Kamil Adamczewski

Yawei Li, Kamil Adamczewski, Wen Li et al. · eth-zurich

Channel (or 3D filter) pruning serves as an effective way to accelerate the inference of neural networks. There has been a flurry of algorithms that try to solve this practical problem, each being claimed effective in some ways. Yet, a benchmark to compare those algorithms directly is lacking, mainly due to the complexity of the algorithms and some custom settings such as the particular network configuration or training procedure. A fair benchmark is important for the further development of channel pruning. Meanwhile, recent investigations reveal that the channel configurations discovered by pruning algorithms are at least as important as the pre-trained weights. This gives channel pruning a new role, namely searching the optimal channel configuration. In this paper, we try to determine the channel configuration of the pruned models by random search. The proposed approach provides a new way to compare different methods, namely how well they behave compared with random pruning. We show that this simple strategy works quite well compared with other channel pruning methods. We also show that under this setting, there are surprisingly no clear winners among different channel importance evaluation methods, which then may tilt the research efforts into advanced channel configuration searching methods.

Kamil Adamczewski, Christos Sakaridis, Vaishakh Patil et al.

Lidar is a vital sensor for estimating the depth of a scene. Typical spinning lidars emit pulses arranged in several horizontal lines and the monetary cost of the sensor increases with the number of these lines. In this work, we present the new problem of optimizing the positioning of lidar lines to find the most effective configuration for the depth completion task. We propose a solution to reduce the number of lines while retaining the up-to-the-mark quality of depth completion. Our method consists of two components, (1) line selection based on the marginal contribution of a line computed via the Shapley value and (2) incorporating line position spread to take into account its need to arrive at image-wide depth completion. Spatially-aware Shapley values (SaS) succeed in selecting line subsets that yield a depth accuracy comparable to the full lidar input while using just half of the lines.

Yilin Yang, Kamil Adamczewski, Danica J. Sutherland et al.

Maximum mean discrepancy (MMD) is a particularly useful distance metric for differentially private data generation: when used with finite-dimensional features it allows us to summarize and privatize the data distribution once, which we can repeatedly use during generator training without further privacy loss. An important question in this framework is, then, what features are useful to distinguish between real and synthetic data distributions, and whether those enable us to generate quality synthetic data. This work considers the using the features of $\textit{neural tangent kernels (NTKs)}$, more precisely $\textit{empirical}$ NTKs (e-NTKs). We find that, perhaps surprisingly, the expressiveness of the untrained e-NTK features is comparable to that of the features taken from pre-trained perceptual features using public data. As a result, our method improves the privacy-accuracy trade-off compared to other state-of-the-art methods, without relying on any public data, as demonstrated on several tabular and image benchmark datasets.

Kamil Adamczewski, Mijung Park

A major challenge in applying differential privacy to training deep neural network models is scalability.The widely-used training algorithm, differentially private stochastic gradient descent (DP-SGD), struggles with training moderately-sized neural network models for a value of epsilon corresponding to a high level of privacy protection. In this paper, we explore the idea of dimensionality reduction inspired by neural network pruning to improve the scalability of DP-SGD. We study the interplay between neural network pruning and differential privacy, through the two modes of parameter updates. We call the first mode, parameter freezing, where we pre-prune the network and only update the remaining parameters using DP-SGD. We call the second mode, parameter selection, where we select which parameters to update at each step of training and update only those selected using DP-SGD. In these modes, we use public data for freezing or selecting parameters to avoid privacy loss incurring in these steps. Naturally, the closeness between the private and public data plays an important role in the success of this paradigm. Our experimental results demonstrate how decreasing the parameter space improves differentially private training. Moreover, by studying two popular forms of pruning which do not rely on gradients and do not incur an additional privacy loss, we show that random selection performs on par with magnitude-based selection when it comes to DP-SGD training.

