Denis Krompaß

Haokun Chen, Yao Zhang, Denis Krompass et al. · deepmind, oxford

Recently, foundation models have exhibited remarkable advancements in multi-modal learning. These models, equipped with millions (or billions) of parameters, typically require a substantial amount of data for finetuning. However, collecting and centralizing training data from diverse sectors becomes challenging due to distinct privacy regulations. Federated Learning (FL) emerges as a promising solution, enabling multiple clients to collaboratively train neural networks without centralizing their local data. To alleviate client computation burdens and communication overheads, previous works have adapted Parameter-efficient Finetuning (PEFT) methods for FL. Hereby, only a small fraction of the model parameters are optimized and communicated during federated communications. Nevertheless, most previous works have focused on a single modality and neglected one common phenomenon, i.e., the presence of data heterogeneity across the clients. Therefore, in this work, we propose a finetuning framework tailored to heterogeneous multi-modal FL, called Federated Dual-Aadapter Teacher (FedDAT). Specifically, our approach leverages a Dual-Adapter Teacher (DAT) to address data heterogeneity by regularizing the client local updates and applying Mutual Knowledge Distillation (MKD) for an efficient knowledge transfer. FedDAT is the first approach that enables an efficient distributed finetuning of foundation models for a variety of heterogeneous Vision-Language tasks. To demonstrate its effectiveness, we conduct extensive experiments on four multi-modality FL benchmarks with different types of data heterogeneity, where FedDAT substantially outperforms the existing centralized PEFT methods adapted for FL.

Yao Zhang, Haokun Chen, Ahmed Frikha et al. · deepmind, oxford

Visual Question Answering (VQA) is a multi-discipline research task. To produce the right answer, it requires an understanding of the visual content of images, the natural language questions, as well as commonsense reasoning over the information contained in the image and world knowledge. Recently, large-scale Vision-and-Language Pre-trained Models (VLPMs) have been the mainstream approach to VQA tasks due to their superior performance. The standard practice is to fine-tune large-scale VLPMs pre-trained on huge general-domain datasets using the domain-specific VQA datasets. However, in reality, the application domain can change over time, necessitating VLPMs to continually learn and adapt to new domains without forgetting previously acquired knowledge. Most existing continual learning (CL) research concentrates on unimodal tasks, whereas a more practical application scenario, i.e, CL on cross-domain VQA, has not been studied. Motivated by this, we introduce CL-CrossVQA, a rigorous Continual Learning benchmark for Cross-domain Visual Question Answering, through which we conduct extensive experiments on 4 VLPMs, 4 CL approaches, and 5 VQA datasets from different domains. In addition, by probing the forgetting phenomenon of the intermediate layers, we provide insights into how model architecture affects CL performance, why CL approaches can help mitigate forgetting in VLPMs to some extent, and how to design CL approaches suitable for VLPMs in this challenging continual learning environment. To facilitate future work on CL for cross-domain VQA, we will release our datasets and code.

Haokun Chen, Denis Krompass, Jindong Gu et al. · deepmind, oxford

Federated Learning (FL) is a distributed machine learning (ML) paradigm, in which multiple clients collaboratively train ML models without centralizing their local data. Similar to conventional ML pipelines, the client local optimization and server aggregation procedure in FL are sensitive to the hyperparameter (HP) selection. Despite extensive research on tuning HPs for centralized ML, these methods yield suboptimal results when employed in FL. This is mainly because their "training-after-tuning" framework is unsuitable for FL with limited client computation power. While some approaches have been proposed for HP-Tuning in FL, they are limited to the HPs for client local updates. In this work, we propose a novel HP-tuning algorithm, called Federated Population-based Hyperparameter Tuning (FedPop), to address this vital yet challenging problem. FedPop employs population-based evolutionary algorithms to optimize the HPs, which accommodates various HP types at both the client and server sides. Compared with prior tuning methods, FedPop employs an online "tuning-while-training" framework, offering computational efficiency and enabling the exploration of a broader HP search space. Our empirical validation on the common FL benchmarks and complex real-world FL datasets, including full-sized Non-IID ImageNet-1K, demonstrates the effectiveness of the proposed method, which substantially outperforms the concurrent state-of-the-art HP-tuning methods in FL.

