Learning Bias Research Articles

A mathematical theory of relational generalization in transitive inference

Humans and animals routinely infer relations between different items or events and generalize these relations to novel combinations of items. This allows them to respond appropriately to radically novel circumstances and is fundamental to advanced cognition. However, how learning systems (including the brain) can implement the necessary inductive biases has been unclear. We investigated transitive inference (TI), a classic relational task paradigm in which subjects must learn a relation ([Formula: see text] and [Formula: see text]) and generalize it to new combinations of items ([Formula: see text]). Through mathematical analysis, we found that a broad range of biologically relevant learning models (e.g. gradient flow or ridge regression) perform TI successfully and recapitulate signature behavioral patterns long observed in living subjects. First, we found that models with item-wise additive representations automatically encode transitive relations. Second, for more general representations, a single scalar "conjunctivity factor" determines model behavior on TI and, further, the principle of norm minimization (a standard statistical inductive bias) enables models with fixed, partly conjunctive representations to generalize transitively. Finally, neural networks in the "rich regime," which enables representation learning and improves generalization on many tasks, unexpectedly show poor generalization and anomalous behavior on TI. We find that such networks implement a form of norm minimization (over hidden weights) that yields a local encoding mechanism lacking transitivity. Our findings show how minimal statistical learning principles give rise to a classical relational inductive bias (transitivity), explain empirically observed behaviors, and establish a formal approach to understanding the neural basis of relational abstraction.

Model inductive bias enhanced deep reinforcement learning for robot navigation in crowded environments

Navigating mobile robots in crowded environments poses a significant challenge and is essential for the coexistence of robots and humans in future intelligent societies. As a pragmatic data-driven approach, deep reinforcement learning (DRL) holds promise for addressing this challenge. However, current DRL-based navigation methods have possible improvements in understanding agent interactions, feedback mechanism design, and decision foresight in dynamic environments. This paper introduces the model inductive bias enhanced deep reinforcement learning (MIBE-DRL) method, drawing inspiration from a fusion of data-driven and model-driven techniques. MIBE-DRL extensively incorporates model inductive bias into the deep reinforcement learning framework, enhancing the efficiency and safety of robot navigation. The proposed approach entails a multi-interaction network featuring three modules designed to comprehensively understand potential agent interactions in dynamic environments. The pedestrian interaction module can model interactions among humans, while the temporal and spatial interaction modules consider agent interactions in both temporal and spatial dimensions. Additionally, the paper constructs a reward system that fully accounts for the robot’s direction and position factors. This system's directional and positional reward functions are built based on artificial potential fields (APF) and navigation rules, respectively, which can provide reasoned evaluations for the robot's motion direction and position during training, enabling it to receive comprehensive feedback. Furthermore, the incorporation of Monte-Carlo tree search (MCTS) facilitates the development of a foresighted action strategy, enabling robots to execute actions with long-term planning considerations. Experimental results demonstrate that integrating model inductive bias significantly enhances the navigation performance of MIBE-DRL. Compared to state-of-the-art methods, MIBE-DRL achieves the highest success rate in crowded environments and demonstrates advantages in navigation time and maintaining a safe social distance from humans.

Performance of off-the-shelf machine learning architectures and biases in low left ventricular ejection fraction detection

Artificial intelligence-machine learning (AI-ML) has demonstrated the ability to extract clinically useful information from electrocardiograms (ECGs) not available using traditional interpretation methods. There exists an extensive body of AI-ML research in fields outside of cardiology including several open-source AI-ML architectures that can be translated to new problems in an "off-the-shelf" manner. We sought to address the limited investigation of which if any of these off-the-shelf architectures could be useful in ECG analysis as well as how and when these AI-ML approaches fail. We applied 6 off-the-shelf AI-ML architectures to detect low left ventricular ejection fraction (LVEF) in a cohort of ECGs from 24,868 patients. We assessed LVEF classification and explored patient characteristics associated with inaccurate (false positive or false negative) LVEF prediction. We found that all of these network architectures produced LVEF detection area under the receiver-operating characteristic curve values above 0.9 (averaged over 5 instances per network), with the ResNet 18 network performing the highest (average area under the receiver-operating characteristic curve of 0.917). We also observed that some patient-specific characteristics such as race, sex, and presence of several comorbidities were associated with lower LVEF prediction performance. This demonstrates the ability of off-the-shelf AI-ML architectures to detect clinically useful information from ECGs with performance matching contemporary custom-build AI-ML architectures. We also highlighted the presence of possible biases in these AI-ML approaches in the context of patient characteristics. These findings should be considered in the pursuit of efficient and equitable deployment of AI-ML technologies moving forward.

