Learning to Learn with Generative Models of Neural Network Checkpoints

We explore a data-driven approach for learning to optimize neural networks. We construct a dataset of neural network checkpoints and train a generative model on the parameters. In particular, our model is a conditional diffusion transformer that, given an initial input parameter vector and a prompted loss, error, or return, predicts the distribution over parameter updates that achieve the desired metric. At test time, it can optimize neural networks with unseen parameters for downstream tasks in just one update. We find that our approach successfully generates parameters for a wide range of loss prompts. Moreover, it can sample multimodal parameter solutions and has favorable scaling properties. We apply our method to different neural network architectures and tasks in supervised and reinforcement learning. Gradient-based optimization is the fuel of modern deep learning. Techniques of this class, such as SGD (Robbins & Monro, 1951) and Adam (Kingma & Ba, 2015), are easy to implement, scale reasonably well and converge to surprisingly good solutions--even in high-dimensional, non-convex neural network loss landscapes. Over the past decade, they have enabled impressive results in computer vision (Krizhevsky et al., 2012; Girshick et al., 2014), natural language processing (Vaswani et al., 2017; Radford et al., 2018) and audio generation (Van Den Oord et al., 2016). While these manual optimization techniques have led to large advances, they suffer from an important limitation: they are unable to improve from past experience. For example, SGD will not converge any faster when used to optimize the same neural network architecture from the same initialization the 100th time versus the first time. Learned optimizers capable of leveraging their past experiences have the potential to overcome this limitation and may accelerate future progress in deep learning.