Papers

Learning and generalizing from limited examples, i.e., few-shot learning, is of core importance to many real-world vision applications. A principal way of achieving few-shot learning is to realize an embedding where samples from different classes are distinctive. Recent studies suggest that embedding via hyperbolic geometry enjoys low distortion for hierarchical and structured data, making it suitable for few-shot learning. In this paper, we propose to learn a context-aware hyperbolic metric to characterize the distance between a point and a set associated with a learned set to set distance. To this end, we formulate the metric as a weighted sum on the tangent bundle of the hyperbolic space and develop a mechanism to obtain the weights adaptively, based on the constellation of the points. This not only makes the metric local but also dependent on the task in hand, meaning that the metric will adapt depending on the samples that it compares. We empirically show that such metric yields robustness in the presence of outliers and achieves a tangible improvement over baseline models. This includes the state-of-the-art results on five popular few-shot classification benchmarks, namely mini-ImageNet, tiered-ImageNet, Caltech-UCSD Birds-200-2011(CUB), CIFAR-FS, and FC100.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

Recent work has shown the potential benefit of selective prediction systems that can learn to defer to a human when the predictions of the AI are unreliable, particularly to improve the reliability of AI systems in high-stakes applications like healthcare or conservation. However, most prior work assumes that human behavior remains unchanged when they solve a prediction task as part of a human-AI team as opposed to by themselves. We show that this is not the case by performing experiments to quantify human-AI interaction in the context of selective prediction. In particular, we study the impact of communicating different types of information to humans about the AI system's decision to defer. Using real-world conservation data and a selective prediction system that improves expected accuracy over that of the human or AI system working individually, we show that this messaging has a significant impact on the accuracy of human judgements. Our results study two components of the messaging strategy: 1) Whether humans are informed about the prediction of the AI system and 2) Whether they are informed about the decision of the selective prediction system to defer. By manipulating these messaging components, we show that it is possible to significantly boost human performance by informing the human of the decision to defer, but not revealing the prediction of the AI. We therefore show that it is vital to consider how the decision to defer is communicated to a human when designing selective prediction systems, and that the composite accuracy of a human-AI team must be carefully evaluated using a human-in-the-loop framework.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

Features, logits, and labels are the three primary data when a sample passes through a deep neural network. Feature perturbation and label perturbation receive increasing attention in recent years. They have been proven to be useful in various deep learning approaches. For example, (adversarial) feature perturbation can improve the robustness or even generalization capability of learned models. However, limited studies have explicitly explored for the perturbation of logit vectors. This work discusses several existing methods related to logit perturbation. Based on a unified viewpoint between positive/negative data augmentation and loss variations incurred by logit perturbation, a new method is proposed to explicitly learn to perturb logits. A comparative analysis is conducted for the perturbations used in our and existing methods. Extensive experiments on benchmark image classification data sets and their long-tail versions indicated the competitive performance of our learning method. In addition, existing methods can be further improved by utilizing our method.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

Large deformations of organs, caused by diverse shapes and nonlinear shape changes, pose a significant challenge for medical image registration. Traditional registration methods need to iteratively optimize an objective function via a specific deformation model along with meticulous parameter tuning, but which have limited capabilities in registering images with large deformations. While deep learning-based methods can learn the complex mapping from input images to their respective deformation field, it is regression-based and is prone to be stuck at local minima, particularly when large deformations are involved. To this end, we present Stochastic Planner-Actor-Critic (spac), a novel reinforcement learning-based framework that performs step-wise registration. The key notion is warping a moving image successively by each time step to finally align to a fixed image. Considering that it is challenging to handle high dimensional continuous action and state spaces in the conventional reinforcement learning (RL) framework, we introduce a new concept `Plan' to the standard Actor-Critic model, which is of low dimension and can facilitate the actor to generate a tractable high dimensional action. The entire framework is based on unsupervised training and operates in an end-to-end manner. We evaluate our method on several 2D and 3D medical image datasets, some of which contain large deformations. Our empirical results highlight that our work achieves consistent, significant gains and outperforms state-of-the-art methods.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

