Simple Multilayer Perceptron Research Articles

Explainable Graph Neural Networks: An Application to Open Statistics Knowledge Graphs for Estimating House Prices

In the rapidly evolving field of real estate economics, the prediction of house prices continues to be a complex challenge, intricately tied to a multitude of socio-economic factors. Traditional predictive models often overlook spatial interdependencies that significantly influence housing prices. The objective of this study is to leverage Graph Neural Networks (GNNs) on open statistics knowledge graphs to model these spatial dependencies and predict house prices across Scotland’s 2011 data zones. The methodology involves retrieving integrated statistical indicators from the official Scottish Open Government Data portal and applying three representative GNN algorithms: ChebNet, GCN, and GraphSAGE. These GNNs are compared against traditional models, including the tabular-based XGBoost and a simple Multi-Layer Perceptron (MLP), demonstrating superior prediction accuracy. Innovative contributions of this study include the use of GNNs to model spatial dependencies in real estate economics and the application of local and global explainability techniques to enhance transparency and trust in the predictions. The global feature importance is determined by a logistic regression surrogate model while the local, region-level understanding of the GNN predictions is achieved through the use of GNNExplainer. Explainability results are compared with those from a previous work that applied the XGBoost machine learning algorithm and the SHapley Additive exPlanations (SHAP) explainability framework on the same dataset. Interestingly, both the global surrogate model and the SHAP approach underscored the comparative illness factor, a health indicator, and the ratio of detached dwellings as the most crucial features in the global explainability. In the case of local explanations, while both methods showed similar results, the GNN approach provided a richer, more comprehensive understanding of the predictions for two specific data zones.

Impact of data for forecasting on performance of model predictive control in buildings with smart energy storage

Data is required to develop forecasting models for use in Model Predictive Control (MPC) schemes in building energy systems. However, data is costly to both collect and exploit. Determining cost optimal data usage strategies requires understanding of the forecast accuracy and resulting MPC operational performance it enables. This study investigates the performance of both simple and state-of-the-art machine learning prediction models for MPC in multi-building energy systems using a simulated case study with historic building energy data. The impact on forecast accuracy of measures to improve model data efficiency is quantified, specifically for: reuse of prediction models, reduction of training data duration, reduction of model data features, and online model training. A simple linear multi-layer perceptron model is shown to provide equivalent forecast accuracy to state-of-the-art models, with greater data efficiency and generalisability. The use of more than 2 years of training data for load prediction models provided no significant improvement in forecast accuracy. Forecast accuracy and data efficiency were improved simultaneously by using change-point analysis to screen training data. Reused models and those trained with 3 months of data had on average 10% higher error than baseline, indicating that deploying MPC systems without prior data collection may be economic.

Detecting images generated by diffusers.

In recent years, the field of artificial intelligence has witnessed a remarkable surge in the generation of synthetic images, driven by advancements in deep learning techniques. These synthetic images, often created through complex algorithms, closely mimic real photographs, blurring the lines between reality and artificiality. This proliferation of synthetic visuals presents a pressing challenge: how to accurately and reliably distinguish between genuine and generated images. This article, in particular, explores the task of detecting images generated by text-to-image diffusion models, highlighting the challenges and peculiarities of this field. To evaluate this, we consider images generated from captions in the MSCOCO and Wikimedia datasets using two state-of-the-art models: Stable Diffusion and GLIDE. Our experiments show that it is possible to detect the generated images using simple multi-layer perceptrons (MLPs), starting from features extracted by CLIP or RoBERTa, or using traditional convolutional neural networks (CNNs). These latter models achieve remarkable performances in particular when pretrained on large datasets. We also observe that models trained on images generated by Stable Diffusion can occasionally detect images generated by GLIDE, but only on the MSCOCO dataset. However, the reverse is not true. Lastly, we find that incorporating the associated textual information with the images in some cases can lead to a better generalization capability, especially if textual features are closely related to visual ones. We also discovered that the type of subject depicted in the image can significantly impact performance. This work provides insights into the feasibility of detecting generated images and has implications for security and privacy concerns in real-world applications. The code to reproduce our results is available at: https://github.com/davide-coccomini/Detecting-Images-Generated-by-Diffusers.

