Stefano Favaro

Alberto Bordino, Stefano Favaro, Sandra Fortini

There is a recent and growing literature on large-width asymptotic and non-asymptotic properties of deep Gaussian neural networks (NNs), namely NNs with weights initialized as Gaussian distributions. For a Gaussian NN of depth $L\geq1$ and width $n\geq1$, it is well-known that, as $n\rightarrow+\infty$, the NN's output converges (in distribution) to a Gaussian process. Recently, some quantitative versions of this result, also known as quantitative central limit theorems (QCLTs), have been obtained, showing that the rate of convergence is $n^{-1}$, in the $2$-Wasserstein distance, and that such a rate is optimal. In this paper, we investigate the use of second-order Poincaré inequalities as an alternative approach to establish QCLTs for the NN's output. Previous approaches consist of a careful analysis of the NN, by combining non-trivial probabilistic tools with ad-hoc techniques that rely on the recursive definition of the network, typically by means of an induction argument over the layers, and it is unclear if and how they still apply to other NN's architectures. Instead, the use of second-order Poincaré inequalities rely only on the fact that the NN is a functional of a Gaussian process, reducing the problem of establishing QCLTs to the algebraic problem of computing the gradient and Hessian of the NN's output, which still applies to other NN's architectures. We show how our approach is effective in establishing QCLTs for the NN's output, though it leads to suboptimal rates of convergence. We argue that such a worsening in the rates is peculiar to second-order Poincaré inequalities, and it should be interpreted as the "cost" for having a straightforward, and general, procedure for obtaining QCLTs.

Alberto Bordino, Stefano Favaro, Sandra Fortini

There is a growing literature on the study of large-width properties of deep Gaussian neural networks (NNs), i.e. deep NNs with Gaussian-distributed parameters or weights, and Gaussian stochastic processes. Motivated by some empirical and theoretical studies showing the potential of replacing Gaussian distributions with Stable distributions, namely distributions with heavy tails, in this paper we investigate large-width properties of deep Stable NNs, i.e. deep NNs with Stable-distributed parameters. For sub-linear activation functions, a recent work has characterized the infinitely wide limit of a suitable rescaled deep Stable NN in terms of a Stable stochastic process, both under the assumption of a ``joint growth" and under the assumption of a ``sequential growth" of the width over the NN's layers. Here, assuming a ``sequential growth" of the width, we extend such a characterization to a general class of activation functions, which includes sub-linear, asymptotically linear and super-linear functions. As a novelty with respect to previous works, our results rely on the use of a generalized central limit theorem for heavy tails distributions, which allows for an interesting unified treatment of infinitely wide limits for deep Stable NNs. Our study shows that the scaling of Stable NNs and the stability of their infinitely wide limits may depend on the choice of the activation function, bringing out a critical difference with respect to the Gaussian setting.

Stefano Favaro, Sandra Fortini, Stefano Peluchetti

There is a recent and growing literature on large-width asymptotic properties of Gaussian neural networks (NNs), namely NNs whose weights are initialized as Gaussian distributions. Two popular problems are: i) the study of the large-width distributions of NNs, which characterizes the infinitely wide limit of a rescaled NN in terms of a Gaussian stochastic process; ii) the study of the large-width training dynamics of NNs, which characterizes the infinitely wide dynamics in terms of a deterministic kernel, referred to as the neural tangent kernel (NTK), and shows that, for a sufficiently large width, the gradient descent achieves zero training error at a linear rate. In this paper, we consider these problems for $α$-Stable NNs, namely NNs whose weights are initialized as $α$-Stable distributions with $α\in(0,2]$. First, for $α$-Stable NNs with a ReLU activation function, we show that if the NN's width goes to infinity then a rescaled NN converges weakly to an $α$-Stable stochastic process. As a difference with respect to the Gaussian setting, our result shows that the choice of the activation function affects the scaling of the NN, that is: to achieve the infinitely wide $α$-Stable process, the ReLU activation requires an additional logarithmic term in the scaling with respect to sub-linear activations. Then, we study the large-width training dynamics of $α$-Stable ReLU-NNs, characterizing the infinitely wide dynamics in terms of a random kernel, referred to as the $α$-Stable NTK, and showing that, for a sufficiently large width, the gradient descent achieves zero training error at a linear rate. The randomness of the $α$-Stable NTK is a further difference with respect to the Gaussian setting, that is: within the $α$-Stable setting, the randomness of the NN at initialization does not vanish in the large-width regime of the training.

