Abstract:Posterior sampling in contextual bandits with a Gaussian prior can be implemented exactly or approximately using the Laplace approximation. The Gaussian prior is computationally efficient but it cannot describe complex distributions. In this work, we propose approximate posterior sampling algorithms for contextual bandits with a diffusion model prior. The key idea is to sample from a chain of approximate conditional posteriors, one for each stage of the reverse process, which are estimated in a closed form using the Laplace approximation. Our approximations are motivated by posterior sampling with a Gaussian prior, and inherit its simplicity and efficiency. They are asymptotically consistent and perform well empirically on a variety of contextual bandit problems.
Abstract:Reproducible research in Machine Learning has seen a salutary abundance of progress lately: workflows, transparency, and statistical analysis of validation and test performance. We build on these efforts and take them further. We offer a principled experimental design methodology, based on linear mixed models, to study and separate the effects of multiple factors of variation in machine learning experiments. This approach allows to account for the effects of architecture, optimizer, hyper-parameters, intentional randomization, as well as unintended lack of determinism across reruns. We illustrate that methodology by analyzing Matching Networks, Prototypical Networks and TADAM on the miniImagenet dataset.
Abstract:Deep learning has mainly thrived by training on large-scale datasets. However, for a continual learning agent it is critical to incrementally update its model in a sample efficient manner. Learning semantic segmentation from few labelled samples can be a significant step toward such goal. We propose a novel method that constructs the new class weights from few labelled samples in the support set without back-propagation, while updating the previously learned classes. Inspiring from the work on adaptive correlation filters, an adaptive masked imprinted weights method is designed. It utilizes a masked average pooling layer on the output embeddings and acts as a positive proxy for that class. Our proposed method is evaluated on PASCAL-5i dataset and outperforms the state of the art in the 5-shot semantic segmentation. Unlike previous methods, our proposed approach does not require a second branch to estimate parameters or prototypes, and enables the adaptation of previously learned weights. Our adaptation scheme is evaluated on DAVIS video segmentation benchmark and our proposed incremental version of PASCAL and has shown to outperform the baseline model.
Abstract:Using variational Bayes neural networks, we develop an algorithm capable of accumulating knowledge into a prior from multiple different tasks. The result is a rich and meaningful prior capable of few-shot learning on new tasks. The posterior can go beyond the mean field approximation and yields good uncertainty on the performed experiments. Analysis on toy tasks shows that it can learn from significantly different tasks while finding similarities among them. Experiments of Mini-Imagenet yields the new state of the art with 74.5% accuracy on 5 shot learning. Finally, we provide experiments showing that other existing methods can fail to perform well in different benchmarks.
Abstract:The recent literature on deep learning offers new tools to learn a rich probability distribution over high dimensional data such as images or sounds. In this work we investigate the possibility of learning the prior distribution over neural network parameters using such tools. Our resulting variational Bayes algorithm generalizes well to new tasks, even when very few training examples are provided. Furthermore, this learned prior allows the model to extrapolate correctly far from a given task's training data on a meta-dataset of periodic signals.
Abstract:This paper presents greedy gossip with eavesdropping (GGE), a novel randomized gossip algorithm for distributed computation of the average consensus problem. In gossip algorithms, nodes in the network randomly communicate with their neighbors and exchange information iteratively. The algorithms are simple and decentralized, making them attractive for wireless network applications. In general, gossip algorithms are robust to unreliable wireless conditions and time varying network topologies. In this paper we introduce GGE and demonstrate that greedy updates lead to rapid convergence. We do not require nodes to have any location information. Instead, greedy updates are made possible by exploiting the broadcast nature of wireless communications. During the operation of GGE, when a node decides to gossip, instead of choosing one of its neighbors at random, it makes a greedy selection, choosing the node which has the value most different from its own. In order to make this selection, nodes need to know their neighbors' values. Therefore, we assume that all transmissions are wireless broadcasts and nodes keep track of their neighbors' values by eavesdropping on their communications. We show that the convergence of GGE is guaranteed for connected network topologies. We also study the rates of convergence and illustrate, through theoretical bounds and numerical simulations, that GGE consistently outperforms randomized gossip and performs comparably to geographic gossip on moderate-sized random geometric graph topologies.