PSVMLP: Point and Shifted Voxel MLP for 3D deep learning

We propose a high-performance 3D feature extraction deep learning network based on point cloud and shifted voxel, named Point and Shifted Voxel MLP (PSVMLP). The main component of PSVMLP is simple Multi-Layer Perceptron (MLP) structure. PSVMLP achieves effective extraction of multi-scale features from 3D data. Specifically, we combine point cloud and voxel-based feature extraction methods. In voxel representation learning, we propose a wide-range geometric feature extraction method based on axial shifting operations and simple MLP structure. The axial shifting operations allow shifting voxels in the depth, height, and width directions, capturing more geometric information. In point cloud representation learning, we use simple MLP structure to extract local features, and we also extract global features by combining transformer structure. By combining point cloud and voxel feature extraction methods, we obtain rich feature representations from different scales, enhancing the model’s expressive power and generalization performance. Applying our designed model to basic geometric feature learning tasks, we achieve excellent results. Despite being built primarily on a simple MLP framework, our model demonstrates remarkable performance on both shape classification and shape part segmentation tasks. Our code is available at https://github.com/hitxraz/psvmlp.

A novel spectral model for prediction of nongray radiative properties in gaseous combustion

Radiative heat transfer during gaseous combustion displays very strong spectral or “nongray” behavior, which is both very difficult to characterize and to evaluate. A novel spectral model based on the full-spectrum correlated k-distribution (FSCK) method combined with the multilayer perceptron (MLP), i.e., NFM model, is developed for the prediction of nongray radiative properties in gaseous combustion. Followed by the nature of FSCK method, the NFM model is designed as a simple MLP structure that has only two outputs to directly give both correlated k-value and corresponding ka-value. This not only avoids on-the-fly calculations of a-values but also saves storage by the simple structure. Moreover, by constructing training data with a number of small mole fractions, the errors caused by nonlinear effects are found to be mitigated during the use of NFM model. Several test cases have been carried out to compare the developed NFM model and other spectral tools including FSCK look-up tables and traditional FSCK MLP (TFM) model. Results show that the NFM model can achieve excellent accuracy that is even better than FSCK look-up tables at a tiny computational cost that is far less than that of TFM model.

CSDFormer: A cloud and shadow detection method for landsat images based on transformer

Cloud and shadow (CS) detection is crucial prerequisite for application of remote sensing images. Current deep learning-based detection algorithms mainly employ Convolutional Neural Networks (CNNs). However, the local receptive field in CNNs cannot effectively capture global contextual information, which hinders accurate characterization of the dependency between clouds and shadows. In vision Transformers, self-attention mechanisms can effectively capture the long-distance dependencies between different regions in an image. Inspired by this, this paper proposed a new CS Detection algorithm based on a Transformer, called CSDFormer. Specifically, we exclusively employed a hierarchical Transformer structure in the encoder stage to extract features of CS. Each Transformer layer contains several multi-head self-attention mechanisms for calculating pixel-wise long-distance connectivity. The designed structure enables the Transformer to better extract global context information, which helps to strengthen the comprehension of the semantic relationships between clouds and shadows. Benefiting from the global feature extraction capability of the encoder stage, we employed several simple multilayer perceptron layers for multi-scale feature map fusion and pixel classification in the decoder stage. The proposed CSDFormer was validated using 898 Landsat 8 Biome images with 512 × 512 pixels, producing an overall accuracy of 95.28 % and a mean intersection over union of 84.08 %, outperforming three state-of-the-art CNN-based algorithms. CSDFormer is consistently more accurate in detection of both clouds and shadows. Owing to the parallel computing capability of the self-attention mechanism, CSDFormer is computationally more efficient than the three CNN-based benchmark methods. For the input spectral bands, the performance of CSDFormer produced can be further enhanced with additional thermal infrared bands.

ModWaveMLP: MLP-Based Mode Decomposition and Wavelet Denoising Model to Defeat Complex Structures in Traffic Forecasting

Traffic prediction is the core issue of Intelligent Transportation Systems. Recently, researchers have tended to use complex structures, such as transformer-based structures, for tasks such as traffic prediction. Notably, traffic data is simpler to process compared to text and images, which raises questions about the necessity of these structures. Additionally, when handling traffic data, researchers tend to manually design the model structure based on the data features, which makes the structure of traffic prediction redundant and the model generalizability limited. To address the above, we introduce the ‘ModWaveMLP’—A multilayer perceptron (MLP) based model designed according to mode decomposition and wavelet noise reduction information learning concepts. The model is based on simple MLP structure, which achieves the separation and prediction of different traffic modes and does not depend on additional features introduced such as the topology of the traffic network. By performing experiments on real-world datasets METR-LA and PEMS-BAY, our model achieves SOTA, outperforms GNN and transformer-based models, and outperforms those that introduce additional feature data with better generalizability, and we further demonstrate the effectiveness of the various parts of the model through ablation experiments. This offers new insights to subsequent researchers involved in traffic model design. The code is available at: https://github.com/Kqingzheng/ModWaveMLP.

