Influence functions (IF) have been seen as a technique for explaining model predictions through the lens of the training data. Their utility is assumed to be in identifying training examples "responsible" for a prediction so that, for example, correcting a prediction is possible by intervening on those examples (removing or editing them) and retraining the model. However, recent empirical studies have shown that the existing methods of estimating IF predict the leave-one-out-and-retrain effect poorly. In order to understand the mismatch between the theoretical promise and the practical results, we analyse five assumptions made by IF methods which are problematic for modern-scale deep neural networks and which concern convexity, numeric stability, training trajectory and parameter divergence. This allows us to clarify what can be expected theoretically from IF. We show that while most assumptions can be addressed successfully, the parameter divergence poses a clear limitation on the predictive power of IF: influence fades over training time even with deterministic training. We illustrate this theoretical result with BERT and ResNet models. Another conclusion from the theoretical analysis is that IF are still useful for model debugging and correcting even though some of the assumptions made in prior work do not hold: using natural language processing and computer vision tasks, we verify that mis-predictions can be successfully corrected by taking only a few fine-tuning steps on influential examples.

We address efficient calculation of influence functions for tracking predictions back to the training data. We propose and analyze a new approach to speeding up the inverse Hessian calculation based on Arnoldi iteration. With this improvement, we achieve, to the best of our knowledge, the first successful implementation of influence functions that scales to full-size (language and vision) Transformer models with several hundreds of millions of parameters. We evaluate our approach in image classification and sequence-to-sequence tasks with tens to a hundred of millions of training examples. Our code is available at https://github.com/google-research/jax-influence.

Feature attribution a.k.a.input salience methods which assign an importance score to a feature are abundant but may produce surprisingly different results for the same model on the same input.While differences are expected if disparate definitions of importance are assumed, most methods claim to provide faithful attributions and point at the features most relevant for a model's prediction.Existing work on faithfulness evaluation is not conclusive and does not provide a clear answer as to how different methods are to be compared.Focusing on text classification and the model debugging scenario, our main contribution is a protocol for faithfulness evaluation that makes use of partially synthetic data to obtain ground truth for feature importance ranking.Following the protocol, we do an in-depth analysis of four standard salience method classes on a range of datasets and lexical shortcuts for BERT and LSTM models.We demonstrate that some of the most popular method configurations provide poor results even for simple shortcuts while a method judged to be too simplistic works remarkably well for BERT.

Unsupervised multi-object scene decomposition is a fast-emerging problem in representation learning. Despite significant progress in static scenes, such models are unable to leverage important dynamic cues present in videos. We propose PROVIDE, a novel unsupervised framework for PRObabilistic VIdeo DEcomposition based on a temporal extension of iterative inference. PROVIDE is powerful enough to jointly model complex individual multi-object representations and explicit temporal dependencies between latent variables across frames. This is achieved by leveraging 2D-LSTM, temporally conditioned inference and generation within the iterative amortized inference for posterior refinement. Our method improves the overall quality of decompositions, encodes information about the objects’ dynamics, and can be used to predict trajectories of each object separately. Additionally, we show that our model has a high accuracy even without color information. We demonstrate the decomposition capabilities of our model and show that it outperforms the state-of-the-art on several benchmark datasets, one of which was curated for this work and will be made publicly available.