InES Seminar Spring Semester 2023

Abstract:

In-context learning was popularized in the original GPT-3 release as a way to use language models to learn tasks with only a few examples. In in-context learning, we give the LM a prompt consisting of a list of input-output pairs that demonstrate a task. At the end of the prompt, we append a test input and allow the LM to make a prediction by simply conditioning on the prompt and predicting the next tokens. In this seminar you study the phenomenon of in-context learning and evaluate different explanations and it’s limitations.

Starting Papers

• Brown et al. Language Models are Few-Shot Learners
• Xie et al. An Explanation of In-Context Learning as Implicit Bayesian Inference
• Akyürek et al. What Learning Algorithm is In-Context Learning? Investigations with Linear Models

Abstract: Action-driven AI is a type of artificial intelligence that focuses on taking actions in response to input data. This approach to AI differs from other methods, such as decision-making AI, which focuses on making decisions based on the data without necessarily taking any direct action. In action-driven AI, the system is designed to take actions in the real world based on the information it receives. This can include things like moving a robotic arm to pick up an object or issuing a command to a drone to fly to a certain location. The goal of action-driven AI is to enable machines to perform tasks in the physical world, making them more useful and versatile. In this seminar you will study methods that are used to enable action-driven AI, the problem of alignment in the context of action-driven AI, and it’s limitations.

Starting Papers:

• Griffith et al. Policy shaping: Integrating human feedback with reinforcement learning
• Christiano et al. Deep reinforcement learning from human preferences
• Ouyang et al. Training language models to follow instructions with human feedback.

Introduction: For longitudinal data, classic neural network architectures for sequential data, such as RNNs do not always perform well. Learning from short and irregular sequences is challenging. Hence, specialized neural network architectures have been proposed in the past to model such data. In the medical area, examples of longitudinal data are hospital patient data with measurements for several days or trajectories of cancer patients presenting at different times. Tasks to be solved with neural networks include trajectory prediction, hospital readmission prediction or mortality prediction.

Goal and Objective: In this seminar, participants will become familiar with longitudinal data in the medical context and the related machine learning challenges. The focus will be to get a deeper understanding of selected specialized neural network architectures beyond RNNs designed particularly for longitudinal medical data. Hence, the final report is expected to be an overview of tasks, datasets and selected methods.

Starting Papers:

• Mandel, F., Ghosh, R. P., & Barnett, I. (2021). Neural networks for clustered and longitudinal data using mixed effects models. Biometrics.
• Nguyen, M., He, T., An, L., Alexander, D. C., Feng, J., Yeo, B. T., & Alzheimer's Disease Neuroimaging Initiative. (2020). Predicting Alzheimer's disease progression using deep recurrent neural networks. NeuroImage, 222, 117203.
• Zhang, J., Kowsari, K., Harrison, J. H., Lobo, J. M., & Barnes, L. E. (2018). Patient2vec: A personalized interpretable deep representation of the longitudinal electronic health record. IEEE Access, 6, 65333-65346.

Abstract:

Prediction tasks over nodes and edges in networks require careful effort in engineering features used by learning algorithms. Recent research in the broader field of representation learning has led to significant progress in automating prediction by learning the features themselves. However, present feature learning approaches are not expressive enough to capture the diversity of connectivity patterns observed in networks. Here we propose node2vec, an algorithmic framework for learning continuous feature representations for nodes in networks. In node2vec, we learn a mapping of nodes to a low-dimensional space of features that maximizes the likelihood of preserving network neighborhoods of nodes.
The aim of this seminar is to summarize the state-of-the-art of node2vec methods and to which problems these methods can be successfully applied.

Starting Papers

• Grover, A., & Leskovec, J. (2016, August). node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 855–864).
• Qiu, J., Dong, Y., Ma, H., Li, J., Wang, K., & Tang, J. (2018, February). Network embedding as matrix factorization: Unifying deepwalk, line, pte, and node2vec. In Proceedings of the eleventh ACM international conference on web search and data mining (pp. 459–467).