Kamil Adamczewski, Yawei Li, Luc van Gool

Neural network pruning is a rich field with a variety of approaches. In this work, we propose to connect the existing pruning concepts such as leave-one-out pruning and oracle pruning and develop them into a more general Shapley value-based framework that targets the compression of convolutional neural networks. To allow for practical applications in utilizing the Shapley value, this work presents the Shapley value approximations, and performs the comparative analysis in terms of cost-benefit utility for the neural network compression. The proposed ranks are evaluated against a new benchmark, Oracle rank, constructed based on oracle sets. The broad experiments show that the proposed normative ranking and its approximations show practical results, obtaining state-of-the-art network compression.

Wojciech Kozłowski, Radosław Kuczbański, Kamil Adamczewski et al.

Deep stochastic processes have recently become a central paradigm for image enhancement, with many methods explicitly conditioning the stochastic trajectory on the degraded input. However, the relationship between these conditional processes and standard diffusion models remains unclear. In this work, we introduce a unified perspective on stochastic image enhancement by classifying recent methods into three families of continuous-time processes: unconditional diffusion models, Ornstein-Uhlenbeck (OU) processes, and diffusion bridges. We show that all of these approaches arise from a common stochastic differential equation (SDE) formulation. This framework makes explicit that seemingly disparate methods differ primarily in their drift and diffusion terms, terminal distributions, and boundary conditions, while schedulers and samplers constitute orthogonal design choices. Leveraging this unification, we conduct a controlled empirical study across multiple image enhancement tasks using identical architectures and training protocols. Our results reveal no consistently dominant method; instead, we identify and disentangle the specific design choices that most strongly influence performance. Finally, we release ItoVision, a modular PyTorch library that implements the unified framework and enables rapid prototyping and fair comparison of stochastic image enhancement methods.

Kamil Adamczewski, Yingchen He, Mijung Park

Scalability is a significant challenge when it comes to applying differential privacy to training deep neural networks. The commonly used DP-SGD algorithm struggles to maintain a high level of privacy protection while achieving high accuracy on even moderately sized models. To tackle this challenge, we take advantage of the fact that neural networks are overparameterized, which allows us to improve neural network training with differential privacy. Specifically, we introduce a new training paradigm that uses \textit{pre-pruning} and \textit{gradient-dropping} to reduce the parameter space and improve scalability. The process starts with pre-pruning the parameters of the original network to obtain a smaller model that is then trained with DP-SGD. During training, less important gradients are dropped, and only selected gradients are updated. Our training paradigm introduces a tension between the rates of pre-pruning and gradient-dropping, privacy loss, and classification accuracy. Too much pre-pruning and gradient-dropping reduces the model's capacity and worsens accuracy, while training a smaller model requires less privacy budget for achieving good accuracy. We evaluate the interplay between these factors and demonstrate the effectiveness of our training paradigm for both training from scratch and fine-tuning pre-trained networks on several benchmark image classification datasets. The tools can also be readily incorporated into existing training paradigms.

Piotr Kubaty, Bartosz Wójcik, Bartłomiej Krzepkowski et al.

Early exits enable the network's forward pass to terminate early by attaching trainable internal classifiers to the backbone network. Existing early-exit methods typically adopt either a joint training approach, where the backbone and exit heads are trained simultaneously, or a disjoint approach, where the heads are trained separately. However, the implications of this choice are often overlooked, with studies typically adopting one approach without adequate justification. This choice influences training dynamics and its impact remains largely unexplored. In this paper, we introduce a set of metrics to analyze early-exit training dynamics and guide the choice of training strategy. We demonstrate that conventionally used joint and disjoint regimes yield suboptimal performance. To address these limitations, we propose a mixed training strategy: the backbone is trained first, followed by the training of the entire multi-exit network. Through comprehensive evaluations of training strategies across various architectures, datasets, and early-exit methods, we present the strengths and weaknesses of the early exit training strategies. In particular, we show consistent improvements in performance and efficiency using the proposed mixed strategy.

