Long Description
Company Name
Photrek
Summary
Photrek will develop and onboard Simulating Risky Worlds. This service enables the simulation of virtual worlds with controllable degrees of uncertainty (temperature) and relative risk (variations in temperature). The service builds from two components:
-
Photrek’s
-
Deep Brain’s Open-Source
The preliminary development of the World Models simulation was completed as part of the research component of Photrek’s Deep Fund 1 Risk-Aware Data Generator for SingularityNET Applications project. The technical risks for the project will be managed through a process of a) identifying gaps between our algorithms capabilities and our customers’ value proposition, b) documenting and prioritizing those risks, and c) mitigating risk through coding improvements and writing requirements for future development.
Funding Amount
$49,480
The Problem to be Solved
The rapid development of AI/ML algorithms, hereafter referred to as Machine Intelligence (MI), has resulted in brittle systems. While impressive results can be reported for tests tightly connected to the statistical distribution of the training set, even slight deviations from the training set can result in catastrophically poor performance. Photrek addresses this shortcoming of Deep Learning by incorporating into the training cost functions a tunable control of relative risk aversion. Increasing the cost of outlier performance enables algorithms trained using the Photrek Risk-Aware process to degrade more gradually. We have also observed in testing image generation algorithms that we achieve higher accuracy than the risk-neutral training, as measured by the log-likelihood of our image generation.
Our Solution
Photrek has demonstrated robust data generation and model learning using the Coupled Variational Autoencoder (CVAE). This algorithm was launched on the SingularityNET marketplace in the first quarter of 2023. Photrek’s testing of the algorithm showed a) impressive gains in accuracy and robustness with the MNIST Handwritten Numerals dataset and b) limitations with more complex datasets such as the CIFAR photographs. We plan to address this short coming by porting the development to PyTorch and incorporating deeper networks for higher resolution.
In addition to its Vision component, the World Models algorithm includes a Memory Module that utilizes a Recurrent Neural Network (RNN) to make forward predictions and a Controller based on a simple linear regression of the image generation and the RNN hidden state. The Simulating Risky Worlds project will focus on onboarding the Vision and Memory components which together provide predictive simulations, while the controller will be proposed for a future project. The World Models algorithm includes a temperature control that allows for modulating the degree of variance in the simulation. Under Photrek’s Deep Fund 1 Risk-Aware Data Generator for SingularityNET (Data Generator) project we integrated a risk-aware cost function into the algorithm. Our report will detail how these two controls coordinate to modify both the temperature and variations in the temperature in the simulation. Fluctuations in variance are the heart of complex, nonlinear systems. These are the domains that cause brittle algorithms to fail. Photrek as described in the roadmap below is seeking to provide a suite of algorithms to address these challenging applications.
Marketing Strategy and Service Benefits
Our third milestone objective is a customer discovery process in which we will demonstrate the Simulating a Risky World to potential customers. Organizations we are currently discussing our Risk-Aware services to include Vantage Risk, a reinsurance broker; the NOAA Severe Storm Laboratory; Rejuve, a SingularityNET startup modeling health risks; SingularityDOA, a SingularityNET DOA building automated trading bots; the SingularityNET sustainability team; and Brown University’s Community Noise Laboratory.
The SingularityNET community will benefit from DC-VAE because it learns to encode complex sequential information suitable for generating representative samples. Via unsupervised training through comparing next-in-sequence generated vs. actual images, such models conveniently require no truth-labeled data. Additionally, correlational relationships among embedding components are captured within a probability model. These capabilities empower such models to forecast likely future instances in a data sequence. A configurable “temperature” parameter controls the sampling variance at run-time. SingularityNET users will be able to generate sequences with a variable level of relative risk.
The DC-VAE architecture, which incorporates Google’s World Model, emphasizes a loosely-coupled modular design with representative spatial (“Vision”), temporal (“Memory”), and probabilistic component models scored against specifiable criteria. Photrek will make this accessible for SingularityNET for (1) incorporation of probabilistic components structured to represent heavy-tailed distributions associated with outliers and (2) scoring performance against Photrek’s Risk-Aware metrics. Such measures will result in a richer probabilistic representation of data so that training may be expedited; outlier events are better represented; and risk is fundamentally expressed in a meaningful, actionable manner.
Our Project Milestones and Cost Breakdown
Milestone 0: Contract Signing
Estimated start: November 1, 2023
Each milestone is planned as one month of effort.
Milestone 1: Integrate CVAE with Dynamics
Photrek is seeking to provide a service for simulating risky scenarios, particularly regarding the financial management of assets impacted by severe hazards. The first milestone will demonstrate the capabilities of DC-VAE achieved under the DF1 portion of the project and prioritize improvements to the service toward our long-term customer engagement goals. For instance, Google’s introduction of the World Model focuses on controlling games, while we are seeking to simulate wind and water damage from severe storms. Initial development will implement a baseline – DC-VAE components and risk-metric scoring into the Google Brain architecture – and test this baseline with similar data. We will perform experiments to quantify the ability to capture essential information about these risky events.
Milestone 2: Onboarding of Simulating a Risky World
The Photrek team has gained detailed expertise in the onboarding process through our Risk-Aware Assessment and Data Generator projects. Now that these services are hosted and usable for customers, we are confident about being able to add to our collection of Risk-Aware services.