MobileDepth: Monocular Depth Estimation Based on Lightweight Vision Transformer

ABSTRACT As deep learning takes off, monocular depth estimation based on convolutional neural networks (CNNs) has made impressive progress. CNNs are superior at extracting local characteristics from a single image; however, they are unable to manage long-range dependence and thus have a substantial impact on the performance of monocular depth estimation. In addition to this, as architectures based on CNNs frequently utilize down sampling operations, numbers of pixel-level features, which are extremely crucial for dense prediction tasks, are lost in the encoder phase. Unlike CNNs, ViT is capable of capturing global feature information, but it requires numbers of parameters and data augmentation owing to its lack of inductive bias. To address the aforementioned difficulties, in this study, we propose a Dilated Self Attention Block (DSAB) as well as a Local and Global Feature Extraction (LGFE) module. The former resolves the inference speed issue of standard ViT models, and we accomplish this by limiting the number of self-attention computations among tokens. The latter combines the advantages of CNNs and ViT, first extracting local representation information in low-dimensional space through standard convolution and then mapping the input tensor to high-dimensional space to capture global information, achieving the simultaneous extraction of global and local characteristics.

Dense Continuous-Time Optical Flow From Event Cameras.

We present a method for estimating dense continuous-time optical flow from event data. Traditional dense optical flow methods compute the pixel displacement between two images. Due to missing information, these approaches cannot recover the pixel trajectories in the blind time between two images. In this work, we show that it is possible to compute per-pixel, continuous-time optical flow using events from an event camera. Events provide temporally fine-grained information about movement in pixel space due to their asynchronous nature and microsecond response time. We leverage these benefits to predict pixel trajectories densely in continuous time via parameterized Bézier curves. To achieve this, we build a neural network with strong inductive biases for this task: First, we build multiple sequential correlation volumes in time using event data. Second, we use Bézier curves to index these correlation volumes at multiple timestamps along the trajectory. Third, we use the retrieved correlation to update the Bézier curve representations iteratively. Our method can optionally include image pairs to boost performance further. To the best of our knowledge, our model is the first method that can regress dense pixel trajectories from event data. To train and evaluate our model, we introduce a synthetic dataset (MultiFlow) that features moving objects and ground truth trajectories for every pixel. Our quantitative experiments not only suggest that our method successfully predicts pixel trajectories in continuous time but also that it is competitive in the traditional two-view pixel displacement metric on MultiFlow and DSEC-Flow. Open source code and datasets are released to the public.

Physics-informed surrogate modeling for a damaged rotating shaft

In the online structural health monitoring framework, surrogate modeling aims to approximate the predictions of the underlying high-fidelity model to facilitate fast simulations. Generally, conventional machine learning techniques based on purely data-driven approaches often behave as black boxes, producing predictions that may lack physical consistency. In an effort to overcome these limitations, we proposed a neural network mapping constrained by the known physical relationships of an engineering model, which can be considered as a physics-informed inductive bias. This approach aims to enhance computational efficiency while reducing the required number of training examples. The physics-informed term guides the training process to yield predictions that exhibit greater physical consistency while maintaining high performance. We here propose a surrogate modeling approach relying on a Physics-Informed Neural Network (PINN) scheme, which is designed to enable robust and real-time monitoring of ballistic impact damage on a rotor shaft. At first, a high fidelity finite element model is developed to simulate the response of the rotor shaft under various operational conditions. As a first reduction step, substructuring is used to decrease the dimensionality of the system's matrices, thereby reducing the computational cost. For comparison and validation, an alternative model based on the Timoshenko beam theory is utilized. Both models examine the effects of impact-induced vibration loads due to ballistic occurrences. The PINN model receives the external load as input and predicts the system's response. The system matrices extracted from both models constitute the physics-informed term. This term is incorporated into the PINN architecture in order to steer the learning process towards more physically consistent and reliable predictions regarding the state and the response of the structure. The algorithm's performance is evaluated using response data derived from both models.