Lucia Pezzetti, Stefano Favaro, Stefano Peluchetti

Bayesian Neural Networks represent a fascinating confluence of deep learning and probabilistic reasoning, offering a compelling framework for understanding uncertainty in complex predictive models. In this paper, we investigate the use of the preconditioned Crank-Nicolson algorithm and its Langevin version to sample from a reparametrised posterior distribution of the neural network's weights, as the widths grow larger. In addition to being robust in the infinite-dimensional setting, we prove that the acceptance probabilities of the proposed algorithms approach 1 as the width of the network increases, independently of any stepsize tuning. Moreover, we examine and compare how the mixing speeds of the underdamped Langevin Monte Carlo, the preconditioned Crank-Nicolson and the preconditioned Crank-Nicolson Langevin samplers are influenced by changes in the network width in some real-world cases. Our findings suggest that, in wide Bayesian Neural Networks configurations, the preconditioned Crank-Nicolson algorithm allows for a scalable and more efficient sampling of the reparametrised posterior distribution, as also evidenced by a higher effective sample size and improved diagnostic results compared with the other analysed algorithms.

Stefano Favaro, Boris Hanin, Domenico Marinucci et al.

We study the distribution of a fully connected neural network with random Gaussian weights and biases in which the hidden layer widths are proportional to a large constant $n$. Under mild assumptions on the non-linearity, we obtain quantitative bounds on normal approximations valid at large but finite $n$ and any fixed network depth. Our theorems show both for the finite-dimensional distributions and the entire process, that the distance between a random fully connected network (and its derivatives) to the corresponding infinite width Gaussian process scales like $n^{-γ}$ for $γ>0$, with the exponent depending on the metric used to measure discrepancy. Our bounds are strictly stronger in terms of their dependence on network width than any previously available in the literature; in the one-dimensional case, we also prove that they are optimal, i.e., we establish matching lower bounds.

Tianmin Xie, Yanfei Zhou, Ziyi Liang et al.

This paper presents a conformal prediction method for classification in highly imbalanced and open-set settings, where there are many possible classes and not all may be represented in the data. Existing approaches require a finite, known label space and typically involve random sample splitting, which works well when there is a sufficient number of observations from each class. Consequently, they have two limitations: (i) they fail to provide adequate coverage when encountering new labels at test time, and (ii) they may become overly conservative when predicting previously seen labels. To obtain valid prediction sets in the presence of unseen labels, we compute and integrate into our predictions a new family of conformal p-values that can test whether a new data point belongs to a previously unseen class. We study these p-values theoretically, establishing their optimality, and uncover an intriguing connection with the classical Good--Turing estimator for the probability of observing a new species. To make more efficient use of imbalanced data, we also develop a selective sample splitting algorithm that partitions training and calibration data based on label frequency, leading to more informative predictions. Despite breaking exchangeability, this allows maintaining finite-sample guarantees through suitable re-weighting. With both simulated and real data, we demonstrate our method leads to prediction sets with valid coverage even in challenging open-set scenarios with infinite numbers of possible labels, and produces more informative predictions under extreme class imbalance.

Francesco Caporali, Stefano Favaro, Dario Trevisan

The asymptotic properties of Bayesian Neural Networks (BNNs) have been extensively studied, particularly regarding their approximations by Gaussian processes in the infinite-width limit. We extend these results by showing that posterior BNNs can be approximated by Student-t processes, which offer greater flexibility in modeling uncertainty. Specifically, we show that, if the parameters of a BNN follow a Gaussian prior distribution, and the variance of both the last hidden layer and the Gaussian likelihood function follows an Inverse-Gamma prior distribution, then the resulting posterior BNN converges to a Student-t process in the infinite-width limit. Our proof leverages the Wasserstein metric to establish control over the convergence rate of the Student-t process approximation.

