Learning Dynamic Models

[GENERAL] Summarize your solution to the problem.

Photrek will upgrade the Coupled VAE (CVAE) algorithm it is developing for the SingularityNETs (SNET) AI Marketplace to incorporate a dynamic model. CVAE is a risk-aware learning method.

[GENERAL] Summarize your relevant experience.

Photrek is an engineering services firm specializing in developing Machine Intelligence for Complex Systems. Photrek’s current projects include designing low-cost environmental sensors, leading peer-reviewed publications, developing risk-aware algorithms, and DAO governance.

[IMPACT] Please describe your proposed solution.

Photrek proposes to host a Dynamic Coupled Variational Autoencoder (DC-VAE) algorithm on the SingularityNet Marketplace [1]. A Variational Autoencoder [2,3] merges the capabilities of deep learning and probabilistic programming. The algorithm provides a method to learn sophisticated models while retaining enough simplicity to enable interpretable manipulation of the model. Photrek invented a risk-aware process called Nonlinear Statistical Coupling [4] which it has applied to the VAE algorithm. As reported in the journal Entropy [5], this innovation significantly improves the robustness of the model and the data generated using the model. Photrek is proposing to integrate its newest design, incorporating dynamic models, onto the SNET marketplace.

This project has a substantial pedigree, originating from 20 years of foundational research on the application of complex systems theory to decision-making by Dr. Nelson and Dr. Thistleton [6]. That research includes a variety of collaborations, including the most recent effort to develop an open-source community called Machine Intelligence for Complex Systems. 

Three funded projects have provided the resources to enable a successful development and integration of the DC-VAE algorithm. The Catalyst Fund 7 project “Forecasting Cardano Native Tokens” [7] provided the opportunity to demonstrate Photrek’s risk-aware algorithms to the SNET community. From those conversations, two SNET Deep Fund 1 projects were awarded. Photrek is currently working on hosting a) a risk-aware assessment of machine learning algorithms [8] that output forecasted probabilities and b) a risk-aware data generator [9] that uses the CVAE to ensure the generation of data with robust properties. The third phase of the project, which we are requesting Catalyst funds to complete, is to host a Dynamic CVAE. The DC-VAE algorithm will enable the generation of time-series simulations that can control for the degree of risk awareness desired by the user. 

The CVAE algorithm [10], which Photrek is currently preparing for the SNET marketplace, was developed by the open-source community Machine Intelligence for Complex Systems (MICS). CVAE utilizes the Google TensorFlow library [11]. The risk-aware innovations are based on a Nonlinear Statistical Coupling python library [12] invented by Dr. Nelson and developed by MICS. The design work for the DC-VAE is funded by a SingularityNet grant. This Catalyst grant will fund the development and integration of the algorithm. As part of the effort, Photrek will explore the possibility of migrating the architecture to the PyTorch [13] and NeuroProphet [14] libraries.

[1] SingularityNet Marketplace, https://beta.singularitynet.io/.

[2] D.P. Kingma, M. Welling, Auto-Encoding Variational Bayes, in: Int. Conf. Learn. Represent. (ICLR), ArXiv1312.6114v10, 2014: https://arxiv.org/pdf/1312.6114.pdf .

[3] D.P. Kingma, M. Welling, An introduction to variational autoencoders, Found. Trends Mach. Learn. 12 (2019) 307–392. https://doi.org/10.1561/2200000056.

[4] K.P. Nelson, S. Umarov, Nonlinear statistical coupling, Phys. A Stat. Mech. Its Appl. 389 (2010) 2157–2163. https://doi.org/10.1016/J.PHYSA.2010.01.044.

[5] S. Cao, J. Li, K.P. Nelson, M.A. Kon, Coupled VAE: Improved Accuracy and Robustness of a Variational Autoencoder, Entropy 2022, Vol. 24, Page 423. 24 (2022) 423. https://doi.org/10.3390/E24030423.

[6] W.J. Thistleton, J.A. Marsh, K. Nelson, C. Tsallis, Generalized Box-Müller method for generating q-Gaussian random deviates, IEEE Trans. Inf. Theory. 53 (2007) 4805–4809, https://ieeexplore.ieee.org/abstract/document/4385787.

