Search

We consider the learning--unlearning paradigm defined as follows. First given a dataset, the goal is to learn a good predictor, such as one minimizing a certain loss. Subsequently, given any subset of examples that wish to be unlearnt, the goal is to learn, without the knowledge of the original training dataset, a good predictor that is identical to the predictor that would have been produced when learning from scratch on the surviving examples. We propose a new ticketed model for learning--unlearning wherein the learning algorithm can send back additional information in the form of a small-sized (encrypted) ``ticket'' to each participating training example, in addition to retaining a small amount of ``central'' information for later. Subsequently, the examples that wish to be unlearnt present their tickets to the unlearning algorithm, which additionally uses the central information to return a new predictor. We provide space-efficient ticketed learning--unlearning schemes for a broad family of concept classes, including thresholds, parities, intersection-closed classes, among others. En route, we introduce the count-to-zero problem, where during unlearning, the goal is to simply know if there are any examples that survived. We give a ticketed learning--unlearning scheme for this problem that relies on the construction of Sperner families with certain properties, which might be of independent interest., Comment: Conference on Learning Theory (COLT) 2023

The optimal learning rate for adaptive gradient methods applied to {\lambda}-strongly convex functions relies on the parameters {\lambda} and learning rate {\eta}. In this paper, we adapt a universal algorithm along the lines of Metagrad, to get rid of this dependence on {\lambda} and {\eta}. The main idea is to concurrently run multiple experts and combine their predictions to a master algorithm. This master enjoys O(d log T) regret bounds.

Neuromorphic sampling is a paradigm shift in analog-to-digital conversion where the acquisition strategy is opportunistic and measurements are recorded only when there is a significant change in the signal. Neuromorphic sampling has given rise to a new class of event-based sensors called dynamic vision sensors or neuromorphic cameras. The neuromorphic sampling mechanism utilizes low power and provides high-dynamic range sensing with low latency and high temporal resolution. The measurements are sparse and have low redundancy making it convenient for downstream tasks. In this paper, we present a sampling-theoretic perspective to neuromorphic sensing of continuous-time signals. We establish a close connection between neuromorphic sampling and time-based sampling - where signals are encoded temporally. We analyse neuromorphic sampling of signals in shift-invariant spaces, in particular, bandlimited signals and polynomial splines. We present an iterative technique for perfect reconstruction subject to the events satisfying a density criterion. We also provide necessary and sufficient conditions for perfect reconstruction. Owing to practical limitations in meeting the sufficient conditions for perfect reconstruction, we extend the analysis to approximate reconstruction from sparse events. In the latter setting, we pose signal reconstruction as a continuous-domain linear inverse problem whose solution can be obtained by solving an equivalent finite-dimensional convex optimization program using a variable-splitting approach. We demonstrate the performance of the proposed algorithm and validate our claims via experiments on synthetic signals.

Solving inverse problems, where we find the input values that result in desired values of outputs, can be challenging. The solution process is often computationally expensive and it can be difficult to interpret the solution in high-dimensional input spaces. In this paper, we use a problem from additive manufacturing to address these two issues with the intent of making it easier to solve inverse problems and exploit their results. First, focusing on Gaussian process surrogates that are used to solve inverse problems, we describe how a simple modification to the idea of tapering can substantially speed up the surrogate without losing accuracy in prediction. Second, we demonstrate that Kohonen self-organizing maps can be used to visualize and interpret the solution to the inverse problem in the high-dimensional input space. For our data set, as not all input dimensions are equally important, we show that using weighted distances results in a better organized map that makes the relationships among the inputs obvious., Comment: 16 figures and 4 tables

