Konstantin Schürholt

Konstantin Schürholt, Boris Knyazev, Xavier Giró-i-Nieto et al. · berkeley

Learning representations of neural network weights given a model zoo is an emerging and challenging area with many potential applications from model inspection, to neural architecture search or knowledge distillation. Recently, an autoencoder trained on a model zoo was able to learn a hyper-representation, which captures intrinsic and extrinsic properties of the models in the zoo. In this work, we extend hyper-representations for generative use to sample new model weights. We propose layer-wise loss normalization which we demonstrate is key to generate high-performing models and several sampling methods based on the topology of hyper-representations. The models generated using our methods are diverse, performant and capable to outperform strong baselines as evaluated on several downstream tasks: initialization, ensemble sampling and transfer learning. Our results indicate the potential of knowledge aggregation from model zoos to new models via hyper-representations thereby paving the avenue for novel research directions.

Konstantin Schürholt, Diyar Taskiran, Boris Knyazev et al. · berkeley

In the last years, neural networks (NN) have evolved from laboratory environments to the state-of-the-art for many real-world problems. It was shown that NN models (i.e., their weights and biases) evolve on unique trajectories in weight space during training. Following, a population of such neural network models (referred to as model zoo) would form structures in weight space. We think that the geometry, curvature and smoothness of these structures contain information about the state of training and can reveal latent properties of individual models. With such model zoos, one could investigate novel approaches for (i) model analysis, (ii) discover unknown learning dynamics, (iii) learn rich representations of such populations, or (iv) exploit the model zoos for generative modelling of NN weights and biases. Unfortunately, the lack of standardized model zoos and available benchmarks significantly increases the friction for further research about populations of NNs. With this work, we publish a novel dataset of model zoos containing systematically generated and diverse populations of NN models for further research. In total the proposed model zoo dataset is based on eight image datasets, consists of 27 model zoos trained with varying hyperparameter combinations and includes 50'360 unique NN models as well as their sparsified twins, resulting in over 3'844'360 collected model states. Additionally, to the model zoo data we provide an in-depth analysis of the zoos and provide benchmarks for multiple downstream tasks. The dataset can be found at www.modelzoos.cc.

Konstantin Schürholt, Boris Knyazev, Xavier Giró-i-Nieto et al. · berkeley

Learning representations of neural network weights given a model zoo is an emerging and challenging area with many potential applications from model inspection, to neural architecture search or knowledge distillation. Recently, an autoencoder trained on a model zoo was able to learn a hyper-representation, which captures intrinsic and extrinsic properties of the models in the zoo. In this work, we extend hyper-representations for generative use to sample new model weights as pre-training. We propose layer-wise loss normalization which we demonstrate is key to generate high-performing models and a sampling method based on the empirical density of hyper-representations. The models generated using our methods are diverse, performant and capable to outperform conventional baselines for transfer learning. Our results indicate the potential of knowledge aggregation from model zoos to new models via hyper-representations thereby paving the avenue for novel research directions.

Christian Nauck, Michael Lindner, Konstantin Schürholt et al.

To mitigate climate change, the share of renewable energies in power production needs to be increased. Renewables introduce new challenges to power grids regarding the dynamic stability due to decentralization, reduced inertia, and volatility in production. Since dynamic stability simulations are intractable and exceedingly expensive for large grids, graph neural networks (GNNs) are a promising method to reduce the computational effort of analyzing the dynamic stability of power grids. As a testbed for GNN models, we generate new, large datasets of dynamic stability of synthetic power grids, and provide them as an open-source resource to the research community. We find that GNNs are surprisingly effective at predicting the highly non-linear targets from topological information only. For the first time, performance that is suitable for practical use cases is achieved. Furthermore, we demonstrate the ability of these models to accurately identify particular vulnerable nodes in power grids, so-called troublemakers. Last, we find that GNNs trained on small grids generate accurate predictions on a large synthetic model of the Texan power grid, which illustrates the potential for real-world applications.

Dominik Honegger, Konstantin Schürholt, Damian Borth · berkeley

With growing size of Neural Networks (NNs), model sparsification to reduce the computational cost and memory demand for model inference has become of vital interest for both research and production. While many sparsification methods have been proposed and successfully applied on individual models, to the best of our knowledge their behavior and robustness has not yet been studied on large populations of models. With this paper, we address that gap by applying two popular sparsification methods on populations of models (so called model zoos) to create sparsified versions of the original zoos. We investigate the performance of these two methods for each zoo, compare sparsification layer-wise, and analyse agreement between original and sparsified populations. We find both methods to be very robust with magnitude pruning able outperform variational dropout with the exception of high sparsification ratios above 80%. Further, we find sparsified models agree to a high degree with their original non-sparsified counterpart, and that the performance of original and sparsified model is highly correlated. Finally, all models of the model zoos and their sparsified model twins are publicly available: modelzoos.cc.

