ICLR 2023 Papers

We present Generalizable NeRF Transformer (GNT), a transformer-based architecture that reconstructs Neural Radiance Fields (NeRFs) and learns to render novel views on the fly from source views. While prior works on NeRFs optimize a scene representation by inverting a handcrafted rendering equation, GNT achieves neural representation and rendering that generalizes across scenes using transformers at two stages. (1) The view transformer leverages multi-view geometry as an inductive bias for attention-based scene representation, and predicts coordinate-aligned features by aggregating information from epipolar lines on the neighboring views. (2) The ray transformer renders novel views using attention to decode the features from the view transformer along the sampled points during ray marching. Our experiments demonstrate that when optimized on a single scene, GNT can successfully reconstruct NeRF without an explicit rendering formula due to the learned ray renderer. When trained on multiple scenes, GNT consistently achieves state-of-the-art performance when transferring to unseen scenes and outperform all other methods by ~10% on average. Our analysis of the learned attention maps to infer depth and occlusion indicate that attention enables learning a physically-grounded rendering. Our results show the promise of transformers as a universal modeling tool for graphics. Please refer to our project page for video results: https://vita-group.github.io/GNT/

We study whether automatically-induced prompts that effectively extract information from a language model can also be used, out-of-the-box, to probe other language models for the same information. After confirming that discrete prompts induced with the AutoPrompt algorithm outperform manual and semi-manual prompts on the slot-filling task, we demonstrate a drop in performance for AutoPrompt prompts learned on a model and tested on another. We introduce a way to induce prompts by mixing language models at training time that results in prompts that generalize well across models. We conduct an extensive analysis of the induced prompts, finding that the more general prompts include a larger proportion of existing English words and have a less order-dependent and more uniform distribution of information across their component tokens. Our work provides preliminary evidence that it's possible to generate discrete prompts that can be induced once and used with a number of different models, and gives insights on the properties characterizing such prompts.

In recent years, deep learning models have set new standards in predicting neural population responses. Most of these models currently focus on predicting the mean response of each neuron for a given input. However, neural variability around this mean is not just noise and plays a central role in several theories on neural computation. To capture this variability, we need models that predict full response distributions for a given stimulus. However, to measure the quality of such models, commonly used correlation-based metrics are not sufficient as they mainly care about the mean of the response distribution. An interpretable alternative evaluation metric for likelihood-based models is \textit{Information Gain} (IG) which evaluates the likelihood of a model relative to a lower and upper bound. However, while a lower bound is usually easy to obtain, constructing an upper bound turns out to be challenging for neural recordings with relatively low numbers of repeated trials, high (shared) variability, and sparse responses. In this work, we generalize the jack-knife oracle estimator for the mean---commonly used for correlation metrics---to a flexible Bayesian oracle estimator for IG based on posterior predictive distributions. We describe and address the challenges that arise when estimating the lower and upper bounds from small datasets. We then show that our upper bound estimate is data-efficient and robust even in the case of sparse responses and low signal-to-noise ratio. We further provide the derivation of the upper bound estimator for a variety of common distributions including the state-of-the-art zero-inflated mixture models, and relate IG to common mean-based metrics. Finally, we use our approach to evaluate such a mixture model resulting in $90\%$ IG performance.

As an intrinsic and fundamental property of big data, data heterogeneity exists in a variety of real-world applications, such as in agriculture, sociology, health care, etc. For machine learning algorithms, the ignorance of data heterogeneity will significantly hurt the generalization performance and the algorithmic fairness, since the prediction mechanisms among different sub-populations are likely to differ. In this work, we focus on the data heterogeneity that affects the prediction of machine learning models, and first formalize the Predictive Heterogeneity, which takes into account the model capacity and computational constraints. We prove that it can be reliably estimated from finite data with PAC bounds even in high dimensions. Additionally, we propose the Information Maximization (IM) algorithm, a bi-level optimization algorithm, to explore the predictive heterogeneity of data. Empirically, the explored predictive heterogeneity provides insights for sub-population divisions in agriculture, sociology, and object recognition, and leveraging such heterogeneity benefits the out-of-distribution generalization performance.

