LEN1
Machine Learning Simulating Identify
After spending some time researching Machine Learning personally, I was astounded with how achievable training was within my skill set. Using GPT-2, a trained model for realistic chatbots, and a dataset containing all of my personal text messages since 2013, I took on the task of recreating myself as a chatbot.
I felt it would be best suited, based on my past experience, to frame this work with a Unity project that allowed you to freely explore a realized future home I may have in the distant future, cluttered with photos of my past that my chatbot could tell you about. While I could not pair the two together given the limitations of Unity and Transformers, I made both the Unity environment and the chatbot available online.
​
Directly below, you can find said Unity environment. Completely modeled by me using both Maya and Gravity Sketch, and interactability using WebGL speech synthesis.
(Use your mouse + WASD to look and move. Use ESCAPE to remove your cursor from the window.)
(Press "E" while looking at a photo to hear a description of it. Hold "F" to zoom in)
THE MODEL
Read here for a quick description of how I made this model, and how to access it
This model was trained using GPT-2, specifically Microsoft's DialoGPT trained model using GPT2. It is a model trained on thousands of interactions on Reddit to simulate real responses. The dataset was taken from a backup of my text messages dating back to 2013. Using Python, I was able to format the texts so that it may be trained using Pytorch from where DialoGPT left off. The model is only trained on 9 forwards and backs of texting, so you may need to reload the page if you'd like to talk further.
​
To access the model, click below. From the top open the "Runtime" menu, press "Run All" or (Ctrl + F9). After some time, scroll to the bottom of the page and you will see an IP such as 120.43.456. Press that link and you can begin chatting.