Orthogonality constraints are ubiquitous in robust and probabilistic machine learning. Unfortunately, current optimizers are computationally expensive and do not scale to problems with hundreds or thousands of constraints. One notable exception is the Landing algorithm (Ablin et al., 2024) which, however comes at the expense of temporarily relaxing orthogonality. In this work, we revisit and improve on the ideas behind Landing, enabling the inclusion of modern adaptive optimizers while ensuring that orthogonal constraints are effectively met. Remarkably, these improvements come at little to no cost, and reduce the number of required hyperparemeters. Our algorithm POGO is fast and GPU-friendly, , and in practice maintains orthogonality at all times. On several challenging benchmarks, POGO greatly outperforms recent optimizers and shows it can optimize problems with thousands of orthogonal matrices in minutes while alternatives would take hours. As such, POGO sets a milestone to finally exploit orthogonality constraints in ML at scale.

Squared tensor networks (TNs) and their extension as computational graphs---squared circuits---have been used as expressive distribution estimators, yet supporting closed-form marginalization. However, the squaring operation introduces additional complexity when computing the partition function or marginalizing variables, which hinders their applicability in ML. To solve this issue, canonical forms of TNs are parameterized via unitary matrices to simplify the computation of marginals. However, these canonical forms do not apply to circuits, as they can represent factorizations that do not directly map to a known TN. Inspired by the ideas of orthogonality in canonical forms and determinism in circuits enabling tractable maximization, we show how to parameterize squared circuits to overcome their marginalization overhead. Our parameterizations unlock efficient marginalization even in factorizations different from TNs, but encoded as circuits, whose structure would otherwise make marginalization computationally hard. Finally, our experiments on distribution estimation show how our proposed conditions in squared circuits come with no expressiveness loss, while enabling more efficient learning.

Tabular data is more challenging to generate than text and images, due to its heterogeneous features and much lower sample sizes. On this task, diffusion-based models are the current state-of-the-art (SotA) model class, achieving almost perfect performance on commonly used benchmarks. In this paper, we question the perception of progress for tabular data generation. First, we highlight the limitations of current protocols to evaluate the fidelity of generated data, and advocate for alternative ones. Next, we revisit a simple baseline – hierarchical mixture models in the form of deep probabilistic circuits (PCs) – which delivers competitive or superior performance to SotA models for a fraction of the cost. PCs are the generative counterpart of decision forests, and as such can natively handle heterogeneous data as well as deliver tractable probabilistic generation and inference. Finally, in a rigorous empirical analysis we show that the apparent saturation of progress for SotA models is largely due to the use of inadequate metrics. As such, we highlight that there is still much to be done to generate realistic tabular data. Code available at https://github.com/april-tools/tabpc.

In machine learning (ML), we often need to choose one among hundreds of trained ML models at hand, based on various objectives such as accuracy, robustness, fairness or scalability. However, it is often unclear how to compare, aggregate and, ultimately, trade-off these objectives, making it a time-consuming task that requires expert knowledge, as objectives may be measured in different units and scales. In this work, we investigate how objectives can be automatically normalized and aggregated to systematically help the user navigate their Pareto front. To this end, we make incomparable objectives comparable using their cumulative functions, approximated by their relative rankings. As a result, our proposed approach, COPA, can aggregate them while matching user-specific preferences, allowing practitioners to meaningfully navigate and search for models in the Pareto front. We demonstrate the potential impact of COPA in both model selection and benchmarking tasks across diverse ML areas such as fair ML, domain generalization, AutoML and foundation models, where classical ways to normalize and aggregate objectives fall short.

We introduce DeCaFlow, a deconfounding causal generative model. Training once per dataset using just observational data and the underlying causal graph, DeCaFlow enables accurate causal inference on continuous variables under the presence of hidden confounders. Specifically, we extend previous results on causal estimation under hidden confounding to show that a single instance of DeCaFlow provides correct estimates for all causal queries identifiable with do-calculus, leveraging proxy variables to adjust for the causal effects when do-calculus alone is insufficient. Moreover, we show that counterfactual queries are identifiable as long as their interventional counterparts are identifiable, and thus are also correctly estimated by DeCaFlow. Our empirical results on diverse settings—including the Ecoli70 dataset, with 3 independent hidden confounders, tens of observed variables and hundreds of causal queries—show that DeCaFlow outperforms existing approaches, while demonstrating its out-of-the-box applicability to any given causal graph.