Christian Nauck, Michael Lindner, Konstantin Schürholt et al.

To mitigate climate change, the share of renewable needs to be increased. Renewable energies introduce new challenges to power grids due to decentralization, reduced inertia and volatility in production. The operation of sustainable power grids with a high penetration of renewable energies requires new methods to analyze the dynamic stability. We provide new datasets of dynamic stability of synthetic power grids and find that graph neural networks (GNNs) are surprisingly effective at predicting the highly non-linear target from topological information only. To illustrate the potential to scale to real-sized power grids, we demonstrate the successful prediction on a Texan power grid model.

Christian Nauck, Rohan Gorantla, Michael Lindner et al.

The geometry of a graph is encoded in dynamical processes on the graph. Many graph neural network (GNN) architectures are inspired by such dynamical systems, typically based on the graph Laplacian. Here, we introduce Dirac--Bianconi GNNs (DBGNNs), which are based on the topological Dirac equation recently proposed by Bianconi. Based on the graph Laplacian, we demonstrate that DBGNNs explore the geometry of the graph in a fundamentally different way than conventional message passing neural networks (MPNNs). While regular MPNNs propagate features diffusively, analogous to the heat equation, DBGNNs allow for coherent long-range propagation. Experimental results showcase the superior performance of DBGNNs over existing conventional MPNNs for long-range predictions of power grid stability and peptide properties. This study highlights the effectiveness of DBGNNs in capturing intricate graph dynamics, providing notable advancements in GNN architectures.

Damian Falk, Léo Meynent, Florence Pfammatter et al.

The availability of large, structured populations of neural networks - called 'model zoos' - has led to the development of a multitude of downstream tasks ranging from model analysis, to representation learning on model weights or generative modeling of neural network parameters. However, existing model zoos are limited in size and architecture and neglect the transformer, which is among the currently most successful neural network architectures. We address this gap by introducing the first model zoo of vision transformers (ViT). To better represent recent training approaches, we develop a new blueprint for model zoo generation that encompasses both pre-training and fine-tuning steps, and publish 250 unique models. They are carefully generated with a large span of generating factors, and their diversity is validated using a thorough choice of weight-space and behavioral metrics. To further motivate the utility of our proposed dataset, we suggest multiple possible applications grounded in both extensive exploratory experiments and a number of examples from the existing literature. By extending previous lines of similar work, our model zoo allows researchers to push their model population-based methods from the small model regime to state-of-the-art architectures. We make our model zoo available at github.com/ModelZoos/ViTModelZoo.

Damian Falk, Konstantin Schürholt, Damian Borth

Re-using trained neural network models is a common strategy to reduce training cost and transfer knowledge. Weight space learning - using the weights of trained models as data modality - is a promising new field to re-use populations of pre-trained models for future tasks. Approaches in this field have demonstrated high performance both on model analysis and weight generation tasks. However, until now their learning setup requires homogeneous model zoos where all models share the same exact architecture, limiting their capability to generalize beyond the population of models they saw during training. In this work, we remove this constraint and propose a modification to a common weight space learning method to accommodate training on heterogeneous populations of models. We further investigate the resulting impact of model diversity on generating unseen neural network model weights for zero-shot knowledge transfer. Our extensive experimental evaluation shows that including models with varying underlying image datasets has a high impact on performance and generalization, for both in- and out-of-distribution settings. Code is available on github.com/HSG-AIML/MultiZoo-SANE.

Yefan Zhou, Jianlong Chen, Qinxue Cao et al.

This paper considers "model diagnosis", which we formulate as a classification problem. Given a pre-trained neural network (NN), the goal is to predict the source of failure from a set of failure modes (such as a wrong hyperparameter, inadequate model size, and insufficient data) without knowing the training configuration of the pre-trained NN. The conventional diagnosis approach uses training and validation errors to determine whether the model is underfitting or overfitting. However, we show that rich information about NN performance is encoded in the optimization loss landscape, which provides more actionable insights than validation-based measurements. Therefore, we propose a diagnosis method called MD tree based on loss landscape metrics and experimentally demonstrate its advantage over classical validation-based approaches. We verify the effectiveness of MD tree in multiple practical scenarios: (1) use several models trained on one dataset to diagnose a model trained on another dataset, essentially a few-shot dataset transfer problem; (2) use small models (or models trained with small data) to diagnose big models (or models trained with big data), essentially a scale transfer problem. In a dataset transfer task, MD tree achieves an accuracy of 87.7%, outperforming validation-based approaches by 14.88%. Our code is available at https://github.com/YefanZhou/ModelDiagnosis.