SEMG-Based Robust Recognition of Grasping Postures with a Machine Learning Approach for Low-Cost Hand Control.

The design and control of artificial hands remains a challenge in engineering. Popular prostheses are bio-mechanically simple with restricted manipulation capabilities, as advanced devices are pricy or abandoned due to their difficult communication with the hand. For social robots, the interpretation of human intention is key for their integration in daily life. This can be achieved with machine learning (ML) algorithms, which are barely used for grasping posture recognition. This work proposes an ML approach to recognize nine hand postures, representing 90% of the activities of daily living in real time using an sEMG human-robot interface (HRI). Data from 20 subjects wearing a Myo armband (8 sEMG signals) were gathered from the NinaPro DS5 and from experimental tests with the YCB Object Set, and they were used jointly in the development of a simple multi-layer perceptron in MATLAB, with a global percentage success of 73% using only two features. GPU-based implementations were run to select the best architecture, with generalization capabilities, robustness-versus-electrode shift, low memory expense, and real-time performance. This architecture enables the implementation of grasping posture recognition in low-cost devices, aimed at the development of affordable functional prostheses and HRI for social robots.

Video-MLP: Convolution-free, attention-free architecture for video classification

Convolutional neural networks (CNNs) and Transformer architectures have traditionally been recognized as the preferred models for addressing computer vision tasks. However, there has been a recent surge in the popularity of networks based on multi-layer perceptron (MLP) structures that do not rely on convolution or attention mechanisms. These MLP architectures have demonstrated exceptional performance in image classification tasks, exhibiting lower time complexity while maintaining high accuracy. In contrast, video classification tasks involve larger amounts of data and necessitate more intricate feature extraction, resulting in increased time and resource consumption. To enhance computational efficiency and minimize resource utilization, we propose a convolution-free and Transformer-free architecture for video classification called Video-MLP for video classification. Video-MLP utilizes a simple MLP structure to learn video features. Specifically, it comprises two types of layers: Spatial-Mixer and Temporal-Mixer, which respectively capture spatial and temporal information. The Spatial-Mixer extracts spatial information from each frame along the height and width dimensions, while the Temporal-Mixer models temporal information for the same spatial positions across frames. To improve the efficiency of spatial-temporal modeling in our network, we used a spatial-temporal information fusion approach to integrate information at different scales. Additionally, we grouped the input data along the time dimension and designed three different grouping schemes when extracting temporal information. The experimental results indicate that Video-MLP achieved accuracy rates of 87.2% on the Kinetics-400 dataset and 75.3% on the SomethingV2 dataset, outperforming models with equivalent computational complexity. Notably, Video-MLP achieved these results without using convolution and attention mechanisms, and without pre-training on large-scale image and video datasets.

Artificial Neural Networks for Determining the Empirical Relationship between Meteorological Parameters and High-Level Cloud Characteristics

The special features of the applicability of artificial neural networks to the task of identifying relationships between meteorological parameters of the atmosphere and optical and geometric characteristics of high-level clouds (HLCs) containing ice crystals are investigated. The existing models describing such relationships do not take into account a number of atmospheric effects, in particular, the orientation of crystalline ice particles due to the simplified physical description of the medium, or within the framework of these models, accounting for such dependencies becomes a highly nontrivial task. Neural networks are able to take into account the complex interaction of meteorological parameters with each other, as well as reconstruct almost any dependence of the HLC characteristics on these parameters. In the process of prototyping the software product, the greatest difficulty was in determining the network architecture, the loss function, and the method of supplying the input parameters (attributes). Each of these problems affected the most important issue of neural networks—the overtraining problem, which occurs when the neural network stops summarizing data and starts to tune to them. Dependence on meteorological parameters was revealed for the following quantities: the altitude of the cloud center; elements m22 and m44 of the backscattering phase matrix (BSPM); and the m33 element of BSPM requires further investigation and expansion of the analyzed dataset. Significantly, the result is not affected by the compression method chosen to reduce the data dimensionality. In almost all cases, the random forest method gave a better result than a simple multilayer perceptron.

Transferability of Temperature Evolution of Dissimilar Wire-Arc Additively Manufactured Components by Machine Learning.