The onboarding task for Simulating a Risky World will focus on selecting the right focus for the UI Dapp that will be attractive to potential customers and support Photrek’s long-term goal of managing financial assets impacted by severe hazards.
Milestone 3: Customer Engagement for utilizing simulation
Photrek utilizes a process of Customer Discovery to test hypotheses regarding the value offered to our customers. In this milestone, we will complete interviews with 10 potential customers and select 2 applications to evaluate the potential value of the Risky World simulation. The process will include an evaluation of how customers would utilize the simulation and how that utilization creates value different from other available options. The report on this milestone will provide guidance both for Photrek’s future development efforts and for the SingularityNET community regarding the identification of best use cases for SingularityNET apps. This process will build from our ongoing discussions with Vantage Risk, a reinsurance broker; the NOAA Severe Storm Laboratory; Rejuve, a SingularityNET startup modeling health risks; SingularityDOA, a SingularityNET DOA building automated trading bots; the SingularityNET sustainability team; and Brown University’s Community Noise Laboratory.
Milestone 4: Hosting Costs
The hosting costs will cover:
- the Google Cloud servers for the development training and the customer usage of the application,
- the Ethereum transaction fees
- AGIX fees for trial use of the application.
Budget
Roadmap and Status of Photrek's Risk-Aware Services
1. Onboarding Risk Assessment Service
As part of Deep Funding Round 1, Photrek proposed and delivered the Risk-Aware Assessment service on the SingularityNET platform. This service uses innovative concepts from information theory to assess the quality of Machine Intelligence algorithms, though the methods can be applied more generaMarketing Strategy and Product Benefitslly to any probabilistic forecasts. In particular, Photrek’s service creates a histogram of performance overlaid with metrics of Robustness, Accuracy, and Decisiveness. These evaluations are applied as test cases to an assessment of FiveThirtyEight’s 2022 midterm elections predictions as well as used to evaluate the performance of the Risk-Aware Data Generator (also funded in Round 1).
2. Demonstrating & Onboarding Coupled VAE
The Coupled VAE (
) uses
to improve the performance of a principal generative algorithm, the
. The CVAE allows learning of probabilistic models that can be used to generate images in a manner that is robust against corruption, such as noise or blur. Photrek is currently demonstrating the algorithm to potential customers.
3. Design of Dynamic and Mixture Model add-ons to Coupled VAE
The research component of our Deep Fund 1 project focused on generalizing a
to incorporate control of relative risk aversion using the methods of
. Our final report included an initial demonstration of the design utilizing Google’s DVAE.
4. Current Proposal: Onboard the Dynamic Coupled VAE
The Simulating Risky Worlds project will onboard the Dynamic Coupled VAE. The project is informed by abundant and lengthy conversations initiated by Photrek with dozens of industry leaders. A very clear signal from a variety of industries was the profound lack of quality data on which to train machine learning and artificial intelligence models. We can add significant value for our customers by producing high-quality, sequential, synthetic data. We will produce data for sequential numerical data, such as financial time series, and sequential image data, such as video.
5. Planned: Control in a Risky World; using simulation to train a controller
The Google World Model includes a simple controller. Photrek will propose at a later time to investigate applying our simulations to the training of control systems. For instance, financial portfolios require management decisions to respond to changes in the marketplace; utility companies require modifications to their disaster recovery plans as storm systems evolve; and agricultural systems need to respond to weather forecasts. While Google’s system utilizes a simple linear controller, the methods of reinforcement learning have developed significantly and should be evaluated for control applications.
Voluntary Revenue
Photrek will support the SingularityNET mission by offering 2.5% of the revenue during months in which the revenue from the application exceeds $5000. This offer is negotiable with the community and was selected based on a review of other project contracts from Deep Fund 1.
Open Source
Photrek’s Simulating Risky Worlds will be based on open-source software. Deep Brain’s World Model software is licensed under the Creative Commons Attribution CC-BY 4.0, which specifies open usage and requires attribution. Photrek utilizes a tiered licensing model for its software. The Nonlinear Statistical Coupling library provides the foundational functions for the coupled algebra and is issued under the Apache 2.0 license, which provides for private and public use. The Coupled VAE and Dynamic Coupled VAE utilize a GNU GPL 3.0 license, which requires users of the software to maintain their code as open source.
Our Team
Advisors
The Photrek team is working closely with Steve Smith, Director of R&D at Vantage Risk, and Harold Brooks, Senior Scientist at NOAA National Severe Storm Laboratories to identify applications for simulating the risks of severe storms. We are collaborating closely with the SingularityNET team including regular discussions with Jan Holdings and Matt Ikle. We thank Ben Goertzel for the vision to launch SingularityNET and for taking the time to discuss the Risk-Aware services.
Related Links
Cao, S., Li, J., Nelson, K. P. & Kon, M. A. Coupled VAE: Improved Accuracy and Robustness of a Variational Autoencoder. Entropy 24, 423 (2022).
Girin, L. et al. Dynamical Variational Autoencoders: A Comprehensive Review. FNT in Machine Learning 15, 1–175 (2021).
Ha, D. & Schmidhuber, J. World Models. (2018)
.