Transfer of statistical learning between tasks.

Recent studies have shown that observers can learn to suppress locations in the visual field with a high distractor probability. Here, we investigated whether this learned suppression resulting from a spatial distractor imbalance transfers to a completely different search task that does not contain any distractors. Observers performed the additional singleton task and learned to suppress the location that was likely to contain a color singleton distractor. Within a block, the additional singleton task would randomly switch to a T-among-L task where observers searched in parallel (Experiment 1) or serially (Experiment 2) for a T among Ls. The upcoming search was either unpredictable (Experiment 1/2A) or cued (Experiment 1/2B). The results show that there was transfer of learning from one to the other task as the learned suppression stayed in place after the switch regardless of whether the T-among-L task was performed via parallel or serial search. Moreover, cueing that the task would switch had no effect on performance. The current findings indicate that implicit learned biases are rather inflexible and remain in place even when the task and the required search strategy are dramatically different and even when participants can anticipate that a change in the search required is imminent. This transfer of the suppression to a different task is consistent with the notion that suppression is proactively applied. Because the location is already suppressed proactively, that is, before display onset, regardless which display and task is presented, the suppressed location competes less for attention than all other locations. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

A Scene Classification Model Based on Global-Local Features and Attention in Lie Group Space

The efficient fusion of global and local multi-scale features is quite important for remote sensing scene classification (RSSC). The scenes in high-resolution remote sensing images (HRRSI) contain many complex backgrounds, intra-class diversity, and inter-class similarities. Many studies have shown that global features and local features are helpful for RSSC. The receptive field of a traditional convolution kernel is small and fixed, and it is difficult to capture global features in the scene. The self-attention mechanism proposed in transformer effectively alleviates the above shortcomings. However, such models lack local inductive bias, and the calculation is complicated due to the large number of parameters. To address these problems, in this study, we propose a classification model of global-local features and attention based on Lie Group space. The model is mainly composed of three independent branches, which can effectively extract multi-scale features of the scene and fuse the above features through a fusion module. Channel attention and spatial attention are designed in the fusion module, which can effectively enhance the crucial features in the crucial regions, to improve the accuracy of scene classification. The advantage of our model is that it extracts richer features, and the global-local features of the scene can be effectively extracted at different scales. Our proposed model has been verified on publicly available and challenging datasets, taking the AID as an example, the classification accuracy reached 97.31%, and the number of parameters is 12.216 M. Compared with other state-of-the-art models, it has certain advantages in terms of classification accuracy and number of parameters.

What makes a stimulus worthy of attention: Cue-outcome correlation and choice relevance in the learned predictiveness effect.

The learned predictiveness effect refers to the tendency for predictive cues to attract greater attention and show faster learning in subsequent tasks. However, in typical designs, the predictiveness of each cue (its objective cue-outcome correlation) is confounded with the degree to which it is informative for making the correct response on each trial (a feature we term choice relevance). In four experiments, we tested the unique contributions of cue-outcome correlation and choice relevance to the learned predictiveness effect by manipulating the outcome choices available on each trial. Experiments 1A and 1B compared two sets of partially predictive cues and found that participants learned more in a transfer phase about the set of cues that were previously choice-relevant. Experiments 2A and 2B used a design in which the cue-outcome correlation was stronger for one set of cues (perfect predictors) than the other set (imperfect predictors). Manipulating the choice relevance of the imperfect predictors in this design did not influence the magnitude of the learning bias toward the perfect predictor. Unlike cue-outcome correlation, choice relevance did not seem to correspond to biases in eye-gaze, suggesting that it operates via a distinct mechanism. Simulations with a modified EXIT model successfully predicted cue-outcome correlation and choice relevance effects by assuming that participants update learning for present outcomes only, but incorrectly predicted additive effects. We conclude that cue-outcome correlation and choice relevance are important factors that can lead to biases in future learning; both were individually sufficient but neither was necessary. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

