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>.
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.
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.