Haokun Chen, Ahmed Frikha, Denis Krompass et al.

Federated Learning (FL) is a decentralized learning paradigm, in which multiple clients collaboratively train deep learning models without centralizing their local data, and hence preserve data privacy. Real-world applications usually involve a distribution shift across the datasets of the different clients, which hurts the generalization ability of the clients to unseen samples from their respective data distributions. In this work, we address the recently proposed feature shift problem where the clients have different feature distributions, while the label distribution is the same. We propose Federated Representation Augmentation (FRAug) to tackle this practical and challenging problem. Our approach generates synthetic client-specific samples in the embedding space to augment the usually small client datasets. For that, we train a shared generative model to fuse the clients knowledge learned from their different feature distributions. This generator synthesizes client-agnostic embeddings, which are then locally transformed into client-specific embeddings by Representation Transformation Networks (RTNets). By transferring knowledge across the clients, the generated embeddings act as a regularizer for the client models and reduce overfitting to the local original datasets, hence improving generalization. Our empirical evaluation on public benchmarks and a real-world medical dataset demonstrates the effectiveness of the proposed method, which substantially outperforms the current state-of-the-art FL methods for non-IID features, including PartialFed and FedBN.

Ihab Ahmed, Denis Krompaß, Cheng Feng et al.

We investigate input normalization methods for Time-Series Foundation Models (TSFMs). While normalization is well-studied in dataset-specific time-series models, it remains overlooked in TSFMs where generalization is critical. Time-series data, unlike text or images, exhibits significant scale variation across domains and channels, coupled with non-stationarity, can undermine TSFM performance regardless of architectural complexity. Through systematic evaluation across four architecturally diverse TSFMs, we empirically establish REVIN as the most efficient approach, reducing zero-shot MASE by 89\% relative to an un-normalized baseline and by 44\% versus other normalization methods, while matching the best in-domain accuracy (0.84 MASE) without any dataset-level preprocessing -- yielding the highest accuracy-efficiency trade-off. Yet its effect utilization depends on architectural design choices and optimization objective, particularly with respect to training loss scale sensitivity and model type (probabilistic, point-forecast, or LLM-based models).

Cheng Feng, Long Huang, Denis Krompass

We present General Time Transformer (GTT), an encoder-only style foundation model for zero-shot multivariate time series forecasting. GTT is pretrained on a large dataset of 200M high-quality time series samples spanning diverse domains. In our proposed framework, the task of multivariate time series forecasting is formulated as a channel-wise next curve shape prediction problem, where each time series sample is represented as a sequence of non-overlapping curve shapes with a unified numerical magnitude. GTT is trained to predict the next curve shape based on a window of past curve shapes in a channel-wise manner. Experimental results demonstrate that GTT exhibits superior zero-shot multivariate forecasting capabilities on unseen time series datasets, even surpassing state-of-the-art supervised baselines. Additionally, we investigate the impact of varying GTT model parameters and training dataset scales, observing that the scaling law also holds in the context of zero-shot multivariate time series forecasting.

Ahmed Frikha, Haokun Chen, Denis Krompaß et al.

In this work, we investigate the unexplored intersection of domain generalization (DG) and data-free learning. In particular, we address the question: How can knowledge contained in models trained on different source domains be merged into a single model that generalizes well to unseen target domains, in the absence of source and target domain data? Machine learning models that can cope with domain shift are essential for real-world scenarios with often changing data distributions. Prior DG methods typically rely on using source domain data, making them unsuitable for private decentralized data. We define the novel problem of Data-Free Domain Generalization (DFDG), a practical setting where models trained on the source domains separately are available instead of the original datasets, and investigate how to effectively solve the domain generalization problem in that case. We propose DEKAN, an approach that extracts and fuses domain-specific knowledge from the available teacher models into a student model robust to domain shift. Our empirical evaluation demonstrates the effectiveness of our method which achieves first state-of-the-art results in DFDG by significantly outperforming data-free knowledge distillation and ensemble baselines.