Jakub Binkowski, Kamil Adamczewski, Tomasz Kajdanowicz

Large language models frequently exhibit hallucinations: fluent and confident outputs that are factually incorrect or unsupported by the input context. While recent hallucination detection methods have explored various features derived from attention maps, the underlying mechanisms they exploit remain poorly understood. In this work, we propose SinkProbe, a hallucination detection method grounded in the observation that hallucinations are deeply entangled with attention sinks - tokens that accumulate disproportionate attention mass during generation - indicating a transition from distributed, input-grounded attention to compressed, prior-dominated computation. Importantly, although sink scores are computed solely from attention maps, we find that the classifier preferentially relies on sinks whose associated value vectors have large norms. Moreover, we show that previous methods implicitly depend on attention sinks by establishing their mathematical relationship to sink scores. Our findings yield a novel hallucination detection method grounded in theory that produces state-of-the-art results across popular datasets and LLMs.

Mikołaj Janusz, Tomasz Wojnar, Yawei Li et al.

Pruning is a core technique for compressing neural networks to improve computational efficiency. This process is typically approached in two ways: one-shot pruning, which involves a single pass of training and pruning, and iterative pruning, where pruning is performed over multiple cycles for potentially finer network refinement. Although iterative pruning has historically seen broader adoption, this preference is often assumed rather than rigorously tested. Our study presents one of the first systematic and comprehensive comparisons of these methods, providing rigorous definitions, benchmarking both across structured and unstructured settings, and applying different pruning criteria and modalities. We find that each method has specific advantages: one-shot pruning proves more effective at lower pruning ratios, while iterative pruning performs better at higher ratios. Building on these findings, we advocate for patience-based pruning and introduce a hybrid approach that can outperform traditional methods in certain scenarios, providing valuable insights for practitioners selecting a pruning strategy tailored to their goals and constraints. Source code is available at https://github.com/janumiko/pruning-benchmark.

Jakub Krajewski, Jan Ludziejewski, Kamil Adamczewski et al.

Mixture of Experts (MoE) models have emerged as a primary solution for reducing the computational cost of Large Language Models. In this work, we analyze their scaling properties, incorporating an expanded range of variables. Specifically, we introduce a new hyperparameter, granularity, whose adjustment enables precise control over the size of the experts. Building on this, we establish scaling laws for fine-grained MoE, taking into account the number of training tokens, model size, and granularity. Leveraging these laws, we derive the optimal training configuration for a given computational budget. Our findings not only show that MoE models consistently outperform dense Transformers but also highlight that the efficiency gap between dense and MoE models widens as we scale up the model size and training budget. Furthermore, we demonstrate that the common practice of setting the size of experts in MoE to mirror the feed-forward layer is not optimal at almost any computational budget.

Jan Małaśnicki, Kamil Ciebiera, Mateusz Boruń et al.

Recent years have seen a growing interest and adoption of LLMs, with Mixture-of-Experts (MoE) emerging as a leading architecture in extremely large models. Currently, the largest open-source models reach over $1$T parameters. At such scales, hyperparameter tuning becomes prohibitively expensive. Precisely for this reason, the $μ$Transfer is becoming a key technique. It allows for seamless transfer of optimal hyperparameters across model scales, resulting in a huge reduction in tuning costs. However, existing work has primarily focused on dense LLMs, leaving MoE architectures unexplored. In this work, we derive a $μ$-Parameterization for MoE, providing theoretical guarantees for feature learning across model widths. Our experiments demonstrate that the optimal learning rate reliably transfers across model sizes, establishing a foundation for efficient hyperparameter tuning in large-scale MoE models.

Kamil Adamczewski, Frederik Harder, Mijung Park

We introduce a simple and intuitive framework that provides quantitative explanations of statistical models through the probabilistic assessment of input feature importance. The core idea comes from utilizing the Dirichlet distribution to define the importance of input features and learning it via approximate Bayesian inference. The learned importance has probabilistic interpretation and provides the relative significance of each input feature to a model's output, additionally assessing confidence about its importance quantification. As a consequence of using the Dirichlet distribution over the explanations, we can define a closed-form divergence to gauge the similarity between learned importance under different models. We use this divergence to study the feature importance explainability tradeoffs with essential notions in modern machine learning, such as privacy and fairness. Furthermore, BIF can work on two levels: global explanation (feature importance across all data instances) and local explanation (individual feature importance for each data instance). We show the effectiveness of our method on a variety of synthetic and real datasets, taking into account both tabular and image datasets. The code is available at https://github.com/kamadforge/featimp_dp.