Non-autoregressive personalized bundle generation

The personalized bundle generation problem, which aims to create a preferred bundle for user from numerous candidate items, receives increasing attention in recommendation. However, existing works ignore the order-invariant nature of the bundle and adopt sequential modeling methods as the solution, which might introduce inductive bias and cause a large latency in prediction. To address this problem, we propose to perform the bundle generation via non-autoregressive mechanism and design a novel encoder–decoder framework named BundleNAT, which can effectively output the targeted bundle in one-shot without relying on any inherent order. In detail, instead of learning sequential dependency, we propose to adopt pre-training techniques and graph neural network to fully embed user-based preference and item-based compatibility information, and use a self-attention based encoder to further extract global dependency pattern. We then design a permutation-equivariant decoding architecture that is able to directly output the desired bundle in a one-shot manner. Experiments on three real-world datasets from Youshu and Netease show the proposed BundleNAT significantly outperforms the current state-of-the-art methods in average by up to 35.92%, 10.97% and 23.67% absolute improvements in Precision, Precision+, and Recall, respectively.

Multi-scale sampling attention graph convolutional networks for skeleton-based action recognition

Skeleton-based action recognition has attracted increasing interest in recent years. With the flexibility of modeling long-range dependency of joints, the self-attention module has served as the basic component in skeleton-based action recognition. However, the global receptive field of self-attention is not conducive to the modeling of skeleton locality, and the self-attention model is imbued with less inductive bias, which leads to overfitting. In this paper, we propose an attention graph convolutional network (AGCN) with multi-scale sampling to effectively model the local and global features of the skeleton. Firstly, we propose two extreme sampling strategies for generating and ordering neighboring nodes of root nodes. A local-first sampling method is introduced to construct local graph windows, and a global-first sampling method is proposed to assemble long-range joints for constructing global graph windows. The local-first sampling and global-first sampling introduce more skeleton-specific inductive biases to regularize the model capacity. Secondly, the AGCN combines the self-attention mechanism with graph convolution operation, which alleviates the over-smoothing of graph convolution and preserves the translation invariant. Based on the multi-scale sampling strategy, the AGCN can effectively model the locality and non-locality of the skeleton. Finally, by coupling the aforementioned proposals, we develop a two-pathway model for multi-scale feature fusion. Extensive experiments demonstrate that our model could achieve comparable performance with state-of-the-art works on the NTU RGB+D 60, NTU RGB+D 120, the UAV-HUMAN and NW-UCLA datasets.

A novel physics-aware graph network using high-order numerical methods in weather forecasting model

In this paper, we present a novel architecture for accurate weather forecasting within the framework of Physics-aware Graph Networks (PaGN), which aims to improve the prediction of weather time series by integrating both data and physical equations defined on sparsely distributed spatial domains. This approach employs geometric deep learning by accounting for both supervised loss and physics loss. Solely in supervised learning, data-driven training demands a large amount of data, whereas integrating physics information can reduce this requirement, resulting in faster training. The physics awareness also provides a stronger inductive bias and better generalization, directing the machine learning process by clarifying the complexity of weather data. Thus, a more accurate discretization of physics equations can enhance the physics awareness of the complex weather dynamics. Meanwhile, weather data is essentially gathered at observatories that are not located densely, since covering the Earth with numerous observatories is prohibitively expensive. Given its reliance on graph structure, the proposed PaGN is well-suited for handling sparsely distributed data particularly defined on arbitrary geometric domain. One of the main contributions of this study is the use of high-order numerical methods approximating physics equations defined on graph neural networks for achieving accuracy improvements. The fractional graph Laplacian operator is further integrated with the physics equations to account for the non-local characteristics of complex weather data. For computational efficiency, the accuracy improved PaGN architecture is smoothly implemented in the embedding space. A series of numerical tests on weather datasets is performed to verify the improved accuracy, robustness, and applicability of the proposed methodology. We first carry out the numerical study on the synthetic datasets obtained from the physics equations with an extra forcing term. We then extensively investigate the accuracies of weather forecasting for various configurations in terms of the geophysical regions and the accuracy-improved PaGN approaches for diffusion and wave equations. Moreover, we emphasize that the fractional graph Laplacian operator incorporated into our PaGN model can further improve the accuracy for weather forecasting.