Image-based survival prediction models can facilitate doctors in diagnosing and treating cancer patients. With the advance of digital pathology technologies, the big whole slide images (WSIs) provide increasing resolution and more details for diagnosis. However, the gigabyte-size WSIs would make most models computationally infeasible. To this end, instead of using the complete WSIs, most of existing models only use a pre-selected subset of key patches or patch clusters as input, which might fail to completely capture the patient's tumor morphology. In this work, we aim to develop a novel survival analysis model to fully utilize the complete WSI information. We show that the use of a Vision Transformer (ViT) backbone, together with convolution operations involved in it, is an effective framework to improve the prediction performance. Additionally, we present a post-hoc explainable method to identify the most salient patches and distinct morphology features, making the model more faithful and the results easier to comprehend by human users. Evaluations on two large cancer datasets show that our proposed model is more effective and has better interpretability for survival prediction.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

Conventional saliency prediction models typically learn a deterministic mapping from images to the corresponding ground truth saliency maps. In this paper, we study the saliency prediction problem from the perspective of generative models by learning a conditional probability distribution over saliency maps given an image, and treating the saliency prediction as a sampling process. Specifically, we propose a generative cooperative saliency prediction framework, where a conditional latent variable model and a conditional energy-based model are jointly trained to predict saliency in a cooperative manner. The latent variable model serves as a fast but coarse predictor to efficiently produce an initial prediction, which is then refined by the iterative Langevin revision of the energy-based model that serves as a fine predictor. Such a coarse-to-fine cooperative saliency prediction strategy offers the best of both worlds. We further generalize our framework to weakly supervised saliency prediction with a cooperative learning while recovering strategy. Experimental results show that our generative model achieves both state-of-the-art performance and reliable sampling processing exploring.

poster session 2 @ red 2, poster session 9 @ red 2, oral session 9 @ red 2, poster session 2, poster session 9, oral session 9

Due to the scarcity of annotated samples, the diversity between support set and query set becomes the main obstacle for few shot semantic segmentation. Most existing prototype-based approaches only exploit the prototype from the support feature and ignore the information from the query sample, failing to remove this obstacle.In this paper, we proposes a dual prototype network (DPNet) to dispose of few shot semantic segmentation from a new perspective. Along with the prototype extracted from the support set, we propose to build the pseudo-prototype based on foreground features in the query image. To achieve this goal, the cycle comparison module is developed to select reliable foreground features and generate the pseudo-prototype with them. Then, a prototype interaction module is utilized to integrate the information of the prototype and the pseudo-prototype based on their underlying correlation. Finally, a multi-scale fusion module is introduced to capture contextual information during the dense comparison between prototype (pseudo-prototype) and query feature. Extensive experiments conducted on two benchmarks demonstrate that our method exceeds previous state-of-the-arts with a sizable margin, verifying the effectiveness of the proposed method.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

The developmental approach, simulating a cognitive development of a human, arises as a way to nurture a human-level commonsense and overcome the limitations of data-driven approaches. However, neither a virtual environment nor an evaluation platform exists for the overall development of core cognitive skills. We present the VECA(Virtual Environment for Cognitive Assessment), which consists of two main components: (i) a first benchmark to assess the overall cognitive development of an AI agent, and (ii) a novel toolkit to generate diverse and distinct cognitive tasks. VECA benchmark virtually implements the cognitive scale of Bayley Scales of Infant and Toddler Development-IV(Bayley-4), the gold-standard developmental assessment for human infants and toddlers. Our VECA toolkit provides a human toddler-like embodied agent with various human-like perceptual features crucial to human cognitive development, e.g., binocular vision, 3D-spatial audio, and tactile receptors. We compare several modern RL algorithms on our VECA benchmark and seek their limitations in modeling human-like cognitive development. We further analyze the validity of the VECA benchmark, as well as the effect of human-like sensory characteristics on cognitive skills.