Wire-arc additive manufacturing (WAAM) is a promising industrial production technique. Without optimization, inherent temperature gradients can cause powerful residual stresses and microstructural defects. There is therefore a need for data-driven methods allowing real-time process optimization for WAAM. This study focuses on machine learning (ML)-based prediction of temperature history for WAAM-produced aluminum bars with different geometries and process parameters, including bar length, number of deposition layers, and heat source movement speed. Finite element (FE) simulations are used to provide training and prediction data. The ML models are based on a simple multilayer perceptron (MLP) and performed well during baseline training and testing, giving a testing mean absolute percentage error (MAPE) of less than 0.7% with an 80/20 train-test split, with low variation in model performance. When using the trained models to predict results from FE simulations with greater length or number of layers, the MAPE increased to an average of 3.22% or less, with greater variability. In the cases of greatest difference, some models still returned a MAPE of less than 1%. For different scanning speeds, the performance was worse, with some outlier models giving a MAPE of up to 14.91%. This study demonstrates the transferability of temperature history for WAAM with a simple MLP approach.

Interferometric image reconstruction using closure invariants and machine learning

Abstract Interferometric closure invariants encode calibration-independent details of an object’s morphology. Excepting simple cases, a direct backward transformation from closure invariants to morphologies is not well established. We demonstrate using simple Machine Learning models that closure invariants can aid in morphological classification and parameter estimation. We consider six phenomenologically parametrized morphologies: point-like, uniform circular disc, crescent, dual disc, crescent with elliptical accretion disc, and crescent with double jet lobes. Using logistic regression (LR), multilayer perceptron (MLP), and random forest models on closure invariants obtained from a sparsely covered aperture, we find that all methods except LR can classify morphologies with $\gtrsim$80 per cent accuracy, which improves with greater aperture coverage. Separately from the classification problem, given an independently confirmed class, we estimate parameters of uniform circular disc, crescent, and dual disc morphologies using simple MLP models, and parametrically reconstruct images. The estimated parameters and images correspond well with inputs, but the accuracy worsens when degeneracies between parameters are present. This independent approach to interferometric imaging under challenging observing conditions such as that faced by the Event Horizon Telescope and Very Long Baseline Interferometry, in general, can complement other methods in robustly constraining an object’s morphology.

A Data-Driven Machine Learning Approach for Turbulent Flow Field Prediction Based on Direct Computational Fluid Dynamics Database

A novel approach is presented for predicting compressible turbulent flow fields using a neural network-based data-driven method. Accurate prediction in turbulent regions heavily relies on the resolution of available data. Traditional methods, employing image-based techniques by mapping scattered computational fluid dynamics (CFD) data onto Cartesian grids, encounter data scarcity in critical areas such as the boundary layer and wake. Recently, convolutional neural networks (CNN) have gained prominence as the most widely referenced technique in fluid dynamics, utilizing flow field images as datasets for flow field prediction. However, CNN requires datasets with a high pixel density to enhance training accuracy in crucial regions, thereby increasing the input data volume and machine training time. To address this challenge, our proposed method deviates from using flow field images and instead generates datasets directly from the flow field properties of CFD grid points. By employing this approach, several advantages are realized. Firstly, the network benefits from the favorable characteristics of unstructured grids, such as varying point spacing near the object surface and in the far field, which effectively reduces the amount of input data and consequently the machine training cost. Secondly, the construction of the training dataset eliminates the need for interpolation or extrapolation, thereby preserving the accuracy of CFD data. In this case, a simple multilayer perceptron can be trained using the proposed dataset. Various flow field properties, including static pressure, turbulent kinetic energy, and velocity components, can be predicted with high accuracy within a few seconds.

Arbitrary Shape Text Detection via Boundary Transformer

In arbitrary shape text detection, locating accurate text boundaries is challenging and non-trivial. Existing methods often suffer from indirect text boundary modeling or complex post-processing. In this paper, we systematically present a unified coarse-to-fine framework via boundary learning for arbitrary shape text detection, which can accurately and efficiently locate text boundaries without post-processing. In our method, we explicitly model the text boundary via an innovative iterative boundary transformer in a coarse-to-fine manner. In this way, our method can directly gain accurate text boundaries and abandon complex post-processing to improve efficiency. Specifically, our method mainly consists of a feature extraction backbone, a boundary proposal module, and an iteratively optimized boundary transformer module. The boundary proposal module consisting of multi-layer dilated convolutions will predict important prior information (including classification map, distance field, and direction field) for generating coarse boundary proposals while guiding the boundary transformer's optimization. The boundary transformer module adopts an encoder-decoder structure, in which the encoder is constructed by multi-layer transformer blocks with residual connection while the decoder is a simple multi-layer perceptron network (MLP). Under the guidance of prior information, the boundary transformer module will gradually refine the coarse boundary proposals via iterative boundary deformation. Furthermore, we propose a novel boundary energy loss (BEL) that introduces an energy minimization constraint and an energy monotonically decreasing constraint to further optimize and stabilize the learning of boundary refinement. Extensive experiments on publicly available and challenging datasets demonstrate the state-of-the-art performance and promising efficiency of our method. The code and model are available at: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/GXYM/TextBPN-Puls-Plus</uri> .