Variational Autoencoders (VAEs) were originally motivated as probabilistic generative models in which one performs approximate Bayesian inference. The proposal of $\beta$-VAEs breaks this interpretation and generalizes VAEs to application domains beyond generative modeling (e.g., representation learning, clustering, or lossy data compression) by introducing an objective function that allows practitioners to trade off between the information content ("bit rate") of the latent representation and the distortion of reconstructed data. In this paper, we reconsider this rate/distortion trade-off in the context of hierarchical VAEs, i.e., VAEs with more than one layer of latent variables. We propose a method to control each layer's contribution to the rate independently. We identify the most general class of inference models to which our proposed method is applicable, and we derive theoretical bounds on the performance of downstream tasks as functions of the individual layers' rates. Our experiments demonstrate that the proposed method allows us to better tune hierarchical VAEs for a diverse set of practical use cases.

This paper uses information-theoretic tools to analyze the generalization error in unsupervised domain adaptation (UDA). We present novel upper bounds for two notions of generalization errors. The first notion measures the gap between the population risk in the target domain and that in the source domain, and the second measures the gap between the population risk in the target domain and the empirical risk in the source domain. While our bounds for the first kind of error are in line with the traditional analysis and give similar insights, our bounds on the second kind of error are algorithm-dependent, which also provide insights into algorithm designs. Specifically, we present two simple techniques for improving generalization in UDA and validate them experimentally.

The information-theoretic framework promises to explain the predictive power of neural networks. In particular, the information plane analysis, which measures mutual information (MI) between input and representation as well as representation and output, should give rich insights into the training process. This approach, however, was shown to strongly depend on the choice of estimator of the MI. The problem is amplified for deterministic networks if the MI between input and representation is infinite. Thus, the estimated values are defined by the different approaches for estimation, but do not adequately represent the training process from an information-theoretic perspective. In this work, we show that dropout with continuously distributed noise ensures that MI is finite. We demonstrate in a range of experiments that this enables a meaningful information plane analysis for a class of dropout neural networks that is widely used in practice.

Denoising diffusion models have spurred significant gains in density modeling and image generation, precipitating an industrial revolution in text-guided AI art generation. We introduce a new mathematical foundation for diffusion models inspired by classic results in information theory that connect Information with Minimum Mean Square Error regression, the so-called I-MMSE relations. We generalize the I-MMSE relations to \emph{exactly} relate the data distribution to an optimal denoising regression problem, leading to an elegant refinement of existing diffusion bounds. This new insight leads to several improvements for probability distribution estimation, including a theoretical justification for diffusion model ensembling. Remarkably, our framework shows how continuous and discrete probabilities can be learned with the same regression objective, avoiding domain-specific generative models used in variational methods.

Interpretability is a pressing issue for machine learning. Common approaches to interpretable machine learning constrain interactions between features of the input, sacrificing model complexity in order to render more comprehensible the effects of those features on the model's output. We approach interpretability from a new angle: constrain the information about the features without restricting the complexity of the model. We use the Distributed Information Bottleneck to optimally compress each feature so as to maximally preserve information about the output. The learned information allocation, by feature and by feature value, provides rich opportunities for interpretation, particularly in problems with many features and complex feature interactions. The central object of analysis is not a single trained model, but rather a spectrum of models serving as approximations that leverage variable amounts of information about the inputs. Information is allocated to features by their relevance to the output, thereby solving the problem of feature selection by constructing a learned continuum of feature inclusion-to-exclusion. The optimal compression of each feature---at every stage of approximation---allows fine-grained inspection of the distinctions among feature values that are most impactful for prediction. We develop a framework for extracting insight from the spectrum of approximate models and demonstrate its utility on a range of tabular datasets.

Temporal networks model a variety of important phenomena involving timed interactions between entities. Existing methods for machine learning on temporal networks generally exhibit at least one of two limitations. First, many methods assume time to be discretized, so if the time data is continuous, the user must determine the discretization and discard precise time information. Second, edge representations can only be calculated indirectly from the nodes, which may be suboptimal for tasks like edge classification. We present a simple method that avoids both shortcomings: construct the line graph of the network, which includes a node for each interaction, and weigh the edges of this graph based on the difference in time between interactions. From this derived graph, edge representations for the original network can be computed with efficient classical methods. The simplicity of this approach facilitates explicit theoretical analysis: we can constructively show the effectiveness of our method's representations for a natural synthetic model of temporal networks. Empirical results on real-world networks demonstrate our method's efficacy and efficiency on both link classification and prediction.

Model compression is vital to the deployment of deep learning on edge devices. Low precision representations, achieved via quantization of weights and activations, can reduce inference time and memory requirements. However, quantifying and predicting the response of a model to the changes associated with this procedure remains challenging. This response is non-linear and heterogeneous throughout the network. Understanding which groups of parameters and activations are more sensitive to quantization than others is a critical stage in maximizing efficiency. For this purpose, we propose FIT. Motivated by an information geometric perspective, FIT combines the Fisher information with a model of quantization. We find that FIT can estimate the final performance of a network without retraining. FIT effectively fuses contributions from both parameter and activation quantization into a single metric. Additionally, FIT is fast to compute when compared to existing methods, demonstrating favourable convergence properties. These properties are validated experimentally across hundreds of quantization configurations, with a focus on layer-wise mixed-precision quantization.