Claudia Contardi, Emanuele Dolera, Stefano Favaro

The unseen-species problem assumes $n\geq1$ samples from a population of individuals belonging to different species, possibly infinite, and calls for estimating the number $K_{n,m}$ of hitherto unseen species that would be observed if $m\geq1$ new samples were collected from the same population. This is a long-standing problem in statistics, which has gained renewed relevance in biological and physical sciences, particularly in settings with large values of $n$ and $m$. In this paper, we adopt a Bayesian nonparametric approach to the unseen-species problem under the Pitman-Yor prior, and propose a novel methodology to derive large $m$ asymptotic credible intervals for $K_{n,m}$, for any $n\geq1$. By leveraging a Gaussian central limit theorem for the posterior distribution of $K_{n,m}$, our method improves upon competitors in two key aspects: firstly, it enables the full parameterization of the Pitman-Yor prior, including the Dirichlet prior; secondly, it avoids the need of Monte Carlo sampling, enhancing computational efficiency. We validate the proposed method on synthetic and real data, demonstrating that it improves the empirical performance of competitors by significantly narrowing the gap between asymptotic and exact credible intervals for any $m\geq1$.

Lorenzo Masoero, Mario Beraha, Thomas Richardson et al.

In online randomized experiments or A/B tests, accurate predictions of participant inclusion rates are of paramount importance. These predictions not only guide experimenters in optimizing the experiment's duration but also enhance the precision of treatment effect estimates. In this paper we present a novel, straightforward, and scalable Bayesian nonparametric approach for predicting the rate at which individuals will be exposed to interventions within the realm of online A/B testing. Our approach stands out by offering dual prediction capabilities: it forecasts both the quantity of new customers expected in future time windows and, unlike available alternative methods, the number of times they will be observed. We derive closed-form expressions for the posterior distributions of the quantities needed to form predictions about future user activity, thereby bypassing the need for numerical algorithms such as Markov chain Monte Carlo. After a comprehensive exposition of our model, we test its performance on experiments on real and simulated data, where we show its superior performance with respect to existing alternatives in the literature.

Mario Beraha, Lorenzo Masoero, Stefano Favaro et al.

Accurately predicting the onset of specific activities within defined timeframes holds significant importance in several applied contexts. In particular, accurate prediction of the number of future users that will be exposed to an intervention is an important piece of information for experimenters running online experiments (A/B tests). In this work, we propose a novel approach to predict the number of users that will be active in a given time period, as well as the temporal trajectory needed to attain a desired user participation threshold. We model user activity using a Bayesian nonparametric approach which allows us to capture the underlying heterogeneity in user engagement. We derive closed-form expressions for the number of new users expected in a given period, and a simple Monte Carlo algorithm targeting the posterior distribution of the number of days needed to attain a desired number of users; the latter is important for experimental planning. We illustrate the performance of our approach via several experiments on synthetic and real world data, in which we show that our novel method outperforms existing competitors.

Emanuele Dolera, Stefano Favaro

Privacy-protecting data analysis investigates statistical methods under privacy constraints. This is a rising challenge in modern statistics, as the achievement of confidentiality guarantees, which typically occurs through suitable perturbations of the data, may determine a loss in the statistical utility of the data. In this paper, we consider privacy-protecting tests for goodness-of-fit in frequency tables, this being arguably the most common form of releasing data, and present a rigorous analysis of the large sample behaviour of a private likelihood-ratio (LR) test. Under the framework of $(\varepsilon,δ)$-differential privacy for perturbed data, our main contribution is the power analysis of the private LR test, which characterizes the trade-off between confidentiality, measured via the differential privacy parameters $(\varepsilon,δ)$, and statistical utility, measured via the power of the test. This is obtained through a Bahadur-Rao large deviation expansion for the power of the private LR test, bringing out a critical quantity, as a function of the sample size, the dimension of the table and $(\varepsilon,δ)$, that determines a loss in the power of the test. Such a result is then applied to characterize the impact of the sample size and the dimension of the table, in connection with the parameters $(\varepsilon,δ)$, on the loss of the power of the private LR test. In particular, we determine the (sample) cost of $(\varepsilon,δ)$-differential privacy in the private LR test, namely the additional sample size that is required to recover the power of the Multinomial LR test in the absence of perturbation. Our power analysis rely on a non-standard large deviation analysis for the LR, as well as the development of a novel (sharp) large deviation principle for sum of i.i.d. random vectors, which is of independent interest.

