We aim to develop novel machine learning and deep learning algorithms based on in-depth research on theories and principles, including(but are not limited to):
- Increasing the efficiency of learning algorithms
- Achieving interpretable AI
- Mitigating algorithmic bias
- Foreseeing mainstream adoption of AI and robots.
First, complex models (e.g., deep learning algorithms that use large neural networks) require a great amount of computational resources, such as power-hungry GPUs. We will seek fundamental solutions to train and utilize complex models more efficiently – spending less energy and memory without sacrificing speed or accuracy.
Next, most of the existing AI approaches are known to be ‘black-box’ solutions in that they perform well at certain tasks but do not provide rationale behind their decisions. The research group will aim to fill this gap and develop neural architectures and training algorithms of high interpretability – it be easy to track the inference process and to identify which factors have led the models to make a certain prediction.
Lastly, many existing training algorithms make a wild generalization of patterns learned from a specific training dataset. An issue arises when the training dataset is biased, reflecting the prejudices within our society, leading to a biased machine learning model. The research group will develop technological solutions that mitigate such bias through, for instance, the Bayesian approach that takes into consideration the pre-existing human biases as prior information.
- Theories and Principles
- Unsupervised annotations using deep learning.
- AI and robot rights.
- Super-resolution GIS using deep learning.
- Modeling sleep via wearable devices.
- Behavioral Modeling
- Anger propagation via social media.
- Dynamics of public discourse on social media.
- Dynamics of content popularity on streaming media.
- Leveraging NLP for sentiment detection.
- Off-line human behavior prediction.
- Consumption patterns of online news portals.
- Designs and Applications
- Extracting economic scales from satellite imagery.
- AI for media literacy.
- Tax fraud detection.