DeepBias is an exploration of machine learning models aimed to predict perceived bias in U.S. broadcast news videos (e.g. CNN, Fox, MSNBC) based on “verbal” and “non-verbal” features.
In Spring 2019, Mike developed a video pre-processing pipeline, built and evaluated machine learning models (such as GRU-RNN, LSTM-RNN, FCN) for his Applied Machine Learning (6.892) class at MIT to study the relationship between extracted low-level features (such as affect and sentiment analysis of closed captions) and high-level features such as the perception of bias (captured with a questionnaire and presented to Amazon Turks). Mike received an A for this project and for the class.
In Summer 2019, Mike is leading three MIT undergraduate engineers to explore other architectures, word and sentence embeddings, and evaluate these models on larger data sets.
Python, Keras, Tensorflow, PyTorch, InferSent, Bert, IBM Watson NLP, Google Cloud Vision API, Affectiva SDK, MongoDB, ffmpeg, Kubernetes, Docker, AWS, Amazon Mechanical Turks, Leadership.