Deep learning is known to be prone to overfitting to the data on which it was trained, especially when data is scarce, thus hindering its generalization capability. One promising direction to help build better models is to incorporate domain-specific knowledge to guide the training process. This aims to help the model identify relevant patterns instead of picking up spurious correlations in the data. An immune repertoire is a set of ~10^8 sequences that correspond to receptors in the immune system, and records all the past and ongoing immune responses in an individual. This makes such data suitable for deep learning algorithms that can be used to extract useful information.
While deep learning has shown promise in classifying these repertoires, even though labels are available only at the individual level and not at the sequence level, the issue of overfitting persists and is particularly problematic in this task if the percentage of disease-related sequences in the repertoires is low. In this work, we try to incorporate domain-specific knowledge by using an additional set of sequences that are experimentally known to be disease-related, in order to guide the whole training process.
Program
11:30 – Doors open and lunch is served
12:00 – "Incorporating Prior Domain Knowledge into Deep Learning Models: A Case of Disease State Prediction" by Ghadi S. Al Hajj (PhD Candidate, Department of Informatics)
This event is open for all PhD candidates and postdocs. No registration needed.
About the seminar series
Once a month, dScience will invite you to join us for lunch, soft drinks and professional talks at the Science Library. In addition to these, we will serve lunch to PhD candidates in our lounge in Kristine Bonnevies hus every Thursday. Due to limited space (40 people), this will be first come, first served. See how to find us here (download).
Our lounge can also be booked by PhDs and Postdocs on a regular basis, whether it is for a meeting or just to hang out – we have fresh coffee all day long! Read more about the seminar series here.