Machine learning is typically framed from a perspective of i.i.d., and more importantly, isolated data. In parts, federated learning lifts this assumption, as it sets out to solve the real-world challenge of collaboratively learning a shared model from data distributed across clients. However, motivated primarily by privacy and computational constraints, the fact that data may change, distributions drift, or even tasks advance individually on clients, is seldom taken into account. The field of continual learning addresses this separate challenge and first steps have recently been taken to leverage synergies in distributed supervised settings, in which several clients learn to solve changing classification tasks over time without forgetting previously seen ones. Motivated by these prior works, we posit that such federated continual learning should be grounded in unsupervised learning of representations that are shared across clients; in the loose spirit of how humans can indirectly leverage others' experience without exposure to a specific task. For this purpose, we demonstrate that masked autoencoders for distribution estimation are particularly amenable to this setup. Specifically, their masking strategy can be seamlessly integrated with task attention mechanisms to enable selective knowledge transfer between clients. We empirically corroborate the latter statement through several continual federated scenarios on both image and binary datasets.

Sampling techniques are used in many fields, including design of experiments, image processing, and graphics. The techniques in each field are designed to meet the constraints specific to that field such as uniform coverage of the range of each dimension or random samples that are at least a certain distance apart from each other. When an application imposes new constraints, for example, by requiring samples in a non-rectangular domain or the addition of new samples to an existing set, a common solution is to modify the algorithm currently in use, often with less than satisfactory results. As an alternative, we propose the concept of intelligent sampling, where we devise algorithms specifically tailored to meet our sampling needs, either by creating new algorithms or by modifying suitable algorithms from other fields. Surprisingly, both qualitative and quantitative comparisons indicate that some relatively simple algorithms can be easily modified to meet the many sampling requirements of surrogate modeling, hyperparameter optimization, and data analysis; these algorithms outperform their more sophisticated counterparts currently in use, resulting in better use of time and computer resources., Comment: 4 Tables, 18 Figures

Urbanization is advancing rapidly, covering less than 2% of Earth's surface yet profoundly influencing global environments and experiencing disproportionate impacts from extreme weather events. Effective urban management and planning require high-resolution, temporally consistent datasets that capture the complexity of urban growth and dynamics. This study presents NDUI+, a novel global urban dataset addressing critical gaps in urban data continuity and quality. NDUI+ integrates data from the Defense Meteorological Satellite Program's Operational Linescan System (DMSP-OLS), VIIRS Nighttime Light, and Landsat 7 NDVI using advanced remote sensing and deep learning techniques. The dataset resolves sensor discontinuity challenges, offering a seamless 30-meter spatial and annual temporal resolution time series from 1999 to the present. NDUI+ demonstrates high precision and granularity, aligning closely with high-resolution satellite data and capturing urban dynamics effectively. The dataset provides valuable insights for urban climate studies, IPCC assessments, and urbanization research, complementing resources like UT-GLOBUS for urban modeling.

A set $D$ of vertices in a graph $G=(V,E)$ is a degree restricted dominating set for $G$ if each vertex $v_i$ in $D$ is dominating atmost $g(d_i)$ vertices of $V-D$, where $g$ is a function restricting the degree value $d_i$ with respect to the given function value $k_i$ for a natural valued function $f$ from the vertex set of the graph. We define three different types of Degree Restricted Domination by varying the way how the restricted function $g(v_i)$ is defined. If $g(d_i)=\big\lceil \frac{d_i}{k_i}\big\rceil$, the corresponding domination is called the ceil degree restricted domination, in short, $CDRD$, and the dominating set obtained in this manner is the $CDRD$-set. If $g(d_i)=\big\lfloor\frac{d_i}{k_i}\big\rfloor$ or $g(d_i)=d_i-k_i+1$, then the corresponding dominations are respectively called the floor degree restricted domination, in short $FDRD$, or the translate degree restricted domination, $TDRD$. The dominating sets obtained in this manner are the $FDRD$-set and the $TDRD$-set respectively. In this paper, we introduce these new generalizations of the domination number in line with the different $DRD$-sets and study these types of domination for some classes of graphs like complete graphs, caterpillar graphs etc. Degree restricted domination has a vital role in retaining the efficiency of nodes in a network and has many interesting applications., Comment: 9 pages

