The Algebraic Value-Editing Conjecture in Deep Reinforcement Learning. Mentor: Andrew Gritsevskiy.The paper ( Alex Turner et al. “Understanding and controlling a maze-solving policy network” (2023)) showed that applying basic arithmetical operations on seemingly random activations of neural networks for different states can change the model’s output behvaior. I researched this theory, the algebraic value-editing conjecture which hypothesizes that the millions of activations in a neural network might have some order to them.

Deep Learning Transformers for Non-cyclical Kinematics. Mentor: Andrew Gritsevskiy.I worked with a relatively untested model in the field of predicting kinematics: transformers. Others have tried this with cyclical data, but the AI could just memorize the cycle rather than actually predict future decisions. Thus, I experimented with noncyclical data–a person cooking–to see if our model could predict their future actions and tested these transformers against state-of-the-art models.

Generate a spreadsheet report to list all services deployed in ICP production environment, grouped by each line of business.Simplify SpringBoot sample OpenAPI application and gRPC application so that they can be deployed in local Kubernetes environment.