Recent years have seen growing interest in learning disentangled representations, in which distinct features, such as size or shape, are represented by distinct neurons. Quantifying the extent to which a given representation is disentangled is not straightforward; multiple metrics have been proposed. In this paper, we identify two failings of existing metrics, which mean they can assign a high score to a model which is still entangled, and we propose two new metrics, which redress these problems. We then consider the task of compositional generalization. Unlike prior works, we treat this as a classification problem, which allows us to use it to measure the disentanglement ability of the encoder, without depending on the decoder. We show that performance on this task is (a) generally quite poor, (b) correlated with most disentanglement metrics, and (c) most strongly correlated with our newly proposed metrics.

Deep learning has redefined the field of artificial intelligence (AI) thanks to the rise of artificial neural networks, which are architectures inspired by their neurological counterpart in the brain. Through the years, this dualism between AI and neuroscience has brought immense benefits to both fields, allowing neural networks to be used in dozens of applications. These networks use an efficient implementation of reverse differentiation, called backpropagation (BP). This algorithm, however, is often criticized for its biological implausibility (e.g., lack of local update rules for the parameters). Therefore, biologically plausible learning methods that rely on predictive coding (PC), a framework for describing information processing in the brain, are increasingly studied. Recent works prove that these methods can approximate BP up to a certain margin on multilayer perceptrons (MLPs), and asymptotically on any other complex model, and that zero-divergence inference learning (Z-IL), a variant of PC, is able to exactly implement BP on MLPs. However, the recent literature shows also that there is no biologically plausible method yet that can exactly replicate the weight update of BP on complex models. To fill this gap, in this paper, we generalize (PC and) Z-IL by directly defining them on computational graphs, and show that it can perform exact reverse differentiation. What results is the first biologically plausible algorithm that is equivalent to BP in the way of updating parameters on any neural network, providing a bridge between the interdisciplinary research of neuroscience and deep learning.

Bias-measuring datasets play a critical role in detecting biased behavior of language models and in evaluating progress of bias mitigation methods. In this work, we focus on evaluating gender bias through coreference resolution, where previous datasets are either hand-crafted or fail to reliably measure an explicitly defined bias. To overcome these shortcomings, we propose a novel method to collect diverse, natural, and minimally distant text pairs via counterfactual generation, and construct Counter-GAP, an annotated dataset consisting of 4008 instances grouped into 1002 quadruples. We further identify a bias cancellation problem in previous group-level metrics on Counter-GAP, and propose to use the difference between inconsistency across genders and within genders to measure bias at a quadruple level. Our results show that four pre-trained language models are significantly more inconsistent across different gender groups than within each group, and that a name-based counterfactual data augmentation method is more effective to mitigate such bias than an anonymization-based method.

The Winograd Schema Challenge - a set of twin sentences involving pronoun reference disambiguation that seem to require the use of commonsense knowledge - was proposed by Hector Levesque in 2011. By 2019, a number of AI systems, based on large pre-trained transformer-based language models and fine-tuned on these kinds of problems, achieved better than 90% accuracy. In this paper, we review the history of the Winograd Schema Challenge and discuss the lasting contributions of the flurry of research that has taken place on the WSC in the last decade. We discuss the significance of various datasets developed for WSC, and the research community's deeper understanding of the role of surrogate tasks in assessing the intelligence of an AI system.

Explaining the predictions of AI models is paramount in safety-critical applications, such as in legal or medical domains. One form of explanation for a prediction is an extractive rationale, i.e., a subset of features of an instance that lead the model to give its prediction on that instance. For example, the subphrase ``he stole the mobile phone'' can be an extractive rationale for the prediction of ``Theft''. Previous works on generating extractive rationales usually employ a two-phase model: a selector that selects the most important features (i.e., the rationale) followed by a predictor that makes the prediction based exclusively on the selected features. One disadvantage of these works is that the main signal for learning to select features comes from the comparison of the answers given by the predictor to the ground-truth answers. In this work, we propose to squeeze more information from the predictor via an information calibration method. More precisely, we train two models jointly: one is a typical neural model that solves the task at hand in an accurate but black-box manner, and the other is a selector-predictor model that additionally produces a rationale for its prediction. The first model is used as a guide for the second model. We use an adversarial technique to calibrate the information extracted by the two models such that the difference between them is an indicator of the missed or over-selected features. In addition, for natural language tasks, we propose a language-model-based regularizer to encourage the extraction of fluent rationales. Experimental results on a sentiment analysis task, a hate speech recognition task as well as on three tasks from the legal domain show the effectiveness of our approach to rationale extraction.