Performant vision-language (VL) models like CLIP represent captions using a single vector. How much information about language is lost in this bottleneck? We first curate CompPrompts, a set of increasingly compositional image captions that VL models should be able to capture (e.g., single object, to object+property, to multiple interacting objects). Then, we train text-only recovery probes that aim to reconstruct captions from single-vector text representations produced by several VL models. This approach does not require images, allowing us to test on a broader range of scenes compared to prior work. We find that: 1) CLIP's text encoder falls short on more compositional inputs, including object relationships, attribute-object association, counting, and negations; 2) some text encoders work significantly better than others; and 3) text-only recovery performance predicts multi-modal matching performance on ControlledImCaps: a new evaluation benchmark we collect and release consisting of fine-grained compositional images and captions. Specifically, our results suggest text-only recoverability is a necessary (but not sufficient) condition for modeling compositional factors in contrastive VL models. We release our datasets and code., Comment: EMNLP 2023

Poincar\'e plots, also called Poincar\'e maps, are used by plasma physicists to understand the behavior of magnetically confined plasma in numerical simulations of a tokamak. These plots are created by the intersection of field lines with a two-dimensional poloidal plane that is perpendicular to the axis of the torus representing the tokamak. A plot is composed of multiple orbits, each created by a different field line as it goes around the torus. Each orbit can have one of four distinct shapes, or classes, that indicate changes in the topology of the magnetic fields confining the plasma. Given the (x,y) coordinates of the points that form an orbit, the analysis task is to assign a class to the orbit, a task that appears ideally suited for a machine learning approach. In this paper, we describe how we overcame two major challenges in solving this problem - creating a high-quality training set, with few mislabeled orbits, and converting the coordinates of the points into features that are discriminating, despite the variation within the orbits of a class and the apparent similarities between orbits of different classes. Our automated approach is not only more objective and accurate than visual classification, but is also less tedious, making it easier for plasma physicists to analyze the topology of magnetic fields from numerical simulations of the tokamak., Comment: 19 pages, 12 Figures, 4 Tables

Understanding the spatio-temporal evolution of thermal gradient (G) at and velocity (R) of the solid-liquid and liquid-vapor interfaces is critical for the control of site-specific microstructures in additive manufacturing. In-situ dynamic x-ray radiography (DXR) has been used in recent years to probe the evolution of R with high spatial and temporal resolutions. However, the current methods used to measure the temperature (and therefore, G) are inadequate for sub-melt-pool surface measurement (e.g. thermography) or have lower resolution or limited field-of-view (e.g. x-ray diffraction). In this study, we demonstrate a novel approach to estimate the sub-surface temperature distribution and its time evolution with significantly higher resolutions using DXR data. This methodology uses the Beer-Lambert's law as a physical basis and is demonstrated using an in-situ laser spot-melting experiment on Ti-6Al-4V alloy.

We introduce a new mechanism for stochastic convex optimization (SCO) with user-level differential privacy guarantees. The convergence rates of this mechanism are similar to those in the prior work of Levy et al. (2021); Narayanan et al. (2022), but with two important improvements. Our mechanism does not require any smoothness assumptions on the loss. Furthermore, our bounds are also the first where the minimum number of users needed for user-level privacy has no dependence on the dimension and only a logarithmic dependence on the desired excess error. The main idea underlying the new mechanism is to show that the optimizers of strongly convex losses have low local deletion sensitivity, along with an output perturbation method for functions with low local deletion sensitivity, which could be of independent interest.