In machine learning (ML), it is common to account for multiple objectives when, e.g., selecting a model to deploy. However, it is often unclear how one should compare, aggregate and, ultimately, trade-off these objectives, as they might be measured in different units or scales. For example, when deploying large language models (LLMs), we might not only care about their performance, but also their CO2 consumption. In this work, we investigate how objectives can be sensibly compared and aggregated to navigate their Pareto front. To do so, we propose to make incomparable objectives comparable via their CDFs, approximated by their relative rankings. This allows us to aggregate them while matching user-specific preferences, allowing practitioners to meaningfully navigate and search for models in the Pareto front. We demonstrate the potential impact of our methodology in diverse areas such as LLM selection, domain generalization, and AutoML benchmarking, where classical ways to aggregate and normalize objectives fail.

In this work, we deepen on the use of normalizing flows for causal reasoning. Specifically, we first leverage recent results on non-linear ICA to show that causal models are identifiable from observational data given a causal ordering, and thus can be recovered using autoregressive normalizing flows (NFs). Second, we analyze different design and learning choices for *causal normalizing flows* to capture the underlying causal data-generating process. Third, we describe how to implement the *do-operator* in causal NFs, and thus, how to answer interventional and counterfactual questions. Finally, in our experiments, we validate our design and training choices through a comprehensive ablation study; compare causal NFs to other approaches for approximating causal models; and empirically demonstrate that causal NFs can be used to address real-world problems—where the presence of mixed discrete-continuous data and partial knowledge on the causal graph is the norm. The code for this work can be found at https://github.com/psanch21/causal-flows.

Existing Graph Neural Networks (GNNs) compute the message exchange between nodes by either aggregating uniformly (convolving) the features of all the neighboring nodes, or by applying a non-uniform score (attending) to the features. Recent works have shown the strengths and weaknesses of the resulting GNN architectures, respectively, GCNs and GATs. In this work, we aim at exploiting the strengths of both approaches to their full extent. To this end, we first introduce the graph convolutional attention layer (CAT), which relies on convolutions to compute the attention scores. Unfortunately, as in the case of GCNs and GATs, we show that there exists no clear winner between the three (neither theoretically nor in practice) as their performance directly depends on the nature of the data (i.e., of the graph and features). This result brings us to the main contribution of our work, the learnable graph convolutional attention network (L-CAT): a GNN architecture that automatically interpolates between GCN, GAT and CAT in each layer, by adding only two scalar parameters. Our results demonstrate that L-CAT is able to efficiently combine different GNN layers along the network, outperforming competing methods in a wide range of datasets, and resulting in a more robust model that reduces the need of cross-validating.

A number of variational autoencoders (VAEs) have recently emerged with the aim of modeling multimodal data, e.g., to jointly model images and their corresponding captions. Still, multimodal VAEs tend to focus solely on a subset of the modalities, e.g., by fitting the image while neglecting the caption. We refer to this limitation as modality collapse. In this work, we argue that this effect is a consequence of conflicting gradients during multimodal VAE training. We show how to detect the sub-graphs in the computational graphs where gradients conflict (impartiality blocks), as well as how to leverage existing gradient-conflict solutions from multitask learning to mitigate modality collapse. That is, to ensure impartial optimization across modalities. We apply our training framework to several multimodal VAE models, losses and datasets from the literature, and empirically show that our framework significantly improves the reconstruction performance, conditional generation, and coherence of the latent space across modalities.

GradNorm (Chen et al., 2018) is a broadly used gradient-based approach for training multitask networks, where different tasks share, and thus compete during learning, for the network parameters. GradNorm eases the fitting of all individual tasks by dynamically equalizing the contribution of each task to the overall gradient magnitude. However, it does not prevent the individual tasks’ gradients from conflicting, i.e., pointing towards opposite directions, and thus resulting in a poor multitask performance. In this work we propose Rotograd, an extension to GradNorm that addresses this problem by dynamically homogenizing not only the gradient magnitudes but also their directions across tasks. For this purpose,Rotograd adds a layer of task-specific rotation matrices that aligns all the task gradients. Importantly, we then analyze Rotograd (and its predecessor) through the lens of game theory, providing theoretical guarantees on the algorithm stability and convergence. Finally, our experiments on several real-world datasets and network architectures show that Rotograd outperforms previous approaches for multitask learning.