Stefano Favaro, Sandra Fortini, Stefano Peluchetti

In modern deep learning, there is a recent and growing literature on the interplay between large-width asymptotic properties of deep Gaussian neural networks (NNs), i.e. deep NNs with Gaussian-distributed weights, and Gaussian stochastic processes (SPs). Such an interplay has proved to be critical in Bayesian inference under Gaussian SP priors, kernel regression for infinitely wide deep NNs trained via gradient descent, and information propagation within infinitely wide NNs. Motivated by empirical analyses that show the potential of replacing Gaussian distributions with Stable distributions for the NN's weights, in this paper we present a rigorous analysis of the large-width asymptotic behaviour of (fully connected) feed-forward deep Stable NNs, i.e. deep NNs with Stable-distributed weights. We show that as the width goes to infinity jointly over the NN's layers, i.e. the ``joint growth" setting, a rescaled deep Stable NN converges weakly to a Stable SP whose distribution is characterized recursively through the NN's layers. Because of the non-triangular structure of the NN, this is a non-standard asymptotic problem, to which we propose an inductive approach of independent interest. Then, we establish sup-norm convergence rates of the rescaled deep Stable NN to the Stable SP, under the ``joint growth" and a ``sequential growth" of the width over the NN's layers. Such a result provides the difference between the ``joint growth" and the ``sequential growth" settings, showing that the former leads to a slower rate than the latter, depending on the depth of the layer and the number of inputs of the NN. Our work extends some recent results on infinitely wide limits for deep Gaussian NNs to the more general deep Stable NNs, providing the first result on convergence rates in the ``joint growth" setting.

Daniele Bracale, Stefano Favaro, Sandra Fortini et al.

In this paper, we consider fully connected feed-forward deep neural networks where weights and biases are independent and identically distributed according to Gaussian distributions. Extending previous results (Matthews et al., 2018a;b; Yang, 2019) we adopt a function-space perspective, i.e. we look at neural networks as infinite-dimensional random elements on the input space $\mathbb{R}^I$. Under suitable assumptions on the activation function we show that: i) a network defines a continuous Gaussian process on the input space $\mathbb{R}^I$; ii) a network with re-scaled weights converges weakly to a continuous Gaussian process in the large-width limit; iii) the limiting Gaussian process has almost surely locally $γ$-Hölder continuous paths, for $0 < γ<1$. Our results contribute to recent theoretical studies on the interplay between infinitely wide deep neural networks and Gaussian processes by establishing weak convergence in function-space with respect to a stronger metric.

Emanuele Dolera, Stefano Favaro, Stefano Peluchetti

The count-min sketch (CMS) is a time and memory efficient randomized data structure that provides estimates of tokens' frequencies in a data stream of tokens, i.e. point queries, based on random hashed data. A learning-augmented version of the CMS, referred to as CMS-DP, has been proposed by Cai, Mitzenmacher and Adams (\textit{NeurIPS} 2018), and it relies on Bayesian nonparametric (BNP) modeling of the data stream of tokens via a Dirichlet process (DP) prior, with estimates of a point query being obtained as suitable mean functionals of the posterior distribution of the point query, given the hashed data. While the CMS-DP has proved to improve on some aspects of CMS, it has the major drawback of arising from a ``constructive" proof that builds upon arguments tailored to the DP prior, namely arguments that are not usable for other nonparametric priors. In this paper, we present a ``Bayesian" proof of the CMS-DP that has the main advantage of building upon arguments that are usable, in principle, within a broad class of nonparametric priors arising from normalized completely random measures. This result leads to develop a novel learning-augmented CMS under power-law data streams, referred to as CMS-PYP, which relies on BNP modeling of the data stream of tokens via a Pitman-Yor process (PYP) prior. Under this more general framework, we apply the arguments of the ``Bayesian" proof of the CMS-DP, suitably adapted to the PYP prior, in order to compute the posterior distribution of a point query, given the hashed data. Applications to synthetic data and real textual data show that the CMS-PYP outperforms the CMS and the CMS-DP in estimating low-frequency tokens, which are known to be of critical interest in textual data, and it is competitive with respect to a variation of the CMS designed for low-frequency tokens. An extension of our BNP approach to more general queries is also discussed.