Jan Ludziejewski, Maciej Pióro, Jakub Krajewski et al.

Mixture of Experts (MoE) architectures have significantly increased computational efficiency in both research and real-world applications of large-scale machine learning models. However, their scalability and efficiency under memory constraints remain relatively underexplored. In this work, we present joint scaling laws for dense and MoE models, incorporating key factors such as the number of active parameters, dataset size, and the number of experts. Our findings provide a principled framework for selecting the optimal MoE configuration under fixed memory and compute budgets. Surprisingly, we show that MoE models can be more memory-efficient than dense models, contradicting conventional wisdom. To derive and validate the theoretical predictions of our scaling laws, we conduct over 280 experiments with up to 2.7B active parameters and up to 5B total parameters. These results offer actionable insights for designing and deploying MoE models in practical large-scale training scenarios.

Adam Pardyl, Michał Wronka, Maciej Wołczyk et al.

Active Visual Exploration (AVE) is a task that involves dynamically selecting observations (glimpses), which is critical to facilitate comprehension and navigation within an environment. While modern AVE methods have demonstrated impressive performance, they are constrained to fixed-scale glimpses from rigid grids. In contrast, existing mobile platforms equipped with optical zoom capabilities can capture glimpses of arbitrary positions and scales. To address this gap between software and hardware capabilities, we introduce AdaGlimpse. It uses Soft Actor-Critic, a reinforcement learning algorithm tailored for exploration tasks, to select glimpses of arbitrary position and scale. This approach enables our model to rapidly establish a general awareness of the environment before zooming in for detailed analysis. Experimental results demonstrate that AdaGlimpse surpasses previous methods across various visual tasks while maintaining greater applicability in realistic AVE scenarios.

Jan Ludziejewski, Jan Małaśnicki, Maciej Pióro et al.

In this work, we introduce a novel approach for optimizing LLM training by adjusting learning rates across weights of different components in Transformer models. Traditional methods often apply a uniform learning rate across all network layers, potentially overlooking the unique dynamics of each part. Remarkably, our introduced relative learning rates, RLRS, method accelerates the training process by up to $23\%$, particularly in complex models such as Mixture of Experts (MoE). Hyperparameters of RLRS can be efficiently tuned on smaller models and then effectively reused on models up to $27\times$ larger. This simple and effective method results in a substantial reduction in training time and computational resources, offering a practical and scalable solution for optimizing large-scale neural networks.

Maciej Stefaniak, Michał Krutul, Jan Małaśnicki et al.

Large language models have steadily increased in size to achieve improved performance; however, this growth has also led to greater inference time and computational demands. Consequently, there is rising interest in model size reduction methods. To address this issue, we propose Projected Compression, a novel model compression technique, that reduces model weights by utilizing projection modules. Specifically, we first train additional trainable projections weights and preserve access to all the original model parameters. Subsequently, these projections are merged into a lower-dimensional product matrix, resulting in a reduced-size standard Transformer-based model. Unlike alternative approaches that require additional computational overhead, our method matches the base model's per-token computation step in FLOPs. Experimental results show that Projected Compression outperforms the comparable hard pruning and retraining approach on higher quality models. Moreover, the performance margin scales well with the number of tokens.

Margarita Vinaroz, Mohammad-Amin Charusaie, Frederik Harder et al.

Kernel mean embedding is a useful tool to represent and compare probability measures. Despite its usefulness, kernel mean embedding considers infinite-dimensional features, which are challenging to handle in the context of differentially private data generation. A recent work proposes to approximate the kernel mean embedding of data distribution using finite-dimensional random features, which yields analytically tractable sensitivity. However, the number of required random features is excessively high, often ten thousand to a hundred thousand, which worsens the privacy-accuracy trade-off. To improve the trade-off, we propose to replace random features with Hermite polynomial features. Unlike the random features, the Hermite polynomial features are ordered, where the features at the low orders contain more information on the distribution than those at the high orders. Hence, a relatively low order of Hermite polynomial features can more accurately approximate the mean embedding of the data distribution compared to a significantly higher number of random features. As demonstrated on several tabular and image datasets, Hermite polynomial features seem better suited for private data generation than random Fourier features.