Many properties of circumbinary discs around evolved post-asymptotic giant branch (post-AGB) binary systems are similar to those of protoplanetary discs around young stars. The deficits of near-infrared (near-IR) flux in the spectral energy distributions (SEDs) of these systems hints towards large dust-free cavities that are reminiscent of transition discs as are commonly observed around young stars. We aim to assess the size of the inner rim of 6 post-AGB binary systems with lack in the near-IR like this. We used resolved mid-infrared (mid-IR) high-angular resolution observations of VLTI/MATISSE and VLTI/MIDI. We compared these inner rim sizes to 5 systems with available MATISSE data that were identified to host a disc starting at the dust sublimation radius. We used geometric ring models to estimate the inner rim sizes, the relative flux contributions of the star, the ring, and an over-resolved emission, the orientation of the ring, and the spectral dependences of the components. We find that the inner dust rims of the targets with a lack of near-IR excess in their SEDs are 2.5 to 7.5 times larger than the theoretical dust sublimation radii, and inner rim sizes of the systems that do not show this deficit are similar to those of their theoretical dust sublimation radii. The physical radii of the inner rims of these transition discs around post-AGB binaries are 3-25 au, which are larger than the disc sizes inferred for transition discs around young stars with VLTI/MIDI. With mid-IR interferometric data, we directly confirm the transition disc nature of six circumbinary discs around post-AGB binary systems. Future observational and modelling efforts are needed to progress in our understanding of the structure, origin, and evolution of these transition discs, Comment: accepted for publication in A&A. 13 pages, including appendices

In this paper, we propose a data-driven approach to train a Generative Adversarial Network (GAN) conditioned on "soft-labels" distilled from the penultimate layer of an audio classifier trained on a target set of audio texture classes. We demonstrate that interpolation between such conditions or control vectors provides smooth morphing between the generated audio textures, and shows similar or better audio texture morphing capability compared to the state-of-the-art methods. The proposed approach results in a well-organized latent space that generates novel audio outputs while remaining consistent with the semantics of the conditioning parameters. This is a step towards a general data-driven approach to designing generative audio models with customized controls capable of traversing out-of-distribution regions for novel sound synthesis., Comment: accepted to ICASSP 2023

Large pre-trained language models (LLMs) have been shown to have significant potential in few-shot learning across various fields, even with minimal training data. However, their ability to generalize to unseen tasks in more complex fields, such as biology, has yet to be fully evaluated. LLMs can offer a promising alternative approach for biological inference, particularly in cases where structured data and sample size are limited, by extracting prior knowledge from text corpora. Our proposed few-shot learning approach uses LLMs to predict the synergy of drug pairs in rare tissues that lack structured data and features. Our experiments, which involved seven rare tissues from different cancer types, demonstrated that the LLM-based prediction model achieved significant accuracy with very few or zero samples. Our proposed model, the CancerGPT (with $\sim$ 124M parameters), was even comparable to the larger fine-tuned GPT-3 model (with $\sim$ 175B parameters). Our research is the first to tackle drug pair synergy prediction in rare tissues with limited data. We are also the first to utilize an LLM-based prediction model for biological reaction prediction tasks.

In this article, we present a detailed review of current practices and state-of-the-art methodologies in the field of differential privacy (DP), with a focus of advancing DP's deployment in real-world applications. Key points and high-level contents of the article were originated from the discussions from "Differential Privacy (DP): Challenges Towards the Next Frontier," a workshop held in July 2022 with experts from industry, academia, and the public sector seeking answers to broad questions pertaining to privacy and its implications in the design of industry-grade systems. This article aims to provide a reference point for the algorithmic and design decisions within the realm of privacy, highlighting important challenges and potential research directions. Covering a wide spectrum of topics, this article delves into the infrastructure needs for designing private systems, methods for achieving better privacy/utility trade-offs, performing privacy attacks and auditing, as well as communicating privacy with broader audiences and stakeholders.