Emanuele Dolera, Stefano Favaro, Stefano Peluchetti

The count-min sketch (CMS) is a randomized data structure that provides estimates of tokens' frequencies in a large data stream using a compressed representation of the data by random hashing. In this paper, we rely on a recent Bayesian nonparametric (BNP) view on the CMS to develop a novel learning-augmented CMS under power-law data streams. We assume that tokens in the stream are drawn from an unknown discrete distribution, which is endowed with a normalized inverse Gaussian process (NIGP) prior. Then, using distributional properties of the NIGP, we compute the posterior distribution of a token's frequency in the stream, given the hashed data, and in turn corresponding BNP estimates. Applications to synthetic and real data show that our approach achieves a remarkable performance in the estimation of low-frequency tokens. This is known to be a desirable feature in the context of natural language processing, where it is indeed common in the context of the power-law behaviour of the data.

Daniele Bracale, Stefano Favaro, Sandra Fortini et al.

The interplay between infinite-width neural networks (NNs) and classes of Gaussian processes (GPs) is well known since the seminal work of Neal (1996). While numerous theoretical refinements have been proposed in the recent years, the interplay between NNs and GPs relies on two critical distributional assumptions on the NN's parameters: A1) finite variance; A2) independent and identical distribution (iid). In this paper, we consider the problem of removing A1 in the general context of deep feed-forward convolutional NNs. In particular, we assume iid parameters distributed according to a stable distribution and we show that the infinite-channel limit of a deep feed-forward convolutional NNs, under suitable scaling, is a stochastic process with multivariate stable finite-dimensional distributions. Such a limiting distribution is then characterized through an explicit backward recursion for its parameters over the layers. Our contribution extends results of Favaro et al. (2020) to convolutional architectures, and it paves the way to expand exciting recent lines of research that rely on classes of GP limits.

Stefano Peluchetti, Stefano Favaro

Modern neural networks (NN) featuring a large number of layers (depth) and units per layer (width) have achieved a remarkable performance across many domains. While there exists a vast literature on the interplay between infinitely wide NNs and Gaussian processes, a little is known about analogous interplays with respect to infinitely deep NNs. NNs with independent and identically distributed (i.i.d.) initializations exhibit undesirable forward and backward propagation properties as the number of layers increases. To overcome these drawbacks, Peluchetti and Favaro (2020) considered fully-connected residual networks (ResNets) with network's parameters initialized by means of distributions that shrink as the number of layers increases, thus establishing an interplay between infinitely deep ResNets and solutions to stochastic differential equations, i.e. diffusion processes, and showing that infinitely deep ResNets does not suffer from undesirable forward-propagation properties. In this paper, we review the results of Peluchetti and Favaro (2020), extending them to convolutional ResNets, and we establish analogous backward-propagation results, which directly relate to the problem of training fully-connected deep ResNets. Then, we investigate the more general setting of doubly infinite NNs, where both network's width and network's depth grow unboundedly. We focus on doubly infinite fully-connected ResNets, for which we consider i.i.d. initializations. Under this setting, we show that the dynamics of quantities of interest converge, at initialization, to deterministic limits. This allow us to provide analytical expressions for inference, both in the case of weakly trained and fully trained ResNets. Our results highlight a limited expressive power of doubly infinite ResNets when the unscaled network's parameters are i.i.d. and the residual blocks are shallow.

Stefano Favaro, Sandra Fortini, Stefano Peluchetti

We consider fully connected feed-forward deep neural networks (NNs) where weights and biases are independent and identically distributed as symmetric centered stable distributions. Then, we show that the infinite wide limit of the NN, under suitable scaling on the weights, is a stochastic process whose finite-dimensional distributions are multivariate stable distributions. The limiting process is referred to as the stable process, and it generalizes the class of Gaussian processes recently obtained as infinite wide limits of NNs (Matthews at al., 2018b). Parameters of the stable process can be computed via an explicit recursion over the layers of the network. Our result contributes to the theory of fully connected feed-forward deep NNs, and it paves the way to expand recent lines of research that rely on Gaussian infinite wide limits.

