Rejuve.AI, the world’s first artificial intelligence fueled Longevity Research network on blockchain, is dedicated to using cutting edge AI in the service of decentralized science. We have our own Generative AI model, similar to ChatGPT, GPT-4, and DALL-E, but dedicated to creating “Deep Reals” of science rather than deep fakes or other “hallucinations” as Generative AI is commonly used for now. We focus on decentralized assembling of different Generative AI models created by Rejuve network members, so that network members own the end product rather than Big Tech. Our tokenomics ensures that these AI modelers and data scientists are compensated for their efforts with each sale of a product made possible by their efforts. Starting in Q2, Rejuve’s next task will be to expand the framework that provides insights in our longevity app. The framework is known as the “Generative Cooperative Network,” and it is a framework for crowdsourcing the scientific community with the goal of creating a Systems Biology model: a multiresolutional simulation of the human body.
By “multiresolutional,” we mean that this simulation will include multiple levels, from the most basic, such as genetic expression, to the most complex, such as chronic aging conditions. This framework will crowdsource, cull, and validate generative models from the scientific community, including Generative Neural Network models, Bayes Expert models, and simulation models. It will use them to find personalized longevity treatments using app user data, data from our methuselah flies, and other open source datasets. We have already developed the simulation code for the GCN, known as the “Singularity net simulation,” and have applied it to generative natural language and image neural models; all that remains is to apply it to longevity.
We chose to focus on ensembles of generative neural network models because they are the most cutting-edge application of Generative AI in science. The ability to recombine ideas into new ideas is one of the foundations of human cognition, according to Smolensky, a neural net pioneer and linguist. Smolensky believes that the trend in deep learning towards neurocompositionality is driving real progress in AI. (1) This is especially true of transformers, which are used in Google’s Linguistic Mixtures of Experts (LIMoE). (2) The most important applications of AI in computational biology are also due to transformer neurocomposition: in OpenAI’s seminal protein folding application Alpha Fold 2 (3), or in its message passing neural cousin Halicin (4), which created the first antibiotic treatment by an AI. Diffusion transformers have recently been combined with bottom-up simulations to design proteins. (5)
We take this approach also because we believe that the answers to longevity can be found in complex biological networks that operate not only at the level of individual relationships between biological phenomena, but also at the patterns that these relationships form when they all act together. Because no single entity can be an expert in all of those studies and ideas about biological relationships, we seek the assistance of the scientific community to correctly model those ideas so that our framework can reassemble them all in ways that ensure the model and data agree with each other and are accurate.
Perhaps some of you recognize our name “Generative Cooperative Network” or GCN from the popular Neural Network known as the “Generative Adversarial Network” or GAN. Our GCN has multiple agents, each with an AI that can be a neural network, similar to the GAN, which has two neural network-equipped agents. Our GCN generates a model, in this case a multiresolutional simulation of the human body, through agent competition and cooperation. Similarly, the GAN uses competition to generate data that is indistinguishable from a real life image, such as a picture of an imaginary celebrity. To generate these things, both the GCN and the GAN use coevolution, or machine learning algorithms that learn from one another.
However, our algorithm employs both cooperation and competition, bringing together models from the scientific community. Both the GCN and the GAN employ scaffolding, which allows one agent to instruct another on how to reach a solution through feedback. Our GCN, on the other hand, does so via a social coordination algorithm in which agents form institutions that guide other agents, whereas the GAN relies on intensive feedback from a single critic.
Our GCN, like our Rejuve tokenomics design, simulates a market for problem solving. In our tokenomics design, longevity challenges are presented to the scientific community, and those who are part of winning teams receive tokens. Individual scientists and health data contributors, on the other hand, do not have to solve the entire problem to win (though they can if they want to). Rather, our GCN will combine their models and contributed data to create a model capable of solving a problem.
It solves the problem by creating a mini, simulated market in which agents combine scientist-contributed models and app user data by forming teams and solving a variety of longevity-related problems at multiple levels of difficulty, including automatically generated problems like masking and predicting internal variables, or comparing the patterns produced by one level of simulation to real-world data. The real market differs from the simulated market in that there are many more challenges, which are applied and won thousands of times until the agents adapt to each other and fall into roles that are general and reflective of biological concepts for the multiresolutional simulation.
Agents form their teams using two main algorithms: one that uses signs and institutions in a “intuitive” way, and the other that uses logic from our premier neuro-symbolic AI, Hyperon. With more model and data contributions over many iterations and over time, a multiresolutional simulation is created.
This multiresolutional simulation can assist in determining what causes some individuals of a species to live longer than others. In the GCN, we plan to use a variety of data sources, including app users, supercentenarians, and methuselah flies. Say the key to living a long life is to protect the telomere ends of the chromosomes, the second “hallmark of aging.” By involving those concepts in the accurate solutions to many challenges, the contributed models and data would help to produce accurate biological concepts around this hallmark.
One challenge could be to find a mechanistic explanation for the genetic expression data involving fly biological networks and telomere ends that, when reused, would fit other data such as supercentenarian longevity data. The targets identified in the dynamic biological network model could then be compared to the results of supplementation taken by app users. The data from flies is “flat,” but the simulation of biological networks behind it, made up of models contributed by scientists, would explain and bring it to life.
With each challenge completed, the group of models assembled to solve a problem is validated or culled. As this growing group of crowdsourced models improves at a variety of tasks related to longevity, it will become a clearer causal and mechanistic representation of the human body’s processes, as well as become better at creating longevity treatment hypotheses. For more information on our Generative Cooperative Network, please see our whitepaper.
References
Smolensky, P. et al. (2022). Neurocompositional computing: From the Central Paradox of Cognition to a new generation of AI systems. Available at: arXiv:2205.01128v1 [cs.AI].
Mustafa, B. and Riquelme, C. (2022). LIMoE: Learning Multiple Modalities with One Sparse Mixture-of-Experts Model.
Skolnick, J. et al (2021) “AlphaFold 2: Why It Works and Its Implications for Understanding the Relationships of Protein Sequence, Structure, and Function”. J Chem Inf Model. vol. 61, no. 10, pp. 4827–4831.
Booq R. et al. (2021) “Assessment of the Antibacterial Efficacy of Halicin against Pathogenic Bacteria”. Antibiotics, Vol. 10 no. 12, pp 1480. Available at: https://pubmed.ncbi.nlm.nih.gov/34943692/
Baker Lab. A diffusion model for protein design., https://www.bakerlab.org/2022/11/30/diffusion-model-for-protein-design/