Multitask learning is being increasingly adopted in applications domains like computer vision and reinforcement learning. However, optimally exploiting its advantages remains a major challenge due to the effect of negative transfer. Previous works have tracked down this issue to the disparities in gradient magnitudes and directions across tasks, when optimizing the shared network parameters. While recent work has acknowledged that negative transfer is a two-fold problem, existing approaches fall short as they only focus on either homogenizing the gradient magnitude across tasks; or greedily change the gradient directions, overlooking future conflicts. In this work, we introduce RotoGrad, an algorithm that tackles negative transfer as a whole: it jointly homogenizes gradient magnitudes and directions, while ensuring training convergence. We show that RotoGrad outperforms competing methods in complex problems, including multi-label classification in CelebA and computer vision tasks in the NYUv2 dataset. A Pytorch implementation can be found in https://github.com/adrianjav/rotograd.

Deep learning methods are gaining popularity in different application domains, and especially in natural language processing. It is commonly believed that using a large enough dataset and an adequate network architecture, almost any processing problem can be solved. A frequent and widely used typology is the encoder-decoder architecture, where the input data is transformed into an intermediate code by means of an encoder, and then a decoder takes this code to produce its output. Different types of networks can be used in the encoder and the decoder, depending on the problem of interest, such as convolutional neural networks (CNN) or long-short term memories (LSTM). This paper uses for the encoder a method recently proposed, called Causal Feature Extractor (CFE). It is based on causal convolutions (i.e., convolutions that depend only on one direction of the input), dilatation (i.e., increasing the aperture size of the convolutions) and bidirectionality (i.e., independent networks in both directions). Some preliminary results are presented on three different tasks and compared with state-of-the-art methods: bilingual translation, LaTeX decompilation and audio transcription. The proposed method achieves promising results, showing its ubiquity to work with text, audio and images. Moreover, it has a shorter training time, requiring less time per iteration, and a good use of the attention mechanisms based on attention matrices.

Learning expressive probabilistic models correctly describing the data is a ubiquitous problem in machine learning. A popular approach for solving it is mapping the observations into a representation space with a simple joint distribution, which can typically be written as a product of its marginals — thus drawing a connection with the field of nonlinear independent component analysis. Deep density models have been widely used for this task, but their maximum likelihood based training requires estimating the log-determinant of the Jacobian and is computationally expensive, thus imposing a trade-off between computation and expressive power. In this work, we propose a new approach for exact training of such neural networks. Based on relative gradients, we exploit the matrix structure of neural network parameters to compute updates efficiently even in high-dimensional spaces; the computational cost of the training is quadratic in the input size, in contrast with the cubic scaling of naive approaches. This allows fast training with objective functions involving the log-determinant of the Jacobian, without imposing constraints on its structure, in stark contrast to autoregressive normalizing flows.

The encoder–decoder architecture is a well-established, effective and widely used approach in many tasks of natural language processing (NLP), among other domains. It consists of two closely-collaborating components: An encoder that transforms the input into an intermediate form, and a decoder producing the output. This paper proposes a new method for the encoder, named Causal Feature Extractor (CFE), based on three main ideas: Causal convolutions, dilatations and bidirectionality. We apply this method to text normalization, which is a ubiquitous problem that appears as the first step of many text-to-speech (TTS) systems. Given a text with symbols, the problem consists in writing the text exactly as it should be read by the TTS system. We make use of an attention-based encoder–decoder architecture using a fine-grained character-level approach rather than the usual word-level one. The proposed CFE is compared to other common encoders, such as convolutional neural networks (CNN) and long-short term memories (LSTM). Experimental results show the feasibility of CFE, achieving better results in terms of accuracy, number of parameters, convergence time, and use of an attention mechanism based on attention matrices. The obtained accuracy ranges from 83.5% to 96.8% correctly normalized sentences, depending on the dataset. Moreover, the proposed method is generic and can be applied to different types of input such as text, audio and images.

Probabilistic learning is increasingly being tackled as an optimization problem, with gradient-based approaches as predominant methods. When modelling multivariate likelihoods, a usual but undesirable outcome is that the learned model fits only a subset of the observed variables, overlooking the rest. In this work, we study this problem through the lens of multitask learning (MTL), where similar effects have been broadly studied. While MTL solutions do not directly apply in the probabilistic setting (as they cannot handle the likelihood constraints) we show that similar ideas may be leveraged during data preprocessing. First, we show that data standardization often helps under common continuous likelihoods, but it is not enough in the general case, specially under mixed continuous and discrete likelihood models. In order for balance multivariate learning, we then propose a novel data preprocessing, Lipschitz standardization, which balances the local Lipschitz smoothness across variables. Our experiments on real-world datasets show that Lipschitz standardization leads to more accurate multivariate models than the ones learned using existing data preprocessing techniques. The models and datasets employed in the experiments can be found in https://github.com/adrianjav/lipschitz-standardization.