Kordel Kade France

---

Uncertainty of the environment limits the circumstances with which any optimization problem can provide meaningful information. Multiple optimizers can combat this problem by communicating different information through cooperative coevolution. In reinforcement learning (RL), uncertainty can be reduced by applying learned policies collaboratively with another agent. Here, we propose DeadReckon to evolve an optimal policy for such scenarios. We hypothesize that self-paced co-training can allow factored particle swarms with imperfect knowledge to consolidate knowledge from each of their imperfect policies in order to approximate a single optimal policy. Additionally, we show how the performance of co-training swarms of RL agents can be maximized through the specific use of Expected SARSA as the policy learner. We evaluate DeadReckon against comparable RL algorithms and attempt to establish limits for which our procedure does and does not provide benefit. Our results indicate that Particle Swarm Optimization (PSO) is effective in training multiple agents under uncertainty and that FEA reduces swarm and policy updates. This paper contributes to the field of cooperative co-evolutionary algorithms by proposing a method by which factored evolutionary techniques can significantly improve how multiple RL agents collaborate under extreme uncertainty to solve complex tasks faster than a single agent can under identical conditions.

Since high school, I have been enthralled with trying to find order in the chaos of the markets through algorithms. A couple of years ago, I put my money where my code was and incorporated a legitimate algorithmic trading fund to pursue this interest deeper. The fund operates in full autonomy and trades through artificial intelligence software I built called Goldilox (all AI algorithms are effectively just  optimization procedures, and the story of Goldilocks and the Three Bears is the original tale of optimization). 

I've turned the problem of stock trading into a controls problem, leveraging control theory, Kalman filters, and reinforcement learning to take on the market with a slightly different angle.

At the time of writing, the fund is divided into five different "sub-funds", each one analyzing a different AI algorithm and its rate of return. I publish my findings on these algorithms as meaningful results are achieved. What started out as an experiment has turned into a very rewarding part-time job, with Goldilox retraining models and sifting through data while the markets are closed and it isn't trading. Goldilox uses A LOT of NLP and reinforcement learning algorithms and retrains itself several times per week with a custom federated learning pipeline--all built by me. Goldilox also contains its own back-testing framework to fully vet each algorithm under historical scenarios before deployment. 

Each share of the fund is tokenized by a cryptocurrency token I built. In other words, the fund is legally tied to a smart contract. Here is a link to the smart contract: https://etherscan.io/token/0xe1ef87078afb008e17dd8c3d604999f332b2724f

Nearly every waking moment not spent on work or other projects is spent maturing Goldilox and its capabilities. I do not believe the markets can be "solved" but I do believe AI can find patterns not totally obvious to humans.   https://goldilox.ai

Seekar TRACE (Target Recognition and Acquisition in Complex Environments) was a target recognition platform that was originally developed for military applications. By adding one hell of a pre-processing stack, we built a very impressive target recognition platform that could identify targets through smoke, fog, rain, forestry, cracked camera lenses and more. Even more, we were able to compress it well enough to run on edge devices like phones, tablets, microcontrollers, and drone cameras. We coupled it with radar and electrochemical sensors to build a multi-functional sensor stack capable of performing autonomous navigation for small vehicles. Eventually, we forked Seekar TRACE to also be used in medical imaging to identify difficult patterns in X-Rays.  Some early highlights of the platform can be found here: https://seekartech.com/defense

This is a pet project that has turned into a significant blend of NLP, computer vision, and Deep RL. On its third iteration now, I've built and programmed a robotic arm to exactly mirror me by interpreting the movement of my left arm through a camera. There is a computer vision model using stacked hourglass networks (https://arxiv.org/pdf/1603.06937.pdf) running on an iOS app, a SARSA reinforcement learning model, and a lot of closed-feedback control loops. As a result, the project requires extreme fluency in three programming languages - Swift, Python, and C++. Eventually, I plan to have the robot arm learn new tasks (i.e. use a screwdriver) by analyzing YouTube videos and learning the actions performed.

If  a  human  is  injured,  the  biology  of its  body  immediately  begins  repairing  itself.  Blood starts to clot in flesh wounds, antibodies are produced as a response to a virus, and neural structure  changes  to  accommodate  brain  damage through  neuroplasticity.  Achieving an accurate model of artificial  neuroplasticity is my ongoing focus with this project. In order to grant robotics  the  same  liberties  that  humans  enjoy  in regards to self-repair, I show how this can effectively  be  accomplished  through  the cooperation and reorganization of two neural networks. The process takes principles from few-shot learning in Natural Language Processing (NLP - https://arxiv.org/abs/2001.07676) and  general adversarial  networks  (GANs - https://arxiv.org/abs/1406.2661) .  If  a  sensor  sending input to an artificial neural network becomes damaged, the neural network of the damaged sensor  will  repurpose  itself  to  be  used  for  another task through self-reorganization and on-the-fly transfer learning. In other words, a robot can exhibit self-repair by borrowing from  principles  of  biological neuroplasticity.   Read more here: http://dx.doi.org/10.13140/RG.2.2.27593.06248

This is a pet project that has turned into a significant blend of NLP, computer vision, and Deep RL. On its third iteration now, I've built and programmed a robotic arm to exactly mirror me by interpreting the movement of my left arm through a camera. There is a computer vision model using stacked hourglass networks (https://arxiv.org/pdf/1603.06937.pdf) running on an iOS app, a SARSA reinforcement learning model, and a lot of closed-feedback control loops. As a result, the project requires extreme fluency in three programming languages - Swift, Python, and C++. Eventually, I plan to have the robot arm learn new tasks (i.e. use a screwdriver) by analyzing YouTube videos and learning the actions performed.