poster session 2 @ red 2, poster session 9 @ red 2, oral session 2 @ red 2, poster session 2, poster session 9, oral session 2

Spiking Neural Networks (SNNs) have been attached great importance due to the distinctive properties of low power consumption, biological plausibility, and adversarial robustness. The most effective way to train deep SNNs is through ANN-to-SNN conversion, which have yielded the best performance in deep network structure and large-scale datasets. However, there is a trade-off between accuracy and latency. In order to achieve high precision as original ANNs, a long simulation time is needed to match the firing rate of a spiking neuron with the activation value of an analog neuron, which impedes the practical application of SNN. In this paper, we aim to achieve high-performance converted SNNs with extremely low latency (fewer than 32 time-steps). We start by theoretically analyzing ANN-to-SNN conversion and show that scaling the thresholds does play a similar role as weight normalization. Instead of introducing constraints that facilitate ANN-to-SNN conversion at the cost of model capacity, we applied a more direct way by optimizing the initial membrane potential to reduce the conversion loss in each layer. Besides, we demonstrate that optimal initialization of membrane potentials can implement expected error-free ANN-to-SNN conversion. We evaluate our algorithm on the CIFAR-10 dataset and CIFAR-100 dataset and achieve state-of-the-art accuracy, using fewer time-steps. For example, we reach top-1 accuracy of 93.38% on CIFAR-10 with 16 time-steps. Moreover, our method can be applied to other ANN-SNN conversion methodologies and remarkably promote performance when the time-steps is small.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

Interactive recommendation with natural-language feedback can provide richer user feedback and has demonstrated advantages over traditional recommender systems. However, the classical online paradigm involves iteratively collecting experience via interaction with users, which is expensive and risky. We propose an offline interactive recommendation to exploit arbitrary experience collected by multiple unknown devices (policies). Direct policy learning with such fixed experience suffers from the distribution shift. To tackle this issue, we leverage a neural correction estimator via adversarial training, which can be seamlessly incorporated into the standard RL objective. Empirical results on real-world datasets demonstrate the effectiveness of our proposed framework.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

As an essential step in the pathological diagnosis, histochemical staining can show specific tissue structure information and, consequently, assist pathologists in making accurate diagnoses. Clinical kidney histopathological analyses usually employ more than one type of staining: H&E, MAS, PAS, PASM, etc. However, due to the interference of colors among multiple stains, it is not easy to perform multiple staining simultaneously on one biological tissue. To address this problem, we propose a network based on unpaired training data to virtually generate multiple types of staining from one staining. Our method can preserve the content of input images while transferring them to multiple target styles accurately. To efficiently control the direction of stain transfer, we propose a style guided normalization (SGN). Furthermore, a multiple style encoding (MSE) is devised to represent the relationship among different staining styles dynamically. An improved one-hot label is also proposed to enhance the generalization ability and extendibility of our method. Vast experiments have demonstrated that our model can achieve superior performance on a tiny dataset. The results exhibit not only good performance but also great visualization and interpretability. Especially, our method also achieves satisfactory results over cross-tissue, cross-staining as well as cross-task. We believe that our method will significantly influence clinical stain transfer and reduce the workload greatly for pathologists.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

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

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

When humans collaborate with each other, they often make decisions by observing others and considering the consequences that their actions may have on the entire team, instead of greedily doing what is best for just themselves. We would like our AI agents to effectively collaborate in a similar way by capturing a model of their partners. In this work, we propose and analyze a decentralized Multi-Armed Bandit (MAB) problem with coupled rewards as an abstraction of more general multi-agent collaboration. We demonstrate that naive extensions of single-agent optimal MAB algorithms fail when applied for decentralized bandit teams. Instead, we propose a Partner-Aware strategy for joint sequential decision-making that extends the well-known single-agent Upper Confidence Bound algorithm. We analytically show that our proposed strategy achieves logarithmic regret, and provide extensive experiments involving human-AI and human-robot collaboration to validate our theoretical findings. Our results show that the proposed partner-aware strategy outperforms other known methods, and our human subject studies suggest humans prefer to collaborate with AI agents implementing our partner-aware strategy.