Using games to understand the mind.

Board, card or video games have been played by virtually every individual in the world. Games are popular because they are intuitive and fun. These distinctive qualities of games also make them ideal for studying the mind. By being intuitive, games provide a unique vantage point for understanding the inductive biases that support behaviour in more complex, ecological settings than traditional laboratory experiments. By being fun, games allow researchers to study new questions in cognition such as the meaning of 'play' and intrinsic motivation, while also supporting more extensive and diverse data collection by attracting many more participants. We describe the advantages and drawbacks of using games relative to standard laboratory-based experiments and lay out a set of recommendations on how to gain the most from using games to study cognition. We hope this Perspective will lead to a wider use of games as experimental paradigms, elevating the ecological validity, scale and robustness of research on the mind.

Towards transferable metamodels for water distribution systems with edge-based graph neural networks

Data-driven metamodels reproduce the input-output mapping of physics-based models while significantly reducing simulation times. Such techniques are widely used in the design, control, and optimization of water distribution systems. Recent research highlights the potential of metamodels based on Graph Neural Networks as they efficiently leverage graph-structured characteristics of water distribution systems. Furthermore, these metamodels possess inductive biases that facilitate generalization to unseen topologies. Transferable metamodels are particularly advantageous for problems that require an efficient evaluation of many alternative layouts or when training data is scarce. However, the transferability of metamodels based on GNNs remains limited, due to the lack of representation of physical processes that occur on edge level, i.e. pipes. To address this limitation, our work introduces Edge-Based Graph Neural Networks, which extend the set of inductive biases and represent link-level processes in more detail than traditional Graph Neural Networks. Such an architecture is theoretically related to the constraints of mass conservation at the junctions. To verify our approach, we test the suitability of the edge-based network to estimate pipe flowrates and nodal pressures emulating steady-state EPANET simulations. We first compare the effectiveness of the metamodels on several benchmark water distribution systems against Graph Neural Networks. Then, we explore transferability by evaluating the performance on unseen systems. For each configuration, we calculate model performance metrics, such as coefficient of determination and speed-up with respect to the original numerical model. Our results show that the proposed method captures the pipe-level physical processes more accurately than node-based models. When tested on unseen water networks with a similar distribution of demands, our model retains a good generalization performance with a coefficient of determination of up to 0.98 for flowrates and up to 0.95 for predicted heads. Further developments could include simultaneous derivation of pressures and flowrates.

FairGAT: Fairness-Aware Graph Attention Networks

Graphs can facilitate modeling various complex systems such as gene networks and power grids as well as analyzing the underlying relations within them. Learning over graphs has recently attracted increasing attention, particularly graph neural network (GNN)–based solutions, among which graph attention networks (GATs) have become one of the most widely utilized neural network structures for graph-based tasks. Although it is shown that the use of graph structures in learning results in the amplification of algorithmic bias, the influence of the attention design in GATs on algorithmic bias has not been investigated. Motivated by this, the present study first carries out a theoretical analysis in order to demonstrate the sources of algorithmic bias in GAT-based learning for node classification. Then, a novel algorithm, FairGAT, which leverages a fairness-aware attention design, is developed based on the theoretical findings. Experimental results on real-world networks demonstrate that FairGAT improves group fairness measures while also providing comparable utility to the fairness-aware baselines for node classification and link prediction.