[7] B. Zubillaga and K. Nelson, “Forecasting Cardano Native Tokens”, Cardano Catalyst Proposal, https://cardano.ideascale.com/c/idea/383178.

[8] K. Nelson and W. Thistleton, SingularityNet DF1 Risk-Aware Assessments for AI Applications, SingularityNet Deep Fund Proposal, https://proposals.deepfunding.ai/graduated/accepted/7f31c306-2605-4f73-9a49-8e2793a75eec

[9] K. Nelson and W. Thistleton, SingularityNet DF 1 Risk-Aware Data Generator for SingularityNet Applications, SingularityNet Deep Fund Proposal, https://proposals.deepfunding.ai/graduated/accepted/5860fb2a-57c5-4230-9638-9284299e12db

[10] K. Chen, J. Clements, D. Svoboda, X.Y. Hong, C. Wloka, W. Thistleton, K. Nelson, Photrek/Coupled-VAE, (2021). https://github.com/Photrek/Coupled-VAE.

[11] TensorFlow, https://www.tensorflow.org/.

[12] J. Clements, D. Svoboda, X.Y. Hong, W. Thistleton, K. Nelson, Photrek/Nonlinear-Statistical-Coupling, (2021), https://github.com/Photrek/Nonlinear-Statistical-Coupling.

[13] PyTorch, https://pytorch.org/.

[14] NeuralProphet, https://neuralprophet.com/html/index.html.

[IMPACT] Please describe how your proposed solution will address the Challenge that you have submitted it in.

As the SingularityNet Marketplace migrates to Cardano, it's important to have foundational Machine Intelligence applications that will attract a high volume of API calls. Each API call will stimulate AGIX transactions. While the marketplace is currently on the ERC20 network, the AGIX coin is now a Cardano native token and the marketplace is being transitioned to the Cardano network. Photrek is seeking to establish a fundamental capability for a variety of AI applications. Deep Learning has contributed to world-record performance in classification and generation tasks, while Probabilistic Programming has established rigorous methods for constructing interpretable models. The combination, Deep Probabilistic Programming, is a foundational method to learn interpretable models. Photrek adds to this approach the ability to incorporate risk-awareness into the training of these models. 

[IMPACT] What are the main risks that could prevent you from delivering the project successfully and please explain how you will mitigate each risk?

The main risks for the Learning Dynamic Models project are: 

a) identifying potential customers for the algorithm 

b) determining an effective design for the dynamic model

c) successfully completing the Python development

d) successfully integrating the algorithm into the SNET marketplace

Customer Development Risk - Photrek is engaged with three communities to mitigate the risk of customer engagement. Within the Catalyst community, Photrek has identified organizations such as Drip Dropz and Liqwid that are keenly interested in analytical forecasting for the purpose of mitigating decision risks. Within the SNET community, Photrek has had lengthy discussions with three potential customers. These include SNET itself which is engaged in developing models of environmental sustainability; Rejuve, a spin-off developing models of longevity based on health monitoring; and SingularityDAO, which is focused on algorithmic financial trading. Finally, Photrek has had extensive conversations with financial institutions impacted by extreme weather events. These include Insurance brokers, commodities traders, and energy suppliers.

Design Risk - Photrek is mitigating the algorithmic design risk in two ways. 1) Its SNET DF1 contract includes a component to complete the DC-VAE design, so we anticipate having extensive analysis of the design completed, and 2) Photrek coordinates an open-source collaboration called Machine Intelligence of Complex Systems. This community includes academic scientists investigating the best approaches to learning dynamic models.

Code Development Risk - Like the design risk, Photrek will manage the code development risk in two manners. Photrek employs an exceptional team of Python developers who have contributed to a variety of machine learning projects. This team will be building from the successful architecture of the CVAE using TensorFlow, which outside investigators have already begun testing [1]. Furthermore, Photrek continues to develop alternative approaches. For instance, on its Catalyst Fund 7 project, we are investigating the use of NueroProphet and the PyTorch platforms.

Integration Risk - Photrek plans to mitigate the risk of successfully integrating the algorithm into the SingularityNet marketplace prior to the start of this project. Photrek has already met with the SNET marketplace engineering team. Photrek has funding to integrate two other algorithms onto the marketplace which will give the team the experience needed to integrate the DC-VAE algorithm.