poster session 2 @ red 2, poster session 9 @ red 2, oral session 2 @ red 2, poster session 2, poster session 9, oral session 2

Human emotions involve basic and compound facial expressions. However, current research on facial expression recognition (FER) mainly focuses on basic expressions, and thus fails to address the diversity of human emotions in practical scenarios. Meanwhile, existing work on compound FER relies heavily on abundant labeled compound expression training data, which are often laboriously collected under the professional instruction of psychology. In this paper, we study compound FER in the cross-domain few-shot learning setting, where only a few images of novel classes from the target domain are required as a reference. In particular, we aim to identify unseen compound expressions with the model trained on easily accessible basic expression datasets. To alleviate the problem of limited base classes in our FER task, we propose a novel Emotion Guided Similarity Network (EGS-Net), consisting of an emotion branch and a similarity branch, based on a two-stage learning framework. Specifically, in the first stage, the similarity branch is jointly trained with the emotion branch in a multi-task fashion. With the regularization of the emotion branch, we prevent the similarity branch from overfitting to sampled base classes that are highly overlapped across different episodes. In the second stage, the emotion branch and the similarity branch play a “two-student game” to alternately learn from each other, thereby further improving the inference ability of the similarity branch on unseen compound expressions. Experimental results on both in-the-lab and in-the-wild compound expression datasets demonstrate the superiority of our proposed method against several state-of-the-art methods.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

With the help of deep neural networks (DNNs), deep reinforcement learning (DRL) has achieved great success on many complex tasks, from games to robotic control. Compared to DNNs with partial brain-inspired structures and functions, spiking neural networks (SNNs) consider more biological features, including spiking neurons with complex dynamics and learning paradigms with biologically plausible plasticity principles. Inspired by the efficient computation of cell assembly in the biological brain, whereby memory-based coding is much more complex than readout, we propose a multiscale dynamic coding improved spiking actor network (MDC-SAN) for reinforcement learning to achieve effective decision-making. The population coding at the network scale is integrated with the dynamic neurons coding (containing 2nd-order neuronal dynamics) at the neuron scale towards a powerful spatial-temporal state representation. Extensive experimental results show that our MDC-SAN performs better than its counterpart deep actor network (based on DNNs) on four continuous control tasks from OpenAI gym. We think this is a significant attempt to improve SNNs from the perspective of efficient coding towards effective decision-making, just like that in biological networks.

poster session 2 @ red 2, poster session 9 @ red 2, oral session 2 @ red 2, poster session 2, poster session 9, oral session 2

Several methods for discovering interpretable directions in the latent space of pre-trained GANs have been proposed. Latent semantics discovered by unsupervised methods are less disentangled than supervised methods since they do not use pre-trained attribute classifiers. We propose Scale Ranking Estimator (SRE), which is trained using self-supervision. SRE enhances the disentanglement in directions obtained by existing unsupervised disentanglement techniques. These directions are updated to preserve the ordering of variation within each direction in latent space. Qualitative and quantitative evaluation of the discovered directions demonstrates that our proposed method significantly improves disentanglement in various datasets. We also show that the learned SRE can be used to perform Attribute-based image retrieval task without any training.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

