Research Project Title:
Towards Continuous Lifelong Learning
abstract:Despite the intriguing performances of the deep neural networks, they are far away from human intelligence, evidenced by their vulnerability to transformation. Moreover, the neural networks should be trained on a large dataset for multiple iterations to learn the complex probability distribution, while humans can infer from few examples. Also, training a neural network towards a new objective alters the learned representation, a phenomenon called catastrophic forgetting. This project is aimed at exploring the frontiers of few-shot learning, including the use of graph neural networks and amortized inference strategies. While previous works focused on few-shot learning accuracy, I would also like to emphasize whether the model can learn new information without disrupting previous knowledge.
“I decided to participate in SuperUROP after two years of great research experience in the machine learning field. I hope I will gain more insights into the field after this extensive research project.”