Efficient approximate multipliers utilizing compact and low-power compressors for error-resilient applications

Approximate computing is one of the promising techniques in error-resilient applications to overcome high-density integration challenges, such as energy consumption and performance. Multipliers constitute a significant portion of computer arithmetic units, leading to considerable energy and time consumption. In this paper, we propose low-power and compact approximate compressors for composing approximate Dadda multiplier structures, including compressors, half adders, and full adders, which utilize three-phase partial product compression: truncated, approximation, and exact columns. In approximate columns, we consider approximate compressors derived from the truth table of the exact 4:2 compressor and simplified K-map entries based on the probability of each combination of inputs. An error-correcting module (ECM) is designed to distinguish specific cases and reduce the error metrics. All circuits were simulated using ModelSim and then synthesized using Design Compiler with the 15 nm FinFET technology. When compared to state-of-the-art works, our multipliers exhibit approximately 30%, 43%, and 10% reductions in power, area, and delay, respectively. To evaluate their functionality, we conducted image multiplication and implemented a simple multi-layer perceptron (MLP) neural network using the modified National Institute of Standards and Technology (MNIST) dataset in MATLAB with 0.998 mean structural similarity index metric (MSSIM), 51 dB peak-signal noise ratio (PSNR), and 95% classification accuracy.

UniG-Encoder: A universal feature encoder for graph and hypergraph node classification

Despite the decent performance and fruitful applications of Graph Neural Networks (GNNs), Hypergraph Neural Networks (HGNNs), and their well-designed variants, on some commonly used benchmark graphs and hypergraphs, they are outperformed by even a simple Multi-Layer Perceptron. This observation motivates a reexamination of the design paradigm of the current GNNs and HGNNs and poses challenges of extracting graph features effectively. In this work, a universal feature encoder for both graph and hypergraph representation learning is designed, called UniG-Encoder. The architecture starts with a forward transformation of the topological relationships of connected nodes into edge or hyperedge features via a normalized projection matrix. The resulting edge/hyperedge features, together with the original node features, are fed into a neural network. The encoded node embeddings are then derived from the reversed transformation, described by the transpose of the projection matrix, of the network’s output, which can be further used for tasks such as node classification. The designed projection matrix, encoding the graph features, is intuitive and interpretable. The proposed architecture, in contrast to the traditional spectral-based and/or message passing approaches, simultaneously and comprehensively exploits the node features and graph/hypergraph topologies in an efficient and unified manner, covering both heterophilic and homophilic graphs. Furthermore, a variant version, UniG-Encoder II, is devised to leverage multi-hop node information. Extensive experiments are conducted and demonstrate the superior performance of the proposed framework on twelve representative hypergraph datasets and six real-world graph datasets, compared to the state-of-the-art methods. Our implementation is available online at https://github.com/MinhZou/UniG-Encoder.

Applying multi-layer perceptron and ResNet for handwritten digits recognition

The task of handwritten digit recognition is to recognizing the handwritten digits from pictures. Applying machine learning based models to automatically perform handwritten digit recognition task can significantly improve efficiency. This paper applies two machine learning based models, including multi-layer perceptron and residual neural network, for such a task. Firstly, this paper introduces the basic concept of the simple multi-layer perceptron model and then presents the structure of the residual neural network model. Subsequently, such two models are trained on the MNIST corpus, one of the classical dataset for the handwritten digit recognition task. The data pre-processing, like the splitting of training and test set, is described. Also, the processes of testing and training of the two models are presented. According to the experiments on the test set of MNIST, it is observed that the residual neural network can achieve better performance where the accuracy score is 99.240%, while the accuracy score of the multi-layer perceptron model is 97.260%.

