Cyanide Cinema

• 

  Screenshot of accuracy and loss of the model at 10,000 epochs

Inspiration

We understand the impact of piracy on the movie industry, and we want to create something that can decrease the rates of piracy and ensure higher-profits and better reputations for these companies. At the same time, we also wanted to do something integrated with Machine Learning (ML); we were able to find a dataset and decided to go for it. Thus, we brought Cyanide Cinema into the world.

What it does

Cyanide Cinema utilizes our advanced Machine Learning algorithm to determine the percent likelihood of a movie being pirated. Cyanide Cinema also creates unique textual output for every piracy report, with options ranging from normal all the way to pirate mode (which imitates a pirate). The piracy probability, combined with detailed textual output, provides the user with everything they need to take action against piracy of their movie.

How we built it

In order to create Cyanide Cinema, we utilized Python, ReactJS, GPT-3.5, API and Tensorflow.js.

• Utilized Python for:

  • Laying the basis for our model
  • Training our model

• Used Tensorflow.js to:

  • Host and implement trained model
  • Estimate the likelihood of piracy

• Utilized GPT-3.5 to:

  • Create the textual outputs that accompany the figures
  • Includes Normal, Poetic, and Pirate mode
    • In order of increasing levels of humor

• Utilized ReactJS for:

  • developing the user interface
  • Hosting the final application

Challenges we ran into

Some challenges that we encountered were related to the development of the web app and the functionality of the model. Firstly, when developing the web app, we wanted to utilize a free LLAMA API or GPT-2 server and host it on GitHub to process report requests. However, it was difficult to find an effective model under 1 GB, which led us to resort to other sources like the GPT API.

Regarding the model, we struggled to create a dataset that would allow us to easily train the model while maintaining high accuracy. Nevertheless, through the use of manual and script-based imputation, we were able to create an ideal dataset and a highly effective model.

Accomplishments that we're proud of

By adjusting the constraints of the Machine Learning algorithm behind Cyanide Cinema, we were able to achieve a very low loss rate of under 0.5. Simultaneously, during our best trial—which took over 10 hours of runtime overnight—we achieved an accuracy of approximately 92.6%. Additionally, our team engineered prompts for the GPT-3.5 API to generate unique, humorous, and informative text output. The team behind Cyanide Cinema also succeeded in creating a high-end, functional, and interactive web application.

What we learned

We gained an understanding of advanced data processing techniques alongside effective tactics for Binary Classification (Sigmoids, 0-1 regression, etc.). Additionally, we gained valuable skills in manipulating AI via prompt engineering to achieve requested report results. These will allow us to further our understanding of deep learning.

What's next for Cyanide Cinema

In the future, we hope to increase the accuracy and precision of Cyanide Cinema through the analyzation of a larger and more comprehensive dataset, manual changes to the constraints of the Machine Learning model, and the addition of new features such as an estimated number of pirated copies.

Built With

• gpt-3.5-api
• python
• react
• tensorflow.js