Predicting the responses of a cell under perturbations may bring important benefits to drug discovery and personalized therapeutics. In this work, we propose a novel graph variational Bayesian causal inference framework to predict a cell's gene expressions under counterfactual perturbations (perturbations that this cell did not factually receive), leveraging information representing biological knowledge in the form of gene regulatory networks (GRNs) to aid individualized cellular response predictions. Aiming at a data-adaptive GRN, we also developed an adjacency matrix updating technique for graph convolutional networks and used it to refine GRNs during pre-training, which generated more insights on gene relations and enhanced model performance. Additionally, we propose a robust estimator within our framework for the asymptotically efficient estimation of marginal perturbation effect, which is yet to be carried out in previous works. With extensive experiments, we exhibited the advantage of our approach over state-of-the-art deep learning models for individual response prediction.

Steganography is the practice of encoding secret information into innocuous content in such a manner that an adversarial third party would not realize that there is hidden meaning. While this problem has classically been studied in security literature, recent advances in generative models have led to a shared interest among security and machine learning researchers in developing scalable steganography techniques. In this work, we show that a steganography procedure is perfectly secure under Cachin (1998)'s information theoretic-model of steganography if and only if it is induced by a coupling. Furthermore, we show that, among perfectly secure procedures, a procedure is maximally efficient if and only if it is induced by a minimum entropy coupling. These insights yield what are, to the best of our knowledge, the first steganography algorithms to achieve perfect security guarantees with non-trivial efficiency; additionally, these algorithms are highly scalable. To provide empirical validation, we compare a minimum entropy coupling-based approach to three modern baselines---arithmetic coding, Meteor, and adaptive dynamic grouping---using GPT-2, WaveRNN, and Image Transformer as communication channels. We find that the minimum entropy coupling-based approach achieves superior encoding efficiency, despite its stronger security constraints. In aggregate, these results suggest that it may be natural to view information-theoretic steganography through the lens of minimum entropy coupling.

Test-time adaptation (TTA) aims to adapt a trained classifier using online unlabeled test data only, without any information related to the training procedure. Most existing TTA methods adapt the trained classifier using the classifier's prediction on the test data as pseudo-label. However, under test-time domain shift, accuracy of the pseudo labels cannot be guaranteed, and thus the TTA methods often encounter performance degradation at the adapted classifier. To overcome this limitation, we propose a novel test-time adaptation method, called Test-time Adaptation via Self-Training with nearest neighbor information (TAST), which is composed of the following procedures: (1) adds trainable adaptation modules on top of the trained feature extractor; (2) newly defines a pseudo-label distribution for the test data by using the nearest neighbor information; (3) trains these modules only a few times during test time to match the nearest neighbor-based pseudo label distribution and a prototype-based class distribution for the test data; and (4) predicts the label of test data using the average predicted class distribution from these modules. The pseudo-label generation is based on the basic intuition that a test data and its nearest neighbor in the embedding space are likely to share the same label under the domain shift. By utilizing multiple randomly initialized adaptation modules, TAST extracts useful information for the classification of the test data under the domain shift, using the nearest neighbor information. TAST showed better performance than the state-of-the-art TTA methods on two standard benchmark tasks, domain generalization, namely VLCS, PACS, OfficeHome, and TerraIncognita, and image corruption, particularly CIFAR-10/100C.

Graph Neural Networks (GNNs) have shown great promise in various graph learning tasks. However, the computational overheads of fitting GNNs to large-scale graphs grow rapidly, posing obstacles to GNNs from scaling up to real-world applications. To tackle this issue, Graph Lottery Ticket (GLT) hypothesis articulates that there always exists a sparse subnetwork/subgraph with admirable performance in GNNs with random initialization. Such a pair of core subgraph and sparse subnetwork (called graph lottery tickets) can be uncovered by iteratively applying a novel sparsification method. While GLT provides new insights for GNN compression, it requires a full pretraining process to obtain graph lottery tickets, which is not universal and friendly to real-world applications. Moreover, the graph sparsification in GLT utilizes sampling techniques, which may result in massive information loss and aggregation failure. In this paper, we explore the searching of graph lottery tickets from a complementary perspective -- transforming a random ticket into a graph lottery ticket, which allows us to more comprehensively explore the relationships between the original network/graph and their sparse counterpart. To achieve this, we propose regularization-based network pruning and hierarchical graph sparsification, leading to our Dual Graph Lottery Ticket (DGLT) framework for a joint sparsification of network and graph. Compared to GLT, our DGLT helps achieve a triple-win situation of graph lottery tickets with high sparsity, admirable performance, and good explainability. More importantly, we rigorously prove that our model can eliminate noise and maintain reliable information in substructures using the graph information bottleneck theory. Extensive experimental results on various graph-related tasks validate the effectiveness of our framework.