Cognitive processing signals can be used to improve natural language processing (NLP) tasks. However, it is not clear how these signals correlate with linguistic information. Bridging between human language processing and linguistic features has been widely studied in neurolinguistics, usually via single-variable controlled experiments with highly-controlled stimuli. Such methods not only compromises the authenticity of natural reading, but also are time-consuming and expensive. In this paper, we propose a data-driven method to investigate the relationship between cognitive processing signals and linguistic features. Specifically, we present a unified attentional framework that is composed of embedding, attention, encoding and predicting layers to selectively map cognitive processing signals to linguistic features. We define the mapping procedure as a bridging task and develop 12 bridging tasks for lexical, syntactic and semantic features. The proposed framework only requires cognitive processing signals recorded under natural reading as inputs, and can be used to detect a wide range of linguistic features with a single cognitive dataset. Observations from experiment results resonate with previous neuroscience findings. In addition to this, our experiments also reveal a number of interesting findings, such as the correlation between contextual eye-tracking features and tense of sentence.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

Generalizing the medical image segmentation algorithms to unseen domains is an important research topic for computer- aided diagnosis and surgery. Most existing methods require a fully labeled dataset in each source domain. Although (Liu et al. 2021b) developed a semi-supervised domain general- ized method, it still requires the domain labels. This paper presents a novel confidence-aware cross pseudo supervision algorithm for semi-supervised domain generalized medical image segmentation. The main goal is to enhance the pseudo label quality for unlabeled images from unknown distribu- tions. To achieve it, we perform the Fourier transformation to learn low-level statistic information across domains and augment the images to incorporate cross-domain information. With these augmentations as perturbations, we feed the input to a confidence-aware cross pseudo supervision network to measure the variance of pseudo labels and regularize the network to learn with more confident pseudo labels. Our method sets new records on public datasets, i.e., M&Ms and SCGM. Notably, without using domain labels, our method surpasses the prior art that even uses domain labels by 11.67% on Dice on M&Ms dataset with 2% labeled data.Code is available at https://github.com/XMed-Lab/EPL SemiDG.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

The Shapley value is one of the most important normative division scheme in cooperative game theory, satisfying basic axioms. However, some allocation according to the Shapley value may seem unfair to humans. In this paper, we develop an automatic method that generates intuitive explanations for a Shapley-based payoff allocation, which utilizes the basic axioms. Given a coalitional game, our method decomposes it to sub-games, for which it is easy to generate verbal explanations, and shows that the given game is composed of the sub-games. Since the payoff allocation for each sub-game is perceived as fair, the Shapley-based payoff allocation for the given game should seem fair as well. We run an experiment with 210 human participants and show that when applying our method, humans perceive Shapley-based payoff allocation as significantly more fair than when using a general standard explanation.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

It is important to estimate the local average treatment effect (LATE) when compliance with a treatment assignment is incomplete. The previously proposed methods for LATE estimation required all relevant variables to be jointly observed in a single dataset; however, it is sometimes difficult or even impossible to collect such data in many real-world problems for technical or privacy reasons. We consider a novel problem setting in which LATE, as a function of covariates, is nonparametrically identified from the combination of separately observed datasets. For estimation, we show that the direct least squares method, which was originally developed for estimating the average treatment effect under complete compliance, is applicable to our setting. However, model selection and hyperparameter tuning for the direct least squares estimator can be unstable in practice since it is defined as a solution to the minimax problem. We then propose a weighted least squares estimator that enables simpler model selection by avoiding the minimax objective formulation. Unlike the inverse probability weighted (IPW) estimator, the proposed estimator directly uses the pre-estimated weight without inversion, avoiding the problems caused by the IPW methods. We demonstrate the effectiveness of our method through experiments using synthetic and real-world datasets.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

Building a socially intelligent agent involves many challenges, one of which is to teach the agent to speak guided by its value like a human. However, value-driven chatbots are still understudied in the area of dialogue systems. Most existing datasets focus on commonsense reasoning or social norm modeling. In this work, we present a new large-scale human value dataset called ValueNet, which contains human attitudes on 21,374 text scenarios. The dataset is organized in ten dimensions that conform to the basic human value theory in intercultural research. We further develop a Transformer-based value regression model on ValueNet to learn the utility distribution. Comprehensive empirical results show that the learned value model could benefit a wide range of dialogue tasks. For example, by teaching a generative agent with reinforcement learning and the rewards from the value model, our method attains state-of-the-art performance on the personalized dialog generation dataset: Persona-Chat. With values as additional features, existing emotion recognition models enable capturing rich human emotions in the context, which further improves the empathetic response generation performance in the EmpatheticDialogues dataset. To the best of our knowledge, ValueNet is the first large-scale text dataset for human value modeling, and we are the first one trying to incorporate a value model into emotionally intelligent dialogue systems. The dataset is available at https://liang-qiu.github.io/ValueNet/.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9