We present an end-to-end computer vision pipeline to detect non-nutritive sucking (NNS) -- an infant sucking pattern with no nutrition delivered -- as a potential biomarker for developmental delays, using off-the-shelf baby monitor video footage. One barrier to clinical (or algorithmic) assessment of NNS stems from its sparsity, requiring experts to wade through hours of footage to find minutes of relevant activity. Our NNS activity segmentation algorithm solves this problem by identifying periods of NNS with high certainty -- up to 94.0\% average precision and 84.9\% average recall across 30 heterogeneous 60 s clips, drawn from our manually annotated NNS clinical in-crib dataset of 183 hours of overnight baby monitor footage from 19 infants. Our method is based on an underlying NNS action recognition algorithm, which uses spatiotemporal deep learning networks and infant-specific pose estimation, achieving 94.9\% accuracy in binary classification of 960 2.5 s balanced NNS vs. non-NNS clips. Tested on our second, independent, and public NNS in-the-wild dataset, NNS recognition classification reaches 92.3\% accuracy, and NNS segmentation achieves 90.8\% precision and 84.2\% recall.

As general purpose vision models get increasingly effective at a wide set of tasks, it is imperative that they be consistent across the tasks they support. Inconsistent AI models are considered brittle and untrustworthy by human users and are more challenging to incorporate into larger systems that take dependencies on their outputs. Measuring consistency between very heterogeneous tasks that might include outputs in different modalities is challenging since it is difficult to determine if the predictions are consistent with one another. As a solution, we introduce a benchmark dataset, CocoCon, where we create contrast sets by modifying test instances for multiple tasks in small but semantically meaningful ways to change the gold label and outline metrics for measuring if a model is consistent by ranking the original and perturbed instances across tasks. We find that state-of-the-art vision-language models suffer from a surprisingly high degree of inconsistent behavior across tasks, especially for more heterogeneous tasks. To alleviate this issue, we propose a rank correlation-based auxiliary training objective, computed over large automatically created cross-task contrast sets, that improves the multi-task consistency of large unified models while retaining their original accuracy on downstream tasks., Comment: TMLR 2024; Project Website: https://adymaharana.github.io/cococon/

Indiscriminate data poisoning attacks aim to decrease a model's test accuracy by injecting a small amount of corrupted training data. Despite significant interest, existing attacks remain relatively ineffective against modern machine learning (ML) architectures. In this work, we introduce the notion of model poisoning reachability as a technical tool to explore the intrinsic limits of data poisoning attacks towards target parameters (i.e., model-targeted attacks). We derive an easily computable threshold to establish and quantify a surprising phase transition phenomenon among popular ML models: data poisoning attacks can achieve certain target parameters only when the poisoning ratio exceeds our threshold. Building on existing parameter corruption attacks and refining the Gradient Canceling attack, we perform extensive experiments to confirm our theoretical findings, test the predictability of our transition threshold, and significantly improve existing indiscriminate data poisoning baselines over a range of datasets and models. Our work highlights the critical role played by the poisoning ratio, and sheds new insights on existing empirical results, attacks and mitigation strategies in data poisoning., Comment: Accepted to ICML 2023

Integral to the problem of detecting communities through graph clustering is the expectation that they are "well connected". In this respect, we examine five different community detection approaches optimizing different criteria: the Leiden algorithm optimizing the Constant Potts Model, the Leiden algorithm optimizing modularity, Iterative K-Core Clustering (IKC), Infomap, and Markov Clustering (MCL). Surprisingly, all these methods produce, to varying extents, communities that fail even a mild requirement for well connectedness. To remediate clusters that are not well connected, we have developed the "Connectivity Modifier" (CM), which, at the cost of coverage, iteratively removes small edge cuts and re-clusters until all communities produced are well connected. Results from real-world and synthetic networks illustrate a tradeoff users make between well connected clusters and coverage, and raise questions about the "clusterability" of networks and models of community structure.

Differentially private stochastic gradient descent privatizes model training by injecting noise into each iteration, where the noise magnitude increases with the number of model parameters. Recent works suggest that we can reduce the noise by leveraging public data for private machine learning, by projecting gradients onto a subspace prescribed by the public data. However, given a choice of public datasets, it is not a priori clear which one may be most appropriate for the private task. We give an algorithm for selecting a public dataset by measuring a low-dimensional subspace distance between gradients of the public and private examples. We provide theoretical analysis demonstrating that the excess risk scales with this subspace distance. This distance is easy to compute and robust to modifications in the setting. Empirical evaluation shows that trained model accuracy is monotone in this distance.