Quality Diversity (QD) has emerged as a powerful alternative optimization paradigm that aims at generating large and diverse collections of solutions, notably with its flagship algorithm MAP-ELITES (ME) which evolves solutions through mutations and crossovers. While very effective for some unstructured problems, early ME implementations relied exclusively on random search to evolve the population of solutions, rendering them notoriously sample-inefficient for high-dimensional problems, such as when evolving neural networks. Follow-up works considered exploiting gradient information to guide the search in order to address these shortcomings through techniques borrowed from either Black-Box Optimization (BBO) or Reinforcement Learning (RL). While mixing RL techniques with ME unlocked state-of-the-art performance for robotics control problems that require a good amount of exploration, it also plagued these ME variants with limitations common among RL algorithms that ME was free of, such as hyperparameter sensitivity, high stochasticity as well as training instability, including when the population size increases as some components are shared across the population in recent approaches. Furthermore, existing approaches mixing ME with RL tend to be tied to a specific RL algorithm, which effectively prevents their use on problems where the corresponding RL algorithm fails. To address these shortcomings, we introduce a flexible framework that allows the use of any RL algorithm and alleviates the aforementioned limitations by evolving populations of agents (whose definition include hyperparameters and all learnable parameters) instead of just policies. We demonstrate the benefits brought about by our framework through extensive numerical experiments on a number of robotics control problems, some of which with deceptive rewards, taken from the QD-RL literature. We open source an efficient JAX-based implementation of our algorithm in the QDax library.

As federated learning (FL) matures, privacy attacks against FL systems in turn become more numerous and complex. Attacks on language models have progressed from recovering single sentences in simple classification tasks to recovering larger parts of user data. Current attacks against federated language models are sequence-agnostic and aim to extract as much data as possible from an FL update - often at the expense of fidelity for any particular sequence. Because of this, current attacks fail to extract any meaningful data under large-scale aggregation. In realistic settings, an attacker cares most about a small portion of user data that contains sensitive personal information, for example sequences containing the phrase "my credit card number is ...". In this work, we propose the first attack on FL that achieves targeted extraction of sequences that contain privacy-critical phrases, whereby we employ maliciously modified parameters to allow the transformer itself to filter relevant sequences from aggregated user data and encode them in the gradient update. Our attack can effectively extract sequences of interest even against extremely large-scale aggregation.

The brain effortlessly extracts latent causes of stimuli, but how it does this at the network level remains unknown. Most prior attempts at this problem proposed neural networks that implement independent component analysis, which works under the limitation that latent elements are mutually independent. Here, we relax this limitation and propose a biologically plausible neural network that extracts correlated latent sources by exploiting information about their domains. To derive this network, we choose maximum correlative information transfer from inputs to outputs as the separation objective under the constraint that the outputs are restricted to their presumed sets. The online formulation of this optimization problem naturally leads to neural networks with local learning rules. Our framework incorporates infinitely many source domain choices and flexibly models complex latent structures. Choices of simplex or polytopic source domains result in networks with piecewise-linear activation functions. We provide numerical examples to demonstrate the superior correlated source separation capability for both synthetic and natural sources.

Machine learning models exhibit two seemingly contradictory phenomena: training data memorization and various forms of forgetting. In memorization, models overfit specific training examples and become susceptible to privacy attacks. In forgetting, examples which appeared early in training are forgotten by the end. In this work, we connect these phenomena. We propose a technique to measure to what extent models ``forget'' the specifics of training examples, becoming less susceptible to privacy attacks on examples they have not seen recently. We show that, while non-convexity can prevent forgetting from happening in the worst-case, standard image,speech, and language models empirically do forget examples over time. We identify nondeterminism as a potential explanation, showing that deterministically trained models do not forget. Our results suggest that examples seen early when training with extremely large datasets---for instance those examples used to pre-train a model---may observe privacy benefits at the expense of examples seen later.