[1] F. Wu, Group Fairness for Learning Representation on Coupled Variational Autoencoder, https://windhaunting.github.io/assets/docs/COMPSCI689_project_FubaoWu.pdf.

[FEASIBILITY] Please provide a detailed plan, including timeline and key milestones for delivering your proposal.

[FEASIBILITY] Please provide a detailed budget breakdown.

The budget is based on the principals Nelson and Thistleton working with a data scientist and web developer. The cloud computing services necessary for customers to process calls on the SingularityNet marketplace is estimated as 25% of the project costs.

[FEASIBILITY] Please provide details of the people who will work on the project.

Project Roles:

• Principal Investigator
• Customer Discovery
• Dynamic Model Design

Nelson is an innovative leader in the research, development, and deployment of systems for complex decision-making. Proven record of creative research, team building, and customer-focused development spanning cyber-security algorithms, multi-sensor systems, machine intelligence algorithms, and decentralized governance. Co-inventor of the Coupled Variational Autoencoder designed to improve the learning of robust, accurate models. Served on Cardano Catalyst Circle, a problem-sensing team guiding the development of one of the world’s largest DAOs.

https://www.linkedin.com/in/kenric-nelson-ph-d-7495b77

William Thistleton

Assoc. Professor

SUNY Polytechnic Institute

Project Roles:

• Co-Principal Investigator
• Dynamic Model Scientist
• DC-VAE Developer

Associate Professor of Mathematics teaching classes in analysis, probability, statistics, design of experiments, and data science. Returned Peace Corps Volunteer. SUNY Online Teaching Ambassador. Research and consulting in Machine Learning, Quantum Annealing, and Data Science. Develops and delivers workshops for teachers, employees, and students.

https://www.linkedin.com/in/william-thistleton-874b8010

The project will employ a data scientist and a web developer. Community members with experience developing machine learning algorithms and/or integrating applications into web marketplace services will be invited to apply.

[FEASIBILITY] If you are funded, will you return to Catalyst in a later round for further funding? Please explain why / why not.

Photrek is seeking to develop a commercial service of innovative machine learning algorithms on the SingularityNet Marketplace. We anticipate 2-3 more Catalyst rounds focused on SNET machine learning development to have a fully functioning service that will be self-sustaining.

[FEASIBILITY] Are you or any member of your team working on any other proposals in this Fund9?

[FEASIBILITY] Are you or your team working on any other proposals from previous Funds?

[AUDITABILITY] Please describe what you will measure to track your project's progress, and how will you measure these?

Customer engagement will be measured by: 

a) the number of conversations regarding customer requirements

b) the number of customer-recommended applications

c) the number of customers using the application

Development progress will be measured by:

a) the number of dynamic models evaluated 

b) the applicability of the model to the analysis of cryptocurrency markets

Integration progress will be measured by: 

a) the functioning of the algorithm on the SingularityNet marketplace 

b) the number of customer API calls

[AUDITABILITY] What does success for this project look like?

Successful completion of the project will include: 

a) development of novel insights from potential customers regarding the use cases for learning dynamical models; 

b) designing a dynamical model that cryptocurrency analysts can use to generate datasets and predict trends; and 

c) initial set of customers who provide a trial exploration using the SingularityNet marketplace.

[AUDITABILITY] Please provide information on whether this proposal is a continuation of a previously funded project in Catalyst or an entirely new one.

The Learning Dynamic Models project will build from the following Catalyst and SingularityNet proposals.

Catalyst F7 Forecasting Cardano Native Tokens, which is part of the AI & SingularityNet challenge. This project will be completed in July 2022. The project was used to introduce the concepts of risk-aware machine learning to the SingularityNet community.  

SingularityNet DF1 Risk-Aware Assessments for AI Applications - This project will develop a SingularityNet marketplace application that provides an assessment of risk impacts on the performance of a machine learning algorithm’s probabilistic forecasts.

SingularityNet DF 1 Risk-Aware Data Generator for SingularityNet Applications - This project will develop a SingularityNet marketplace application that generates static datasets such as images that are robust against outlier risks.

Sustainable Development Goals (SDG) Rating

8 - Promote sustained, inclusive, and sustainable economic growth, full and productive employment, and decent work for all

9 - Build resilient infrastructure, promote inclusive and sustainable industrialization, and foster innovation