Kamil Adamczewski, Mijung Park

We introduce Dirichlet pruning, a novel post-processing technique to transform a large neural network model into a compressed one. Dirichlet pruning is a form of structured pruning that assigns the Dirichlet distribution over each layer's channels in convolutional layers (or neurons in fully-connected layers) and estimates the parameters of the distribution over these units using variational inference. The learned distribution allows us to remove unimportant units, resulting in a compact architecture containing only crucial features for a task at hand. The number of newly introduced Dirichlet parameters is only linear in the number of channels, which allows for rapid training, requiring as little as one epoch to converge. We perform extensive experiments, in particular on larger architectures such as VGG and ResNet (45% and 58% compression rate, respectively) where our method achieves the state-of-the-art compression performance and provides interpretable features as a by-product.

Frederik Harder, Kamil Adamczewski, Mijung Park

We propose a differentially private data generation paradigm using random feature representations of kernel mean embeddings when comparing the distribution of true data with that of synthetic data. We exploit the random feature representations for two important benefits. First, we require a minimal privacy cost for training deep generative models. This is because unlike kernel-based distance metrics that require computing the kernel matrix on all pairs of true and synthetic data points, we can detach the data-dependent term from the term solely dependent on synthetic data. Hence, we need to perturb the data-dependent term only once and then use it repeatedly during the generator training. Second, we can obtain an analytic sensitivity of the kernel mean embedding as the random features are norm bounded by construction. This removes the necessity of hyper-parameter search for a clipping norm to handle the unknown sensitivity of a generator network. We provide several variants of our algorithm, differentially-private mean embeddings with random features (DP-MERF) to jointly generate labels and input features for datasets such as heterogeneous tabular data and image data. Our algorithm achieves drastically better privacy-utility trade-offs than existing methods when tested on several datasets.

Kamil Adamczewski, Mijung Park

Convolutional neural networks (CNNs) in recent years have made a dramatic impact in science, technology and industry, yet the theoretical mechanism of CNN architecture design remains surprisingly vague. The CNN neurons, including its distinctive element, convolutional filters, are known to be learnable features, yet their individual role in producing the output is rather unclear. The thesis of this work is that not all neurons are equally important and some of them contain more useful information to perform a given task . Consequently, we quantify the significance of each filter and rank its importance in describing input to produce the desired output. This work presents two different methods: (1) a game theoretical approach based on Shapley value which computes the marginal contribution of each filter; and (2) a probabilistic approach based on what-we-call, the Importance switch using variational inference. Strikingly, these two vastly different methods produce similar experimental results, confirming the general theory that some of the filters are inherently more important that the others. The learned ranks can be readily useable for network compression and interpretability.

Changyong Oh, Kamil Adamczewski, Mijung Park

We propose a new variational family for Bayesian neural networks. We decompose the variational posterior into two components, where the radial component captures the strength of each neuron in terms of its magnitude; while the directional component captures the statistical dependencies among the weight parameters. The dependencies learned via the directional density provide better modeling performance compared to the widely-used Gaussian mean-field-type variational family. In addition, the strength of input and output neurons learned via the radial density provides a structured way to compress neural networks. Indeed, experiments show that our variational family improves predictive performance and yields compressed networks simultaneously.

Kamil Adamczewski, Szymon Matejczyk, Tomasz P. Michalak

The Shapley value has been recently advocated as a method to choose the seed nodes for the process of information diffusion. Intuitively, since the Shapley value evaluates the average marginal contribution of a player to the coalitional game, it can be used in the network context to evaluate the marginal contribution of a node in the process of information diffusion given various groups of already 'infected' nodes. Although the above direction of research seems promising, the current liter- ature is missing a throughout assessment of its performance. The aim of this work is to provide such an assessment of the existing Shapley value-based approaches to information diffusion.