Biological direct-shortcut deep residual learning for sparse visual processing

We show, based on the following three grounds, that the primary visual cortex (V1) is a biological direct-shortcut deep residual learning neural network (ResNet) for sparse visual processing: (1) We first highlight that Gabor-like sets of basis functions, which are similar to the receptive fields of simple cells in the primary visual cortex (V1), are excellent candidates for sparse representation of natural images; i.e., images from the natural world, affirming the brain to be optimized for this. (2) We then prove that the intra-layer synaptic weight matrices of this region can be reasonably first-order approximated by identity mappings, and are thus sparse themselves. (3) Finally, we point out that intra-layer weight matrices having identity mapping as their initial approximation, irrespective of this approximation being also a reasonable first-order one or not, resemble the building blocks of direct-shortcut digital ResNets, which completes the grounds. This biological ResNet interconnects the sparsity of the final representation of the image to that of its intra-layer weights. Further exploration of this ResNet, and understanding the joint effects of its architecture and learning rules, e.g. on its inductive bias, could lead to major advancements in the area of bio-inspired digital ResNets. One immediate line of research in this context, for instance, is to study the impact of forcing the direct-shortcuts to be good first-order approximations of each building block. For this, along with the ℓ1\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$${{\\ell}}_{1}$$\\end{document}-minimization posed on the basis function coefficients the ResNet finally provides at its output, another parallel one could e.g. also be posed on the weights of its residual layers.

Embed systemic equity throughout industrial ecology applications: How to address machine learning unfairness and bias

AbstractRecent calls have been made for equity tools and frameworks to be integrated throughout the research and design life cycle —from conception to implementation—with an emphasis on reducing inequity in artificial intelligence (AI) and machine learning (ML) applications. Simply stating that equity should be integrated throughout, however, leaves much to be desired as industrial ecology (IE) researchers, practitioners, and decision‐makers attempt to employ equitable practices. In this forum piece, we use a critical review approach to explain how socioecological inequities emerge in ML applications across their life cycle stages by leveraging the food system. We exemplify the use of a comprehensive questionnaire to delineate unfair ML bias across data bias, algorithmic bias, and selection and deployment bias categories. Finally, we provide consolidated guidance and tailored strategies to help address AI/ML unfair bias and inequity in IE applications. Specifically, the guidance and tools help to address sensitivity, reliability, and uncertainty challenges. There is also discussion on how bias and inequity in AI/ML affect other IE research and design domains, besides the food system—such as living labs and circularity. We conclude with an explanation of the future directions IE should take to address unfair bias and inequity in AI/ML. Last, we call for systemic equity to be embedded throughout IE applications to fundamentally understand domain‐specific socioecological inequities, identify potential unfairness in ML, and select mitigation strategies in a manner that translates across different research domains.

Deep learning bias correction of GEMS tropospheric NO2: A comparative validation of NO2 from GEMS and TROPOMI using Pandora observations