Ahmed Frikha, Denis Krompaß, Volker Tresp

Machine learning models that can generalize to unseen domains are essential when applied in real-world scenarios involving strong domain shifts. We address the challenging domain generalization (DG) problem, where a model trained on a set of source domains is expected to generalize well in unseen domains without any exposure to their data. The main challenge of DG is that the features learned from the source domains are not necessarily present in the unseen target domains, leading to performance deterioration. We assume that learning a richer set of features is crucial to improve the transfer to a wider set of unknown domains. For this reason, we propose COLUMBUS, a method that enforces new feature discovery via a targeted corruption of the most relevant input and multi-level representations of the data. We conduct an extensive empirical evaluation to demonstrate the effectiveness of the proposed approach which achieves new state-of-the-art results by outperforming 18 DG algorithms on multiple DG benchmark datasets in the DomainBed framework.

Ahmed Frikha, Denis Krompaß, Volker Tresp

Although continual learning and anomaly detection have separately been well-studied in previous works, their intersection remains rather unexplored. The present work addresses a learning scenario where a model has to incrementally learn a sequence of anomaly detection tasks, i.e. tasks from which only examples from the normal (majority) class are available for training. We define this novel learning problem of continual anomaly detection (CAD) and formulate it as a meta-learning problem. Moreover, we propose A Rapid Continual Anomaly Detector (ARCADe), an approach to train neural networks to be robust against the major challenges of this new learning problem, namely catastrophic forgetting and overfitting to the majority class. The results of our experiments on three datasets show that, in the CAD problem setting, ARCADe substantially outperforms baselines from the continual learning and anomaly detection literature. Finally, we provide deeper insights into the learning strategy yielded by the proposed meta-learning algorithm.

Ahmed Frikha, Denis Krompaß, Hans-Georg Köpken et al.

Although few-shot learning and one-class classification (OCC), i.e., learning a binary classifier with data from only one class, have been separately well studied, their intersection remains rather unexplored. Our work addresses the few-shot OCC problem and presents a method to modify the episodic data sampling strategy of the model-agnostic meta-learning (MAML) algorithm to learn a model initialization particularly suited for learning few-shot OCC tasks. This is done by explicitly optimizing for an initialization which only requires few gradient steps with one-class minibatches to yield a performance increase on class-balanced test data. We provide a theoretical analysis that explains why our approach works in the few-shot OCC scenario, while other meta-learning algorithms fail, including the unmodified MAML. Our experiments on eight datasets from the image and time-series domains show that our method leads to better results than classical OCC and few-shot classification approaches, and demonstrate the ability to learn unseen tasks from only few normal class samples. Moreover, we successfully train anomaly detectors for a real-world application on sensor readings recorded during industrial manufacturing of workpieces with a CNC milling machine, by using few normal examples. Finally, we empirically demonstrate that the proposed data sampling technique increases the performance of more recent meta-learning algorithms in few-shot OCC and yields state-of-the-art results in this problem setting.

Yinchong Yang, Denis Krompass, Volker Tresp

The Recurrent Neural Networks and their variants have shown promising performances in sequence modeling tasks such as Natural Language Processing. These models, however, turn out to be impractical and difficult to train when exposed to very high-dimensional inputs due to the large input-to-hidden weight matrix. This may have prevented RNNs' large-scale application in tasks that involve very high input dimensions such as video modeling; current approaches reduce the input dimensions using various feature extractors. To address this challenge, we propose a new, more general and efficient approach by factorizing the input-to-hidden weight matrix using Tensor-Train decomposition which is trained simultaneously with the weights themselves. We test our model on classification tasks using multiple real-world video datasets and achieve competitive performances with state-of-the-art models, even though our model architecture is orders of magnitude less complex. We believe that the proposed approach provides a novel and fundamental building block for modeling high-dimensional sequential data with RNN architectures and opens up many possibilities to transfer the expressive and advanced architectures from other domains such as NLP to modeling high-dimensional sequential data.