Data-driven capacity estimation for lithium-ion batteries with feature matching based transfer learning method

Accurate capacity estimation is essential in the management of lithium-ion batteries, as it guarantees the safety and dependability of battery-powered systems. However, direct measurement of battery capacity is challenging due to the unpredictable working conditions and intricate electrochemical characteristics, which complicates the identification of battery degradation. In this work, through in-depth analysis of battery aging data, an incremental slope (IS) aided feature extraction method is proposed to obtain universal multidimensional features that adapt to different working conditions. With the extracted features, a simple multilayer perceptron (MLP) is used to achieve high-precision capacity estimation. Furthermore, a feature matching based transfer learning (FM-TL) method is proposed to automatically adapt the capacity estimation across different types of batteries that are cycled under various working conditions. 158 batteries covering five material types and 15 working conditions are used to validate the proposed method. Results suggest that the MLP model can provide an accurate capacity estimation, where the overall mean absolute percentage error (MAPE) and root mean square percentage error (RMSPE) are limited to 1.22% and 1.61%, respectively. Furthermore, compared with the traditional fine-tuning method, the overall MAPE and RMSPE under various transfer learning application scenarios respectively decrease by up to 78.23% and 75.31%, indicating that the FM-TL method is promising to construct a reliable transfer learning path, which improves the accuracy and reliability of capacity estimation when applied to various target domains.

Blind Quality Assessment of Images Containing Objects of Interest.

To monitor objects of interest, such as wildlife and people, image-capturing devices are used to collect a large number of images with and without objects of interest. As we are recording valuable information about the behavior and activity of objects, the quality of images containing objects of interest should be better than that of images without objects of interest, even if the former exhibits more severe distortion than the latter. However, according to current methods, quality assessments produce the opposite results. In this study, we propose an end-to-end model, named DETR-IQA (detection transformer image quality assessment), which extends the capability to perform object detection and blind image quality assessment (IQA) simultaneously by adding IQA heads comprising simple multi-layer perceptrons at the top of the DETRs (detection transformers) decoder. Using IQA heads, DETR-IQA carried out blind IQAs based on the weighted fusion of the distortion degree of the region of objects of interest and the other regions of the image; the predicted quality score of images containing objects of interest was generally greater than that of images without objects of interest. Currently, the subjective quality score of all public datasets is in accordance with the distortion of images and does not consider objects of interest. We manually extracted the images in which the five predefined classes of objects were the main contents of the largest authentic distortion dataset, KonIQ-10k, which was used as the experimental dataset. The experimental results show that with slight degradation in object detection performance and simple IQA heads, the values of PLCC and SRCC were 0.785 and 0.727, respectively, and exceeded those of some deep learning-based IQA models that are specially designed for only performing IQA. With the negligible increase in the computation and complexity of object detection and without a decrease in inference speeds, DETR-IQA can perform object detection and IQA via multi-tasking and substantially reduce the workload.

Retrieval of surface solar irradiance from satellite imagery using machine learning: pitfalls and perspectives

Abstract. Knowledge of the spatial and temporal characteristics of solar surface irradiance (SSI) is critical in many domains. While meteorological ground stations can provide accurate measurements of SSI locally, they are sparsely distributed worldwide. SSI estimations derived from satellite imagery are thus crucial to gain a finer understanding of the solar resource. Inferring SSI from satellite images is, however, not straightforward, and it has been the focus of many researchers in the past 30 to 40 years. For long, the emphasis has been on models grounded in physical laws with, in some cases, simple statistical parametrizations. Recently, new satellite SSI retrieval methods have been emerging, which directly infer the SSI from the satellite images using machine learning. Although only a few such works have been published, their practical efficiency has already been questioned. The objective of this paper is to better understand the potential and the pitfalls of this new family of methods. To do so, simple multi-layer-perceptron (MLP) models are constructed with different training datasets of satellite-based radiance measurements from Meteosat Second Generation (MSG) with collocated SSI ground measurements from Météo-France. The performance of the models is evaluated on a test dataset independent from the training set in both space and time and compared to that of a state-of-the-art physical retrieval model from the Copernicus Atmosphere Monitoring Service (CAMS). We found that the data-driven model's performance is very dependent on the training set. Provided the training set is sufficiently large and similar enough to the test set, even a simple MLP has a root mean square error (RMSE) that is 19 % lower than CAMS and outperforms the physical retrieval model at 96 % of the test stations. On the other hand, in certain configurations, the data-driven model can dramatically underperform even in stations located close to the training set: when geographical separation was enforced between the training and test set, the MLP-based model exhibited an RMSE that was 50 % to 100 % higher than that of CAMS in several locations.