning. I want to bring AI to real world problems, and have done so for the US government and healthcare system before coming to Singularity Net. I have written simulations of how corruption tears at social institutions in hybrid warfare, of the political compromises in Bueno De Mesquita’s theory of dictatorship, of the effects of Strategic Information Warfare on populations, and of ACA programs that incentivize healthcare stakeholders towards patient health. I want to inspire the next generation of programmers, and especially women, to go beyond playing video games with AI, towards playing healthcare distribution, best treatment path, and defense against fascism. To take seriously their responsibility to solve social justice issues. My doctorate is from GMU and influenced by their center for market processes but I am not an Austrian economist. For more info see my old school webpage at http://www.scs.gmu.edu/~dduong/ Fav Team : Go Deep State!!
Welcome Deborah! So glad to have you here!
CAS theory is something I studied with the Singaporean CAS Scholar, Dr Liang and I deeply enjoyed the introduction to it and the models faciliated a lot of my work later with the companies i worked at (https://www.worldscientific.com/worldscibooks/10.1142/7326)
It’s some tough heavy hitting stuff to wrap my mind around
I did have a question- could you expand on this so I can :
"Besides these first priorities I am also interested in teasing out cause in observational studies with our machine learning, making Bayesian networks and machine learning actually causal and thus actually semantic with the help of simulation. "
Also if you have any helpful beginner resources, it would be great to put that on my reading list!
Hey Thanks! As for “making machine learning actually causal” its an interest of mine, and I have done a few things in that direction. For example, I have a natural language program called Indra that finds roles and role relations in text, for the purpose of feeding simulations. The intent of the project is not just data driven simulation, but also natural language understanding. As for Bayesian Networks, you pretty much have to give them the causal relations before they can do their causal inference, but they can not find them themselves in observational data. They do not abduct a causal model. I think knowledge of models is necessary to making perception semantic and to turning natural language processing into natural language understanding. I think we need an abduction program like Bacon, but a smart version of it, that finds the features, and makes the model. Could we make a neural network look at, say, a solar system, and then look at the data from the atom, to abduct the Bohr atom as Bohr did in his dream of the solar system? Could a neural network understand that both videos are of the same underlying process? Not the supervised learning version - that’s easy - but the abduction version. If it could , I think we would have semantic perception. The underlying model could be a complex model of human intention, or a simple model of physics like gravity, but the point is to understand either a visual scene or language in terms of, given the context of possible and likely models, what would be happening.
Teasing out cause from observational data is one thing that machine learning and reinforcement learning need to deal with in order to solve real world problems at all. These are subjects that epidemiology and econometrics address, with tools that make sure distributions are in common when looking at the effect of a treatment, that you are looking at the same distributions (as in the potential outcome framework) or if the distributions are different, they are randomized (as in instrumental variable techniques). If off-policy reinforcement learning is to learn the effects of actions in the real world, then it will have to tease out ambiguities like, was the poor result in the diabetes patients who take insulin due to their poverty - insulin is cheap and so are twinkies - or due to the insulin itself? Narrow RL doesnt do that now, and can learn from the social sciences. This is necessary if we are going to play “healthcare” instead of Atari with RL algorithms.
As for beginner’s resources, I think one place to look is the work of David Sontag, who has some neural models with causal regularization.
Hi Deborah! A warm welcome to this forum, I am very proud to have a woman like you as my colleague . I would really love to see the simulations you created to be discussed more here on the forum and to have more people see your wonderful work. I would like to encourage you to share it in Share your Project as I think it will inspire many to see what is possible .
Please also let me know if you feel like the categories we currently have fit the expertise and topics you are able to bring to the forum. Throwing out a quick idea but perhaps Decentralized AI Economy could be reshaped still to Decentralized Societies instead - which would maybe allow for more social science discussions too as well as the economy discussion. Please do share your thoughts . I could create a poll later to evaluate the categories with everyone in the community and perhaps change a thing or two. We are still shaping it and I would be really interested to especially read more discussions about complex adaptive systems in healthcare and warfare as well.