Deep Learning over Heterogeneous Data: a challenge, a solution, and an application to Poly(A) signal prediction
The talk will start with the introduction of the most basic building blocks of deep learning models (especially convolutional neural networks) to build some statistical intuition of what deep learning is supposed to be capable of, and will show some evidence supporting that what is behind the promised human-level understanding of data is partially the model's tendency to capture the high-frequent superficial signals. The second part of the talk will discuss a recent method-development work focusing on overcoming the model's tendency in capturing superficial signals in visual applications. Finally, the talk will conclude with an application of cross-species Poly(A) signal prediction work. This is an ongoing project demonstrating the usage of the proposed method in the second part of the talk.