Kai Wang, Dongwen Tang, Wangbo Zhao et al.

Parameter generation has long struggled to match the scale of today large vision and language models, curbing its broader utility. In this paper, we introduce Recurrent Diffusion for Large Scale Parameter Generation (RPG), a novel framework that generates full neural network parameters up to hundreds of millions on a single GPU. Our approach first partitions a networks parameters into non-overlapping tokens, each corresponding to a distinct portion of the model. A recurrent mechanism then learns the inter token relationships, producing prototypes which serve as conditions for a diffusion process that ultimately synthesizes the full parameters. Across a spectrum of architectures and tasks including ResNets, ConvNeXts and ViTs on ImageNet 1K and COCO, and even LoRA based LLMs RPG achieves performance on par with fully trained networks while avoiding excessive memory overhead. Notably, it generalizes beyond its training set to generate valid parameters for previously unseen tasks, highlighting its flexibility in dynamic and open ended scenarios. By overcoming the longstanding memory and scalability barriers, RPG serves as a critical advance in AI generating AI, potentially enabling efficient weight generation at scales previously deemed infeasible.

Léo Meynent, Ivan Melev, Konstantin Schürholt et al.

The weights of neural networks (NNs) have recently gained prominence as a new data modality in machine learning, with applications ranging from accuracy and hyperparameter prediction to representation learning or weight generation. One approach to leverage NN weights involves training autoencoders (AEs), using contrastive and reconstruction losses. This allows such models to be applied to a wide variety of downstream tasks, and they demonstrate strong predictive performance and low reconstruction error. However, despite the low reconstruction error, these AEs reconstruct NN models with deteriorated performance compared to the original ones, limiting their usability with regard to model weight generation. In this paper, we identify a limitation of weight-space AEs, specifically highlighting that a structural loss, that uses the Euclidean distance between original and reconstructed weights, fails to capture some features critical for reconstructing high-performing models. We analyze the addition of a behavioral loss for training AEs in weight space, where we compare the output of the reconstructed model with that of the original one, given some common input. We show a strong synergy between structural and behavioral signals, leading to increased performance in all downstream tasks evaluated, in particular NN weights reconstruction and generation.

Zhiyuan Liang, Dongwen Tang, Yuhao Zhou et al.

Modern Parameter-Efficient Fine-Tuning (PEFT) methods such as low-rank adaptation (LoRA) reduce the cost of customizing large language models (LLMs), yet still require a separate optimization run for every downstream dataset. We introduce \textbf{Drag-and-Drop LLMs (\textit{DnD})}, a prompt-conditioned parameter generator that eliminates per-task training by mapping a handful of unlabeled task prompts directly to LoRA weight updates. A lightweight text encoder distills each prompt batch into condition embeddings, which are then transformed by a cascaded hyper-convolutional decoder into the full set of LoRA matrices. Once trained in a diverse collection of prompt-checkpoint pairs, DnD produces task-specific parameters in seconds, yielding i) up to \textbf{12,000$\times$} lower overhead than full fine-tuning, ii) average gains up to \textbf{30\%} in performance over the strongest training LoRAs on unseen common-sense reasoning, math, coding, and multimodal benchmarks, and iii) robust cross-domain generalization despite never seeing the target data or labels. Our results demonstrate that prompt-conditioned parameter generation is a viable alternative to gradient-based adaptation for rapidly specializing LLMs. Our project is available at \href{https://jerryliang24.github.io/DnD}{https://jerryliang24.github.io/DnD}.

Konstantin Schürholt, Léo Meynent, Yefan Zhou et al.

Using the weights of trained Neural Network (NN) models as data modality has recently gained traction as a research field - dubbed Weight Space Learning (WSL). Multiple recent works propose WSL methods to analyze models, evaluate methods, or synthesize weights. Weight space learning methods require populations of trained models as datasets for development and evaluation. However, existing collections of models - called `model zoos' - are unstructured or follow a rudimentary definition of diversity. In parallel, work rooted in statistical physics has identified phases and phase transitions in NN models. Models are homogeneous within the same phase but qualitatively differ from one phase to another. We combine the idea of `model zoos' with phase information to create a controlled notion of diversity in populations. We introduce 12 large-scale zoos that systematically cover known phases and vary over model architecture, size, and datasets. These datasets cover different modalities, such as computer vision, natural language processing, and scientific ML. For every model, we compute loss landscape metrics and validate full coverage of the phases. With this dataset, we provide the community with a resource with a wide range of potential applications for WSL and beyond. Evidence suggests the loss landscape phase plays a role in applications such as model training, analysis, or sparsification. We demonstrate this in an exploratory study of the downstream methods like transfer learning or model weights averaging.