Despite advancements in satellite instruments, such as those in geostationary orbit, biases continue to affect the accuracy of satellite data. This research pioneers the use of a deep convolutional neural network to correct bias in tropospheric column density of NO2 (TCDNO2) from the Geostationary Environment Monitoring Spectrometer (GEMS) during 2021–2023. Initially, we validate GEMS TCDNO2 against Pandora observations and compare its accuracy with measurements from the TROPOspheric Monitoring Instrument (TROPOMI). GEMS displays acceptable accuracy in TCDNO2 measurements, with a correlation coefficient (R) of 0.68, an index of agreement (IOA) of 0.79, and a mean absolute bias (MAB) of 5.73321 × 1015 molecules/cm2, though it is not highly accurate. The evaluation showcases moderate to high accuracy of GEMS TCDNO2 across all Pandora stations, with R values spanning from 0.46 to 0.80. Comparing TCDNO2 from GEMS and TROPOMI at TROPOMI overpass time shows satisfactory performance of GEMS TCDNO2 measurements, achieving R, IOA, and MAB values of 0.71, 0.78, and 6.82182 × 1015 molecules/cm2, respectively. However, these figures are overshadowed by TROPOMI’s superior accuracy, which reports R, IOA, and MAB values of 0.81, 0.89, and 3.26769 × 1015 molecules/cm2, respectively. While GEMS overestimates TCDNO2 by 52 % at TROPOMI overpass time, TROPOMI underestimates it by 9 %. The deep learning bias corrected GEMS TCDNO2 (GEMS-DL) demonstrates a marked enhancement in the accuracy of original GEMS TCDNO2 measurements. The GEMS-DL product improves R from 0.68 to 0.88, IOA from 0.79 to 0.93, MAB from 5.73321 × 1015 to 2.67659 × 1015 molecules/cm2, and reduces MAB percentage (MABP) from 64 % to 30 %. This represents a significant reduction in bias, exceeding 50 %. Although the original GEMS product overestimates TCDNO2 by 28 %, the GEMS-DL product remarkably minimizes this error, underestimating TCDNO2 by a mere 1 %. Spatial cross-validation across Pandora stations shows a significant reduction in MABP, from a range of 45 %-105.6 % in original GEMS data to 24 %-59 % in GEMS-DL.

Mitigating machine learning bias between high income and low–middle income countries for enhanced model fairness and generalizability

Collaborative efforts in artificial intelligence (AI) are increasingly common between high-income countries (HICs) and low- to middle-income countries (LMICs). Given the resource limitations often encountered by LMICs, collaboration becomes crucial for pooling resources, expertise, and knowledge. Despite the apparent advantages, ensuring the fairness and equity of these collaborative models is essential, especially considering the distinct differences between LMIC and HIC hospitals. In this study, we show that collaborative AI approaches can lead to divergent performance outcomes across HIC and LMIC settings, particularly in the presence of data imbalances. Through a real-world COVID-19 screening case study, we demonstrate that implementing algorithmic-level bias mitigation methods significantly improves outcome fairness between HIC and LMIC sites while maintaining high diagnostic sensitivity. We compare our results against previous benchmarks, utilizing datasets from four independent United Kingdom Hospitals and one Vietnamese hospital, representing HIC and LMIC settings, respectively.

HTC-retina: A hybrid retinal diseases classification model using transformer-Convolutional Neural Network from optical coherence tomography images

Retinal diseases are among nowadays major public health issues, deservedly needing advanced computer-aided diagnosis. We propose a hybrid model for multi label classification, whereby seven retinal diseases are automatically classified from Optical Coherence Tomography (OCT) images. We show that, by combining the strengths of Convolutional Neural Networks (CNNs) and Visual Transformers (ViTs), we can produce a more powerful type of model for medical image classification, especially when considering local lesion information such as retinal diseases. CNNs are indeed proved to be efficient at parameter utilization and provide the ability to extract local features and multi-scale feature maps through convolutional operations. On the other hand, ViT’s self-attention procedure allows processing long-range and global dependencies within an image. The paper clearly shows that the hybridization of these complementary capabilities (CNNs-ViTs) presents a high image processing potential that is more robust and efficient. The proposed model adopts a hierarchical CNN module called Convolutional Patch and Token Embedding (CPTE) instead of employing a direct tokenization approach using the raw input OCT image in the transformer. The CPTE module’s role is to incorporate an inductive bias, to reduce the reliance on large-scale datasets, and to address the low-level feature extraction challenges of the ViT. In addition, considering the importance of local lesion information in OCT images, the model relies on a parallel module called Residual Depthwise-Pointwise ConvNet (RDP-ConvNet) for extracting high-level features. RDP-ConvNet utilizes depthwise and pointwise convolution layers within a residual network architecture. The overall performance of the HTC-Retina model was evaluated on three datasets: the OCT-2017, OCT-C8, and OCT-2014 ; outperforming previous established models, achieving accuracy rates of 99.40%, 97.00%, and 99.77%, respectively ; and sensitivity rates of 99.41%, 97.00%, and 99.77%, respectively. Notably, the model showed high performance while maintaining computational efficiency.