VisTabNet: Adapting Vision Transformers for Tabular Data

Although deep learning models have had great success in natural language processing and computer vision, we do not observe comparable improvements in the case of tabular data, which is still the most common data type used in biological, industrial and financial applications. In particular, it is challenging to transfer large-scale pre-trained models to downstream tasks defined on small tabular datasets. To address this, we propose VisTabNet -- a cross-modal transfer learning method, which allows for adapting Vision Transformer (ViT) with pre-trained weights to process tabular data. By projecting tabular inputs to patch embeddings acceptable by ViT, we can directly apply a pre-trained Transformer Encoder to tabular inputs. This approach eliminates the conceptual cost of designing a suitable architecture for processing tabular data, while reducing the computational cost of training the model from scratch. Experimental results on multiple small tabular datasets (less than 1k samples) demonstrate VisTabNet's superiority, outperforming both traditional ensemble methods and recent deep learning models. The proposed method goes beyond conventional transfer learning practice and shows that pre-trained image models can be transferred to solve tabular problems, extending the boundaries of transfer learning.

StyleAutoEncoder for manipulating image attributes using pre-trained StyleGAN

Deep conditional generative models are excellent tools for creating high-quality images and editing their attributes. However, training modern generative models from scratch is very expensive and requires large computational resources. In this paper, we introduce StyleAutoEncoder (StyleAE), a lightweight AutoEncoder module, which works as a plugin for pre-trained generative models and allows for manipulating the requested attributes of images. The proposed method offers a cost-effective solution for training deep generative models with limited computational resources, making it a promising technique for a wide range of applications. We evaluate StyleAutoEncoder by combining it with StyleGAN, which is currently one of the top generative models. Our experiments demonstrate that StyleAutoEncoder is at least as effective in manipulating image attributes as the state-of-the-art algorithms based on invertible normalizing flows. However, it is simpler, faster, and gives more freedom in designing neural

FreSh: Frequency Shifting for Accelerated Neural Representation Learning

Implicit Neural Representations (INRs) have recently gained attention as a powerful approach for continuously representing signals such as images, videos, and 3D shapes using multilayer perceptrons (MLPs). However, MLPs are known to exhibit a low-frequency bias, limiting their ability to capture high-frequency details accurately. This limitation is typically addressed by incorporating high-frequency input embeddings or specialized activation layers. In this work, we demonstrate that these embeddings and activations are often configured with hyperparameters that perform well on average but are suboptimal for specific input signals under consideration, necessitating a costly grid search to identify optimal settings. Our key observation is that the initial frequency spectrum of an untrained model's output correlates strongly with the model's eventual performance on a given target signal. Leveraging this insight, we propose frequency shifting (or FreSh), a method that selects embedding hyperparameters to align the frequency spectrum of the model's initial output with that of the target signal. We show that this simple initialization technique improves performance across various neural representation methods and tasks, achieving results comparable to extensive hyperparameter sweeps but with only marginal computational overhead compared to training a single model with default hyperparameters.

Make Interval Bound Propagation great again

In various scenarios motivated by real life, such as medical data analysis, autonomous driving, and adversarial training, we are interested in robust deep networks. A network is robust when a relatively small perturbation of the input cannot lead to drastic changes in output (like change of class, etc.). This falls under the broader scope field of Neural Network Certification (NNC). Two crucial problems in NNC are of profound interest to the scientific community: how to calculate the robustness of a given pre-trained network and how to construct robust networks. The common approach to constructing robust networks is Interval Bound Propagation (IBP). This paper demonstrates that IBP is sub-optimal in the first case due to its susceptibility to the wrapping effect. Even for linear activation, IBP gives strongly sub-optimal bounds. Consequently, one should use strategies immune to the wrapping effect to obtain bounds close to optimal ones. We adapt two classical approaches dedicated to strict computations -- Dubleton Arithmetic and Affine Arithmetic -- to mitigate the wrapping effect in neural networks. These techniques yield precise results for networks with linear activation functions, thus resisting the wrapping effect. As a result, we achieve bounds significantly closer to the optimal level than IBPs.

InfoDisent: Explainability of Image Classification Models by Information Disentanglement

Understanding the decisions made by image classification networks is a critical area of research in deep learning. This task is traditionally divided into two distinct approaches: post-hoc methods and intrinsic methods. Post-hoc methods, such as GradCam, aim to interpret the decisions of pre-trained models by identifying regions of the image where the network focuses its attention. However, these methods provide only a high-level overview, making it difficult to fully understand the network's decision-making process. Conversely, intrinsic methods, like prototypical parts models, offer a more detailed understanding of network predictions but are constrained by specific architectures, training methods, and datasets. In this paper, we introduce InfoDisent, a hybrid model that combines the advantages of both approaches. By utilizing an information bottleneck, InfoDisent disentangles the information in the final layer of a pre-trained deep network, enabling the breakdown of classification decisions into basic, understandable atomic components. Unlike standard prototypical parts approaches, InfoDisent can interpret the decisions of pre-trained classification networks and be used for making classification decisions, similar to intrinsic models. We validate the effectiveness of InfoDisent on benchmark datasets such as ImageNet, CUB-200-2011, Stanford Cars, and Stanford Dogs for both convolutional and transformer backbones.