Learning a disentangled representation is still a challenge in the field of the interpretability of generative adversarial networks (GANs). This paper proposes a generic method to modify a traditional GAN into an interpretable GAN, which ensures that filters in an intermediate layer of the generator encode disentangled localized visual concepts. Each filter in the layer is supposed to consistently generate image regions corresponding to the same visual concept when generating different images. The interpretable GAN learns to automatically discover meaningful visual concepts without any annotations of visual concepts. The interpretable GAN enables people to modify a specific visual concept on generated images by manipulating feature maps of the corresponding filters in the layer. Our method can be broadly applied to different types of GANs. Experiments have demonstrated the effectiveness of our method. The code will be released when the paper is accepted.

poster session 2 @ red 2, poster session 9 @ red 2, oral session 9 @ red 2, poster session 2, poster session 9, oral session 9

Current representation learning methods for whole slide image (WSI) with pyramidal resolutions are inherently homogeneous and flat, which cannot fully exploit the multiscale and heterogeneous diagnostic information of different structures for comprehensive analysis. This paper presents a novel graph neural network-based multiple instance learning framework (i.e., H^2-MIL) to learn hierarchical representation from a heterogeneous graph with different resolutions for WSI analysis. A heterogeneous graph with the “resolution” attribute is constructed to explicitly model the feature and spatial-scaling relationship of multi-resolution patches. We then design a novel resolution-aware attention convolution (RAConv) block to learn compact yet discriminative representation from the graph, which tackles the heterogeneity of node neighbors with different resolutions and yields more reliable message passing. More importantly, to explore the task-related structured information of WSI pyramid, we elaborately design a novel iterative hierarchical pooling (IHPool) module to progressively aggregate the heterogeneous graph based on scaling relationships of different nodes. We evaluated our method on two public WSI datasets from the TCGA project, i.e., esophageal cancer and kidney cancer. Experimental results show that our method clearly outperforms the state-of-the-art methods on both tumor typing and staging tasks.

poster session 2 @ red 2, poster session 9 @ red 2, oral session 9 @ red 2, poster session 2, poster session 9, oral session 9

Multiple medical institutions collaboratively training a model using federated learning (FL) has become a promising solution for maximizing the potential of data-driven models, yet the non-independent and identically distributed (non-iid) data in medical images is still an outstanding challenge in real-world practice. The feature heterogeneity caused by diverse scanners or sensors introduces a drift in the learning process, in both local (client) and global (server) optimizations, which harms the convergence as well as model performance. Many previous works have attempted to address the non-iid issue by tackling the drift locally or globally, but how to jointly solve the two essentially coupled drifts is still unclear. In this work, we concentrate on handling both local and global drifts and introduce a new harmonizing framework called HarmoFL. First, we propose to mitigate the local update drift by normalizing amplitudes of images transformed into the frequency domain to mimic a unified scanner/sensor, in order to generate a harmonized feature space across local clients. Second, based on harmonized features, we design a client weight perturbation guiding each local model to reach a flat optimum, where a neighborhood area of the local optimal solution has a uniformly low loss. Without any extra communication cost, the perturbation assists the global model to optimize towards a converged optimal solution by aggregating several local flat optima. We have theoretically analyzed the proposed method and empirically conducted extensive experiments on three medical image classification and segmentation tasks, showing that HarmoFL outperforms a set of recent state-of-the-art methods with promising convergence behavior.

poster session 2 @ red 2, poster session 9 @ red 2, poster session 2, poster session 9