We provide a probabilistic interpretation of attention and show that the standard dot-product attention in transformers is a special case of Maximum A Posteriori (MAP) inference. The proposed approach suggests the use of Expectation Maximization algorithms for on-line adaptation of key and value model parameters. This approach is useful for cases in which external agents, e.g., annotators, provide inference-time information about the correct values of some tokens, e.g., the semantic category of some pixels, and we need for this new information to propagate to other tokens in a principled manner. We illustrate the approach on an interactive semantic segmentation task in which annotators and models collaborate online to improve annotation efficiency. Using standard benchmarks, we observe that key adaptation boosts model performance ($\sim10\%$ mIoU) in the low feedback regime and value propagation improves model responsiveness in the high feedback regime.

Specifically, we plot the improvement in mask accuracy, i.e. mean IOU relative to ground truth, as a function of

We provide a probabilistic interpretation of attention and show that the standard dot-product attention in transformers is a special case of Maximum A Posteriori (MAP) inference. The proposed approach suggests the use of Expectation Maximization algorithms for on-line adaptation of key and value model parameters. This approach is useful for cases in which external agents, e.g., annotators, provide inference-time information about the correct values of some tokens, e.g., the semantic category of some pixels, and we need for this new information to propagate to other tokens in a principled manner. We illustrate the approach on an interactive semantic segmentation task in which annotators and models collaborate online to improve annotation efficiency. Using standard benchmarks, we observe that key adaptation boosts model performance ( \sim10\% mIoU) in the low feedback regime and value propagation improves model responsiveness in the high feedback regime.

We provide a probabilistic interpretation of attention and show that the standard dot-product attention in transformers is a special case of Maximum A Posteriori (MAP) inference. The proposed approach suggests the use of Expectation Maximization algorithms for on-line adaptation of key and value model parameters. This approach is useful for cases in which external agents, e.g., annotators, provide inference-time information about the correct values of some tokens, e.g., the semantic category of some pixels, and we need for this new information to propagate to other tokens in a principled manner. We illustrate the approach on an interactive semantic segmentation task in which annotators and models collaborate online to improve annotation efficiency. Using standard benchmarks, we observe that key adaptation boosts model performance ( \sim10\% mIoU) in the low feedback regime and value propagation improves model responsiveness in the high feedback regime.

Multiple Instance Learning (MIL) is a weakly supervised learning paradigm that is becoming increasingly popular because it requires less labeling effort than fully supervised methods. This is especially interesting for areas where the creation of large annotated datasets remains challenging, as in medicine. Although recent deep learning MIL approaches have obtained state-of-the-art results, they are fully deterministic and do not provide uncertainty estimations for the predictions. In this work, we introduce the Attention Gaussian Process (AGP) model, a novel probabilistic attention mechanism based on Gaussian Processes for deep MIL. AGP provides accurate bag-level predictions as well as instance-level explainability, and can be trained end-to-end. Moreover, its probabilistic nature guarantees robustness to overfitting on small datasets and uncertainty estimations for the predictions. The latter is especially important in medical applications, where decisions have a direct impact on the patient's health. The proposed model is validated experimentally as follows. First, its behavior is illustrated in two synthetic MIL experiments based on the well-known MNIST and CIFAR-10 datasets, respectively. Then, it is evaluated in three different real-world cancer detection experiments. AGP outperforms state-of-the-art MIL approaches, including deterministic deep learning ones. It shows a strong performance even on a small dataset with less than 100 labels and generalizes better than competing methods on an external test set. Moreover, we experimentally show that predictive uncertainty correlates with the risk of wrong predictions, and therefore it is a good indicator of reliability in practice. Our code is publicly available.

We provide a probabilistic interpretation of attention and show that the standard dot-product attention in transformers is a special case of Maximum A Posteriori (MAP) inference. The proposed approach suggests the use of Expectation Maximization algorithms for online adaptation of key and value model parameters. This approach is useful for cases in which external agents, e.g., annotators, provide inference-time information about the correct values of some tokens, e.g, the semantic category of some pixels, and we need for this new information to propagate to other tokens in a principled manner. We illustrate the approach on an interactive semantic segmentation task in which annotators and models collaborate online to improve annotation efficiency. Using standard benchmarks, we observe that key adaptation boosts model performance ($\sim10\%$ mIoU) in the low feedback regime and value propagation improves model responsiveness in the high feedback regime. A PyTorch layer implementation of our probabilistic attention model will be made publicly available here: https://github.com/apple/ml-probabilistic-attention.

Sequence-to-sequence constituency parsing casts the tree structured prediction problem as a general sequential problem by top-down tree linearization,and thus it is very easy to train in parallel with distributed facilities. Despite its success, it relies on a probabilistic attention mechanism for a general purpose, which can not guarantee the selected context to be informative in the specific parsing scenario. Previous work introduced a deterministic attention to select the informative context for sequence-to-sequence parsing, but it is based on the bottom-up linearization even if it was observed that top-down linearization is better than bottom-up linearization for standard sequence-to-sequence constituency parsing. In this paper, we thereby extend the deterministic attention to directly conduct on the top-down tree linearization. Intensive experiments show that our parser delivers substantial improvements over the bottom-up linearization in accuracy, and it achieves 92.3 Fscore on the Penn English Treebank section 23 and 85.4 Fscore on the Penn Chinese Treebank test dataset, without reranking or semi-supervised training.