GeoGuide: Geometric guidance of diffusion models

Diffusion models are among the most effective methods for image generation. This is in particular because, unlike GANs, they can be easily conditioned during training to produce elements with desired class or properties. However, guiding a pre-trained diffusion model to generate elements from previously unlabeled data is significantly more challenging. One of the possible solutions was given by the ADM-G guiding approach. Although ADM-G successfully generates elements from the given class, there is a significant quality gap compared to a model originally conditioned on this class. In particular, the FID score obtained by the ADM-G-guided diffusion model is nearly three times lower than the class-conditioned guidance. We demonstrate that this issue is partly due to ADM-G providing minimal guidance during the final stage of the denoising process. To address this problem, we propose GeoGuide, a guidance model based on tracing the distance of the diffusion model's trajectory from the data manifold. The main idea of GeoGuide is to produce normalized adjustments during the backward denoising process. As shown in the experiments, GeoGuide surpasses the probabilistic approach ADM-G with respect to both the FID scores and the quality of the generated images.

RetroGFN: Diverse and Feasible Retrosynthesis using GFlowNets

Single-step retrosynthesis aims to predict a set of reactions that lead to the creation of a target molecule, which is a crucial task in molecular discovery. Although a target molecule can often be synthesized with multiple different reactions, it is not clear how to verify the feasibility of a reaction, because the available datasets cover only a tiny fraction of the possible solutions. Consequently, the existing models are not encouraged to explore the space of possible reactions sufficiently. In this paper, we propose a novel single-step retrosynthesis model, RetroGFN, that can explore outside the limited dataset and return a diverse set of feasible reactions by leveraging a feasibility proxy model during the training. We show that RetroGFN achieves competitive results on standard top-k accuracy while outperforming existing methods on round-trip accuracy. Moreover, we provide empirical arguments in favor of using round-trip accuracy which expands the notion of feasibility with respect to the standard top-k accuracy metric.

PrAViC: Probabilistic Adaptation Framework for Real-Time Video Classification

Video processing is generally divided into two main categories: processing of the entire video, which typically yields optimal classification outcomes, and real-time processing, where the objective is to make a decision as promptly as possible. The latter is often driven by the need to identify rapidly potential critical or dangerous situations. These could include machine failure, traffic accidents, heart problems, or dangerous behavior. Although the models dedicated to the processing of entire videos are typically well-defined and clearly presented in the literature, this is not the case for online processing, where a plethora of hand-devised methods exist. To address this, we present \our{}, a novel, unified, and theoretically-based adaptation framework for dealing with the online classification problem for video data. The initial phase of our study is to establish a robust mathematical foundation for the theory of classification of sequential data, with the potential to make a decision at an early stage. This allows us to construct a natural function that encourages the model to return an outcome much faster. The subsequent phase is to demonstrate a straightforward and readily implementable method for adapting offline models to online and recurrent operations. Finally, by comparing the proposed approach to the non-online state-of-the-art baseline, it is demonstrated that the use of \our{} encourages the network to make earlier classification decisions without compromising accuracy.

Sparser, Better, Deeper, Stronger: Improving Sparse Training with Exact Orthogonal Initialization

Static sparse training aims to train sparse models from scratch, achieving remarkable results in recent years. A key design choice is given by the sparse initialization, which determines the trainable sub-network through a binary mask. Existing methods mainly select such mask based on a predefined dense initialization. Such an approach may not efficiently leverage the mask's potential impact on the optimization. An alternative direction, inspired by research into dynamical isometry, is to introduce orthogonality in the sparse subnetwork, which helps in stabilizing the gradient signal. In this work, we propose Exact Orthogonal Initialization (EOI), a novel sparse orthogonal initialization scheme based on composing random Givens rotations. Contrary to other existing approaches, our method provides exact (not approximated) orthogonality and enables the creation of layers with arbitrary densities. We demonstrate the superior effectiveness and efficiency of EOI through experiments, consistently outperforming common sparse initialization techniques. Our method enables training highly sparse 1000-layer MLP and CNN networks without residual connections or normalization techniques, emphasizing the crucial role of weight initialization in static sparse training alongside sparse mask selection. The code is available at https://github.com/woocash2/sparser-better-deeper-stronger

LoRA-XS: Low-Rank Adaptation with Extremely Small Number of Parameters

The recent trend in scaling language models has led to a growing demand for parameter-efficient tuning (PEFT) methods such as LoRA (Low-Rank Adaptation). LoRA consistently matches or surpasses the full fine-tuning baseline with fewer parameters. However, handling numerous task-specific or user-specific LoRA modules on top of a base model still presents significant storage challenges. To address this, we introduce LoRA-XS (Low-Rank Adaptation with eXtremely Small number of parameters), a novel approach leveraging Singular Value Decomposition (SVD) for parameter-efficient fine-tuning. LoRA-XS introduces a small r x r weight matrix between frozen LoRA matrices, which are constructed by SVD of the original weight matrix. Training only r x r weight matrices ensures independence from model dimensions, enabling more parameter-efficient fine-tuning, especially for larger models. LoRA-XS achieves a remarkable reduction of trainable parameters by over 100x in 7B models compared to LoRA. Our benchmarking across various scales, including GLUE, GSM8k, and MATH benchmarks, shows that our approach outperforms LoRA and recent state-of-the-art approaches like VeRA in terms of parameter efficiency while maintaining competitive performance.