Polina Barabanshchikova

Large-scale foundation models have become ubiquitous in machine learning research, but their training is resource-intensive. This creates a need for efficient reusing and fine-tuning of pre-trained models for new tasks. Additionally, many practical problems require knowledge from multiple domains, which a single model cannot fully address. This PhD project investigates methods for combining pre-trained generative models into complex systems tailored to specific tasks. By composing generative processes, we aim to enhance the capacity of individual models and introduce careful control over the resulting distribution. The goal is to construct compositions that can solve tasks entirely unseen during training, with potential applications in areas like drug discovery.