Damian Falk, Konstantin Schürholt, Konstantinos Tzevelekakis et al.

The weights of neural networks have emerged as a novel data modality, giving rise to the field of weight space learning. A central challenge in this area is that learning meaningful representations of weights typically requires large, carefully constructed collections of trained models, typically referred to as model zoos. These model zoos are often trained ad-hoc, requiring large computational resources, constraining the learned weight space representations in scale and flexibility. In this work, we drop this requirement by training a weight space learning backbone on arbitrary models downloaded from large, unstructured model repositories such as Hugging Face. Unlike curated model zoos, these repositories contain highly heterogeneous models: they vary in architecture and dataset, and are largely undocumented. To address the methodological challenges posed by such heterogeneity, we propose a new weight space backbone designed to handle unstructured model populations. We demonstrate that weight space representations trained on models from Hugging Face achieve strong performance, often outperforming backbones trained on laboratory-generated model zoos. Finally, we show that the diversity of the model weights in our training set allows our weight space model to generalize to unseen data modalities. By demonstrating that high-quality weight space representations can be learned in the wild, we show that curated model zoos are not indispensable, thereby overcoming a strong limitation currently faced by the weight space learning community.

Konstantin Schürholt, Michael W. Mahoney, Damian Borth

Learning representations of well-trained neural network models holds the promise to provide an understanding of the inner workings of those models. However, previous work has either faced limitations when processing larger networks or was task-specific to either discriminative or generative tasks. This paper introduces the SANE approach to weight-space learning. SANE overcomes previous limitations by learning task-agnostic representations of neural networks that are scalable to larger models of varying architectures and that show capabilities beyond a single task. Our method extends the idea of hyper-representations towards sequential processing of subsets of neural network weights, thus allowing one to embed larger neural networks as a set of tokens into the learned representation space. SANE reveals global model information from layer-wise embeddings, and it can sequentially generate unseen neural network models, which was unattainable with previous hyper-representation learning methods. Extensive empirical evaluation demonstrates that SANE matches or exceeds state-of-the-art performance on several weight representation learning benchmarks, particularly in initialization for new tasks and larger ResNet architectures.

Konstantin Schürholt, Dimche Kostadinov, Damian Borth

Self-Supervised Learning (SSL) has been shown to learn useful and information-preserving representations. Neural Networks (NNs) are widely applied, yet their weight space is still not fully understood. Therefore, we propose to use SSL to learn hyper-representations of the weights of populations of NNs. To that end, we introduce domain specific data augmentations and an adapted attention architecture. Our empirical evaluation demonstrates that self-supervised representation learning in this domain is able to recover diverse NN model characteristics. Further, we show that the proposed learned representations outperform prior work for predicting hyper-parameters, test accuracy, and generalization gap as well as transfer to out-of-distribution settings.

Christian Nauck, Michael Lindner, Konstantin Schürholt et al.

The prediction of dynamical stability of power grids becomes more important and challenging with increasing shares of renewable energy sources due to their decentralized structure, reduced inertia and volatility. We investigate the feasibility of applying graph neural networks (GNN) to predict dynamic stability of synchronisation in complex power grids using the single-node basin stability (SNBS) as a measure. To do so, we generate two synthetic datasets for grids with 20 and 100 nodes respectively and estimate SNBS using Monte-Carlo sampling. Those datasets are used to train and evaluate the performance of eight different GNN-models. All models use the full graph without simplifications as input and predict SNBS in a nodal-regression-setup. We show that SNBS can be predicted in general and the performance significantly changes using different GNN-models. Furthermore, we observe interesting transfer capabilities of our approach: GNN-models trained on smaller grids can directly be applied on larger grids without the need of retraining.

Konstantin Schürholt, Damian Borth

Safe use of Deep Neural Networks (DNNs) requires careful testing. However, deployed models are often trained further to improve in performance. As rigorous testing and evaluation is expensive, triggers are in need to determine the degree of change of a model. In this paper we investigate the weight space of DNN models for structure that can be exploited to that end. Our results show that DNN models evolve on unique, smooth trajectories in weight space which can be used to track DNN training progress. We hypothesize that curvature and smoothness of the trajectories as well as step length along it may contain information on the state of training as well as potential domain shifts. We show that the model trajectories can be separated and the order of checkpoints on the trajectories recovered, which may serve as a first step towards DNN model versioning.