Predictive coding is a message-passing framework initially developed to model information processing in the brain, and now also topic of research in machine learning due to some interesting properties. One of such properties is the natural ability of generative models to learn robust representations thanks to their peculiar credit assignment rule, that allows neural activities to converge to a solution before updating the synaptic weights. Graph neural networks are also message-passing models, which have recently shown outstanding results in diverse types of tasks in machine learning, providing interdisciplinary state-of-the-art performance on structured data. However, they are vulnerable to imperceptible adversarial attacks, and unfit for out-of-distribution generalization. In this work, we address this by building models that have the same structure of popular graph neural network architectures, but rely on the message-passing rule of predictive coding. Through an extensive set of experiments, we show that the proposed models are (i) comparable to standard ones in terms of performance in both inductive and transductive tasks, (ii) better calibrated, and (iii) robust against multiple kinds of adversarial attacks.

In recent years, deep learning has reached and surpassed human-level performance in a multitude of tasks, such as game playing [Silver et al., 2017, 2016], image recognition [Krizhevsky et al., 2012, He et al., 2016], natural language processing [Chen et al., 2020], and image generation [Ramesh et al., 2022]. These successes are achieved entirely using deep artificial neural networks trained via backpropagation (BP), which is a learning algorithm that is often criticized for its biological implausibilities [Grossberg, 1987, Crick, 1989, Abdelghani et al., 2008, Lillicrap et al., 2016, Roelfsema and Holtmaat, 2018, Whittington and Bogacz, 2019], such as lacking local plasticity and autonomy. In fact, backpropagation requires a global control signal required to trigger computations, since gradients must be sequentially computed backwards through the computation graph. These properties are not only important for biological plausibility: parallelization, locality, and automation are key to build efficient models that can be trained end-to-end on non Von-Neumann machines, such as analog chips [Kendall et al., 2020]. A learning algorithm with most of the above properties is predictive coding (PC). PC is an influential theory of information processing in the brain [Mumford, 1992, Friston, 2005], where learning happens by minimizing the prediction error of every neuron. PC can be shown to approximate backpropagation in layered networks [Whittington and Bogacz, 2017], as well as on any other model [Millidge et al., 2020], and can exactly replicate its weight update if some external control is added [Salvatori et al., 2022a]. Also the differences with BP are interesting, as PC allows for a much more flexible training and testing [Salvatori et al., 2022b], has a rich mathematical formulation [Friston, 2005, Millidge et al., 2022], and is an energy-based model [Bogacz, 2017].

A large amount of recent research has the far-reaching goal of finding training methods for deep neural networks that can serve as alternatives to backpropagation (BP). A prominent example is predictive coding (PC), which is a neuroscience-inspired method that performs inference on hierarchical Gaussian generative models. These methods, however, fail to keep up with modern neural networks, as they are unable to replicate the dynamics of complex layers and activation functions. In this work, we solve this problem by generalizing PC to arbitrary probability distributions, enabling the training of architectures, such as transformers, that are hard to approximate with only Gaussian assumptions. We perform three experimental analyses. First, we study the gap between our method and the standard formulation of PC on multiple toy examples. Second, we test the reconstruction quality on variational autoencoders, where our method reaches the same reconstruction quality as BP. Third, we show that our method allows us to train transformer networks and achieve a performance comparable with BP on conditional language models. More broadly, this method allows neuroscience-inspired learning to be applied to multiple domains, since the internal distributions can be flexibly adapted to the data, tasks, and architectures used.

Neural networks have proven to be very powerful at computer vision tasks. However, they often exhibit unexpected behaviours, violating known requirements expressing background knowledge. This calls for models (i) able to learn from the requirements, and (ii) guaranteed to be compliant with the requirements themselves. Unfortunately, the development of such models is hampered by the lack of datasets equipped with formally specified requirements. In this paper, we introduce the ROad event Awareness Dataset with logical Requirements (ROAD-R), the first publicly available dataset for autonomous driving with requirements expressed as logical constraints. Given ROAD-R, we show that current state-of-the-art models often violate its logical constraints, and that it is possible to exploit them to create models that (i) have a better performance, and (ii) are guaranteed to be compliant with the requirements themselves.

Predictive coding (PC) is an influential theory in computational neuroscience, which argues that the cortex forms unsupervised world models by implementing a hierarchical process of prediction error minimization. PC networks (PCNs) are trained in two phases. First, neural activities are updated to optimize the network's response to external stimuli. Second, synaptic weights are updated to consolidate this change in activity -- an algorithm called \emph{prospective configuration}. While previous work has shown how in various limits, PCNs can be found to approximate backpropagation (BP), recent work has demonstrated that PCNs operating in this standard regime, which does not approximate BP, nevertheless obtain competitive training and generalization performance to BP-trained networks while outperforming them on tasks such as online, few-shot, and continual learning, where brains are known to excel. Despite this promising empirical performance, little is understood theoretically about the properties and dynamics of PCNs in this regime. In this paper, we provide a comprehensive theoretical analysis of the properties of PCNs trained with prospective configuration. We first derive analytical results concerning the inference equilibrium for PCNs and a previously unknown close connection relationship to target propagation (TP). Secondly, we provide a theoretical analysis of learning in PCNs as a variant of generalized expectation-maximization and use that to prove the convergence of PCNs to critical points of the BP loss function, thus showing that deep PCNs can, in theory, achieve the same generalization performance as BP, while maintaining their unique advantages.