Stefano Peluchetti, Stefano Favaro

When the parameters are independently and identically distributed (initialized) neural networks exhibit undesirable properties that emerge as the number of layers increases, e.g. a vanishing dependency on the input and a concentration on restrictive families of functions including constant functions. We consider parameter distributions that shrink as the number of layers increases in order to recover well-behaved stochastic processes in the limit of infinite depth. This leads to set forth a link between infinitely deep residual networks and solutions to stochastic differential equations, i.e. diffusion processes. We show that these limiting processes do not suffer from the aforementioned issues and investigate their properties.

Fadhel Ayed, Marco Battiston, Federico Camerlenghi et al.

Given $n$ samples from a population of individuals belonging to different types with unknown proportions, how do we estimate the probability of discovering a new type at the $(n+1)$-th draw? This is a classical problem in statistics, commonly referred to as the missing mass estimation problem. Recent results by Ohannessian and Dahleh \citet{Oha12} and Mossel and Ohannessian \citet{Mos15} showed: i) the impossibility of estimating (learning) the missing mass without imposing further structural assumptions on the type proportions; ii) the consistency of the Good-Turing estimator for the missing mass under the assumption that the tail of the type proportions decays to zero as a regularly varying function with parameter $α\in(0,1)$. In this paper we rely on tools from Bayesian nonparametrics to provide an alternative, and simpler, proof of the impossibility of a distribution-free estimation of the missing mass. Up to our knowledge, the use of Bayesian ideas to study large sample asymptotics for the missing mass is new, and it could be of independent interest. Still relying on Bayesian nonparametric tools, we then show that under regularly varying type proportions the convergence rate of the Good-Turing estimator is the best rate that any estimator can achieve, up to a slowly varying function, and that minimax rate must be at least $n^{-α/2}$. We conclude with a discussion of our results, and by conjecturing that the Good-Turing estimator is an rate optimal minimax estimator under regularly varying type proportions.

Marco Battiston, Stefano Favaro, Daniel M. Roy et al.

We characterize the class of exchangeable feature allocations assigning probability $V_{n,k}\prod_{l=1}^{k}W_{m_{l}}U_{n-m_{l}}$ to a feature allocation of $n$ individuals, displaying $k$ features with counts $(m_{1},\ldots,m_{k})$ for these features. Each element of this class is parametrized by a countable matrix $V$ and two sequences $U$ and $W$ of non-negative weights. Moreover, a consistency condition is imposed to guarantee that the distribution for feature allocations of $n-1$ individuals is recovered from that of $n$ individuals, when the last individual is integrated out. In Theorem 1.1, we prove that the only members of this class satisfying the consistency condition are mixtures of the Indian Buffet Process over its mass parameter $γ$ and mixtures of the Beta--Bernoulli model over its dimensionality parameter $N$. Hence, we provide a characterization of these two models as the only, up to randomization of the parameters, consistent exchangeable feature allocations having the required product form.

María Lomelí, Stefano Favaro, Yee Whye Teh

We investigate the class of $σ$-stable Poisson-Kingman random probability measures (RPMs) in the context of Bayesian nonparametric mixture modeling. This is a large class of discrete RPMs which encompasses most of the the popular discrete RPMs used in Bayesian nonparametrics, such as the Dirichlet process, Pitman-Yor process, the normalized inverse Gaussian process and the normalized generalized Gamma process. We show how certain sampling properties and marginal characterizations of $σ$-stable Poisson-Kingman RPMs can be usefully exploited for devising a Markov chain Monte Carlo (MCMC) algorithm for making inference in Bayesian nonparametric mixture modeling. Specifically, we introduce a novel and efficient MCMC sampling scheme in an augmented space that has a fixed number of auxiliary variables per iteration. We apply our sampling scheme for a density estimation and clustering tasks with unidimensional and multidimensional datasets, and we compare it against competing sampling schemes.