For local neuropsychology clinics, I build voice and emotion recognition software to detect neuropsychological conditions through natural language processing and computer vision. Software was updated biweekly and advised the neuropsychologist on symptoms detected in diagnostic interviews through a camera and microphone. The AI runs through a mobile app and is intended to help screen for symptoms of neurodegenerative disorders such as Alzheimer's disease and Parkinson's disease. An onboard AI agent tracks eye movement, blink rate, skeletal movement, facial expressions, and even analyzes speech for vocal indicators. The app does not need a cloud connection in order to operate and is designed under HIPAA compliant data practices.  See a demo here: https://seekartech.com/healthcare

At Seekar, I built a computer vision model that was capable of identifying up to seven different respiratory conditions within chest X-rays. We leveraged significant knowledge distillation principles and "cluster models" to beat the performance of one larger model. The result was a hierarchy of much smaller models that we could compress small enough to fit on a mobile app. The app went through an ethics-board approved clinical trial and proved to be 95.5% as accurate as a radiologist in detecting the seven respiratory conditions.

Seekar donated the app as an effort to relieve physician demand during the COVID-19 pandemic.
Read more here: https://www.researchgate.net/publication/345761193_Cluster_Neural_Networks_for_Edge_Intelligence_in_Medical_ImagingSee the product: https://apps.apple.com/us/app/covid-ai/id1505887668

At Seekar, my team and I were recruited to develop a device that could be used for the autonomous detection of explosive compounds such as C4, TNT, RDX, and gunpowder. In collaboration with university professors at the University of Texas at Dallas, we were able to develop a device and sensor capable of detecting these compounds at a distance of nearly 30 feet away. Additionally, we developed a fully custom motherboard with me leading development of all firmware and software. The result was a truly novel device developed entirely in-house that incorporated electrochemical sensing and edge AI.  Pairing this with radar, computer vision, and some control hardware, we will  enable this technology to navigate drones through buildings in order to find explosives.

A significant research effort showing how the TrueDepth camera on an Apple iPhone--more specifically time-of-flight sensors in general--can be reprogrammed as a thermometer. We establish our hypothesis, show our experimental setup, and analyze data obtained from experiments. The results suggest that the TrueDepth sensor has thermometric reading capabilities due to the specific part of the electromagnetic spectrum that the sensor wavelength resides in.
Read more here: https://github.com/KordelFranceTech/Research-TrueDepthCameraAsThermometer

As part of an effort to attain more control over my training performance, I've developed my own framework for implementing neural networks. Activation functions, convolution, SGD, Backdrop, etc are all implemented from scratch with Numpy as the only library. It is effectively my own custom-built version of Tensorflow.

The twist I've added is that I've encoded each neuron with an identifier so I can trace the firing pattern of inference throughout the network. This allows me to see which neurons are more frequently used so that the network can perform "self-pruning" to make itself smaller and more efficient.

Find it here: https://github.com/KordelFranceTech/Academic-DeepNeuralNetsFromScratch

Skoped AI was designed for the purposes of training law enforcement and homeland security officers in mixed reality. My component, called Range Vision, was an iOS application that ran on an iPhone mounted to a rifle scope and provided augmented reality overlays for ballistic trajectory calculations. This allowed the trainee to interpret how estimated wind conditions and range would affect the drop, drift, and spin of the rifle shot by visualizing the shot before it was fired. I eventually developed a 4-degrees-of-freedom (4-DOF) calculator into the software that calculated bullet drop, drift, pitch, spin, and even aerodynamic jump. The app incorporated target recognition and included an optical rangefinder that used two cameras to measure distance instead of a laser.

I also built the software to Bluetooth pair with a Kestrel wind meter to read wind data at the source. One of the most difficult aspects of ballistics calculations is estimating wind conditions at the target so that the marksman can compensate for changing wind pattern. Our last effort on the project was to use topographical information as well as the camera zoomed through the rifle scope to predict wind at the target and increase confidence in the trajectory.

The coolest thing about this project is that it allowed me to translate a heavy dose of aerodynamics equations and mathematics into code. The product was validated at ranges up to 1200 yards, but the calculations theoretically prove it for much longer distances.

Find it here: https://apps.apple.com/us/app/range-vision/id1465211163