Cristóbal Esteban, Volker Tresp, Yinchong Yang et al.

Embedding learning, a.k.a. representation learning, has been shown to be able to model large-scale semantic knowledge graphs. A key concept is a mapping of the knowledge graph to a tensor representation whose entries are predicted by models using latent representations of generalized entities. Knowledge graphs are typically treated as static: A knowledge graph grows more links when more facts become available but the ground truth values associated with links is considered time invariant. In this paper we address the issue of knowledge graphs where triple states depend on time. We assume that changes in the knowledge graph always arrive in form of events, in the sense that the events are the gateway to the knowledge graph. We train an event prediction model which uses both knowledge graph background information and information on recent events. By predicting future events, we also predict likely changes in the knowledge graph and thus obtain a model for the evolution of the knowledge graph as well. Our experiments demonstrate that our approach performs well in a clinical application, a recommendation engine and a sensor network application.

Volker Tresp, Cristóbal Esteban, Yinchong Yang et al.

Embedding learning, a.k.a. representation learning, has been shown to be able to model large-scale semantic knowledge graphs. A key concept is a mapping of the knowledge graph to a tensor representation whose entries are predicted by models using latent representations of generalized entities. Latent variable models are well suited to deal with the high dimensionality and sparsity of typical knowledge graphs. In recent publications the embedding models were extended to also consider time evolutions, time patterns and subsymbolic representations. In this paper we map embedding models, which were developed purely as solutions to technical problems for modelling temporal knowledge graphs, to various cognitive memory functions, in particular to semantic and concept memory, episodic memory, sensory memory, short-term memory, and working memory. We discuss learning, query answering, the path from sensory input to semantic decoding, and the relationship between episodic memory and semantic memory. We introduce a number of hypotheses on human memory that can be derived from the developed mathematical models.

Denis Krompaß, Stephan Baier, Volker Tresp

Large knowledge graphs increasingly add value to various applications that require machines to recognize and understand queries and their semantics, as in search or question answering systems. Latent variable models have increasingly gained attention for the statistical modeling of knowledge graphs, showing promising results in tasks related to knowledge graph completion and cleaning. Besides storing facts about the world, schema-based knowledge graphs are backed by rich semantic descriptions of entities and relation-types that allow machines to understand the notion of things and their semantic relationships. In this work, we study how type-constraints can generally support the statistical modeling with latent variable models. More precisely, we integrated prior knowledge in form of type-constraints in various state of the art latent variable approaches. Our experimental results show that prior knowledge on relation-types significantly improves these models up to 77% in link-prediction tasks. The achieved improvements are especially prominent when a low model complexity is enforced, a crucial requirement when these models are applied to very large datasets. Unfortunately, type-constraints are neither always available nor always complete e.g., they can become fuzzy when entities lack proper typing. We show that in these cases, it can be beneficial to apply a local closed-world assumption that approximates the semantics of relation-types based on observations made in the data.

Volker Tresp, Sonja Zillner, Maria J. Costa et al.

In this paper we define Clinical Data Intelligence as the analysis of data generated in the clinical routine with the goal of improving patient care. We define a science of a Clinical Data Intelligence as a data analysis that permits the derivation of scientific, i.e., generalizable and reliable results. We argue that a science of a Clinical Data Intelligence is sensible in the context of a Big Data analysis, i.e., with data from many patients and with complete patient information. We discuss that Clinical Data Intelligence requires the joint efforts of knowledge engineering, information extraction (from textual and other unstructured data), and statistics and statistical machine learning. We describe some of our main results as conjectures and relate them to a recently funded research project involving two major German university hospitals.