The goal of this study is to reconstruct the evolution and the dust formation processes during the final AGB phases of a sample of carbon-rich, post-AGB Galactic stars, with particular attention to the determination of the past mass-loss history. We study the IR excess of sources classified as single stars by means of dust formation modelling where dust grains form and grow in a static wind and expand from the surface of the star. The method is applied to various evolutionary stages of the final AGB phase of stars with different masses and metallicities. The detailed analysis of the SED of the sources investigated, which included the derivation of the luminosities and the dust properties, is used to infer information on mass loss, efficiency of dust formation, and wind dynamics. We confirm previous results that most of the investigated sources descend from low-mass(M<1.5Msun) progenitors that reached the C-star stage. Metal-poor carbon stars are characterised by higher IR excesses with respect to their more metal-rich counterparts of similar luminosity due to a higher surface carbon-to-oxygen excess. This work confirms previous conclusions that more luminous stars descending from higher-mass progenitors are generally more opaque due to shorter evolutionary timescales that place the dust shell closer to the central object. We also find that the mass-loss rate at the tip of the AGB phase of metal-rich low-mass carbon stars is approximately 1-1.5x10^-5Msun/yr, whereas in the metal-poor domain M~4-5x10^-5Msun/yr is required. These results indicate the need for an upwards revision of the theoretical mass-loss rates of low-mass carbon stars in the available literature, which in turn require a revised determination of carbon dust yields by AGB stars., Comment: 11 pages, 5 figures

We show that the canonical approach for training differentially private GANs -- updating the discriminator with differentially private stochastic gradient descent (DPSGD) -- can yield significantly improved results after modifications to training. Specifically, we propose that existing instantiations of this approach neglect to consider how adding noise only to discriminator updates inhibits discriminator training, disrupting the balance between the generator and discriminator necessary for successful GAN training. We show that a simple fix -- taking more discriminator steps between generator steps -- restores parity between the generator and discriminator and improves results. Additionally, with the goal of restoring parity, we experiment with other modifications -- namely, large batch sizes and adaptive discriminator update frequency -- to improve discriminator training and see further improvements in generation quality. Our results demonstrate that on standard image synthesis benchmarks, DPSGD outperforms all alternative GAN privatization schemes. Code: https://github.com/alexbie98/dpgan-revisit., Comment: 28 pages; revisions and new experiments from TMLR camera-ready + code release at https://github.com/alexbie98/dpgan-revisit

Predicting transport rates of windblown sand is a central problem in aeolian research, with implications for climate, environmental, and planetary sciences. Though studied since the 1930s, the underlying many-body dynamics is still incompletely understood, as underscored by the recent empirical discovery of an unexpected third-root scaling in the particle-fluid density ratio. Here, by means of grain-scale simulations and analytical modeling, we elucidate how a complex coupling between grain-bed collisions and granular creep within the sand bed yields a dilatancy-enhanced bed erodibility. Our minimal saltation model robustly predicts both the observed scaling and a new undersaturated steady transport state that we confirm by simulations for rarefied atmospheres.

Differential privacy (DP) is a rigorous notion of data privacy, used for private statistics. The canonical algorithm for differentially private mean estimation is to first clip the samples to a bounded range and then add noise to their empirical mean. Clipping controls the sensitivity and, hence, the variance of the noise that we add for privacy. But clipping also introduces statistical bias. This tradeoff is inherent: we prove that no algorithm can simultaneously have low bias, low error, and low privacy loss for arbitrary distributions. Additionally, we show that under strong notions of DP (i.e., pure or concentrated DP), unbiased mean estimation is impossible, even if we assume that the data is sampled from a Gaussian. On the positive side, we show that unbiased mean estimation is possible under a more permissive notion of differential privacy (approximate DP) if we assume that the distribution is symmetric.

