MBMagnus Bengtsson

Magnus Bengtsson

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Magnus Bengtsson has a total of 10 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Parameter-Efficient Reference-Based PCA of Polar-Transformed Images

    by Magnus Bengtsson, Jens Wittsten

    Image classification using the MNIST and CIFAR-100 datasets remains a standard benchmark for evaluating machine learning models. This work introduces a reference-based Principal Component Analysis (PCA) framework in which images are represented relative to one or more reference images. For complex datasets, multiple class-specific PCA coordinate systems are constructed and combined through discriminative learning, enabling each sample to be represented relative to several reference bases. For MNIST, transforming images into polar coordinates substantially improves class separability. A direct PCA-based comparison in the polar domain achieved 60% validation accuracy, compared with 30% using the original Cartesian representation. The resulting low-dimensional feature vectors enable a two-layer fully connected network with approximately 2.2k trainable parameters to achieve 94% validation accuracy. For CIFAR-100, a purely PCA-based classification strategy achieved only 1% validation accuracy. However, combining class-specific reference-based PCA representations with a compact convolutional neural network and an argmax-based fusion strategy across multiple PCA coordinate systems yielded 97.5% validation accuracy using approximately 250k trainable parameters. To explain the effectiveness of the proposed approach, we also investigate the spectral structure of polar-transformed images. Empirical analysis reveals exponential decay in both angular Fourier spectra and covariance eigenvalues. We show that this behaviour follows naturally from an anisotropic regularity model, yielding exponential spectral decay and logarithmic bounds on the effective rank of the associated covariance operator. These results demonstrate that reference-based PCA provides an effective dimensionality reduction strategy for compact image classification models.

  • Open Access

    Continuous Temporal Energy Representations for Event-BasedSignals

    by Magnus Bengtsson

    Spatio-temporal signals arising from event-driven biological processes, such as surface electromyography(sEMG), exhibit asynchronous activation patterns that are challenging to modelusing conventional discrete or purely real-valued temporal representations.In this work, we propose the Continuous Temporal Energy Network (CTEN), a continuoustemporal modeling framework for event-driven signals based on latent temporal energy representations.The proposed approach maps asynchronous temporal activity into a continuous latentspace, where temporal structure is represented through phase-modulated activations and energybasedtemporal projections over finite observation windows.By transforming sparse event activity into structured continuous temporal energy patterns,the model enables temporal feature extraction without explicit recurrence, discrete spikesimulation, or surrogate gradient approximations. The resulting representation remains fullydifferentiable and compatible with conventional gradient-based optimization.Experimental evaluation on a synthetic interaural phase difference (IPD) classification taskdemonstrates stable and competitive performance compared to both spike-based and conventionaltemporal neural architectures. Additional experiments on the Ninapro sEMG datasetfurther indicate that the proposed representation generalizes beyond synthetic benchmarks toreal-world multichannel biosignals using rawtemporal input signals with minimal preprocessing.Ablation experiments further suggest that the primary representational strength of theframework originates from the continuous temporal energy representation itself, highlighting theeffectiveness of differentiable continuous temporal embeddings for event-driven learning tasks.

  • Open Access

    Compressing Large Language Models with PCA Without Performance Loss

    by Magnus Bengtsson

    We demonstrate that Principal Component Analysis (PCA), when applied in a structured manner, either to polar-transformed images or segment-wise to token sequences, enables extreme compression of neural models without sacrificing performance. Across three case studies, we show that a one-layer classifier trained on PCA-compressed polar MNIST achieves over 98 percent accuracy using only 840 parameters. A two-layer transformer trained on 70-dimensional PCA-reduced MiniLM embeddings reaches 76.62 percent accuracy on the 20 Newsgroups dataset with just 81000 parameters. A decoder-only transformer generates coherent token sequences from 70-dimensional PCA embeddings while preserving over 97 percent cosine similarity with full MiniLM representations, using less than 17 percent of the parameter count of GPT-2. These results highlight PCA-based input compression as a general and effective strategy for aligning model capacity with information content, enabling lightweight architectures across multiple modalities.

  • Open Access

    Glass Waste as a Supplementary Cementitious Material in Climate Reduced Concrete – A Review

    by Nada Al-Hellali, Magnus Bengtsson, Agnes Nagy, Madumita Sadagopan

    A recycling pathway using flat glass waste as a supplementary cementitious material (SCM) leads to circularity in concrete production through industrial symbiosis. By incorporating glass waste, the environmental impact of cement can be reduced, contributing to the goal of climate-neutral concrete by 2045. This review study highlights the pozzolanic, chemical and physical properties of glass powder (GP) activated through a mechanical process. The relationship between particle size of GP and the strength activity index as well as the hydration phases and its potential to reduce alkali-silica reactions (ASR) are discussed.


  • Open Access

    Selective Comminution Applied to Mineral Processing of a Tantalum Ore: A Technical, Economic Analysis

    by Lorena Guldris Leon, Magnus Bengtsson

    There is an increasing demand to simulate and optimize the performance and profit of comminution circuits, especially in low-grade ore processing, as is the case with critical metals minerals. Recent research has shown that the optimization result is greatly influenced by quality aspects of the products, such as cost, profit, and capacity. This paper presents a novel approach to performing a multi-objective technical and economic analysis of tantalum ore processing to increase the production of critical metals minerals. The article starts with mineral composition analysis to highlight the potential of strategies for balancing the process layout for maximized production. The introduction of a combined technical and economic analysis presents the possibility of improving the profit by rearranging the mass flow given the rock’s mineral composition. Results show that selective comminution can improve process capacity by 23% and decrease production cost by 10% for the presented case.


  • Open Access

    Understanding Mineral Liberation during Crushing Using Grade-by-Size Analysis—A Case Study of the Penuota Sn-Ta Mineralization, Spain

    by Lorena Guldris Leon, K. Johan Hogmalm, Magnus Bengtsson

    Coarse comminution test-work and modeling are powerful tools in the design and optimization of mineral processing plants and provide information on energy consumption. Additional information on mineral liberation characteristics can be used for assessing the potential of pre-concentration stages or screens in the plant design. In ores of high-value metals (e.g., Ta, W), standard techniques—such as the mineralogical quantification of grain mounts by quantitative evaluation of minerals by scanning electron microscopy (QEMSCAN) or chemical analysis by X-ray fluorescence (XRF) can be challenging, due to the low relative abundance of such valuable minerals. The cost of QEMSCAN is also a limiting factor, especially considering the large number of samples required for the optimization of coarse comminution. In this study, we present an extended analytical protocol to a well-established mechanical test of interparticle breakage to improve the assessment of coarse mineral liberation characteristics. The liberation of ore minerals is a function of the rock texture and the difference in size and mechanical properties of the valuable minerals relative to gangue minerals and they may fraction in certain grain sizes if they behave differently during comminution. By analyzing the bulk-chemistry of the different grain size fractions produced after compressional testing, and by generating element by size diagrams, it is possible to understand the liberation characteristics of an ore. We show, based on a case study performed on a tantalum ore deposit, that element distribution can be used to study the influence of mechanical parameters on mineral liberation. This information can direct further mineralogical investigation and test work.


  • Model of banana screen for robust performance

    by Gauti Asbjörnsson, Magnus Bengtsson, Erik Hulthén, Magnus Evertsson