Fluorophores emitting in the NIR-IIb wavelength range (1.5 micron - 1.7 micron) show great potential for bioimaging due to their large tissue penetration. However, current fluorophores suffer from poor emission with quantum yields ~2% in aqueous solvents. In this work, we report the synthesis of HgSe/CdSe core/shell quantum dots emitting at 1.7 microns through the interband transition. Growth of a thick shell led to a drastic increase in the photoluminescence quantum yield, with a value of 55% in nonpolar solvents. The quantum yields of our QDs and other reported QDs are explained well by a model of Forster resonance energy transfer to ligands and solvent molecules. The model predicts a quantum yield >6% when these HgSe/CdSe QDs are solubilized in water. Our work demonstrates the importance of a thick type-I shell to obtain bright emission in the NIR-IIb region, Comment: Main text: 9 pages, 4 figures. Supplementary Information: 31 pages, 21 figures

(abridged) Stable circumbinary discs around evolved post-Asymptotic Giant branch (post-AGB) binary systems show many similarities with protoplanetary discs around young stellar objects. These discs can provide constraints on both binary evolution and the formation of macrostructures within circumstellar discs. Here we focus on one post-AGB binary system: IRAS08544-4431. We aim to refine the physical model of IRAS08544-4431 with a radiative transfer treatment and continue the near-infrared and mid-infrared interferometric analysis covering the H-, K-, L-, and N-bands. We aim to capture the previously detected amount of over-resolved flux and the radial intensity profile at and beyond the inner dust disc rim to put constraints on the physical processes in the inner disc regions. We used a Monte Carlo radiative transfer code to investigate the physical structure of the disc by reproducing both the photometry and the multi-wavelength infrared interferometric data set. We developed a strategy to identify the models which perform best to reproduce our data set. We found a family of models that successfully fit the infrared photometric and interferometric data in all bands. Some over-resolved flux component was recovered in all bands but the optimised models still fall short to explain all the over-resolved flux. This suggests that another dusty structure within the system plays a role. Multi-wavelength infrared interferometric observations of circumstellar discs allow to study the inner disc regions in unprecedented detail. The refined physical models can reproduce most of the investigated features, including the photometric characteristics, the radial extent, and the overall shape of the visibility curves. Our multi-wavelength interferometric observations combined with photometry show that the disc is similar to protoplanetary discs with similar dust masses and efficient dust growth., Comment: 18 pages, 13 figures (including apppendix), accepted for publication in A&A

The formation of silicates in circumstellar envelopes of stars evolving through the AGB is still debated given the uncertainties affecting stellar evolution modelling, the description of the dust formation process, and the capability of silicate grains to accelerate stellar outflows via radiation pressure. We study the formation of dust in the winds of intermediate mass (M $\geq 4 M_{\odot}$) stars of solar metallicity while evolving through the AGB phase. We tested the different treatments of the mass-loss mechanism by this class of stars, with the aim of assessing their contribution to the general enrichment of silicates of the interstellar medium of galaxies. We consider a sub-sample of AGB stars, whose SED is characterised by deep absorption features at $10$ and $18\mu$m, which can be regarded as the class of stars providing the most relevant contribution to the silicates' production across the Universe. Results from stellar evolution and dust formation modelling were used to fit the observed SED and to reproduce, at the same time, the detected pulsation periods and the derived surface chemical composition. This analysis leads to the derivation of tight constraints on the silicates' production rates experienced by these sources during the final AGB stages. Two out of the four sources investigated are interpreted as stars currently undergoing HBB, evolving through phases close to the stage when the mass-loss rate is largest. The remaining two stars are likely evolving through the very final AGB phases, after HBB was turned off by the gradual consumption of the convective mantle. Mass-loss rates of the order of $1-2\times 10^{-4} M_{\odot}/$yr are required when looking for consistency with the observational evidence. These results indicate the need for a revision of the silicate yields by intermediate mass stars, which are found to be $\sim 3$ times higher than previously determined.