Search

Denoising Diffusion Probabilistic Models (DDPMs) achieve state-of-the-art performance in synthesizing new images from random noise, but they lack meaningful latent space that encodes data into features. Recent DDPM-based editing techniques try to mitigate this issue by inverting images back to their approximated staring noise. In this work, we study the relation between the initial Gaussian noise, the samples generated from it, and their corresponding latent encodings obtained through the inversion procedure. First, we interpret their spatial distance relations to show the inaccuracy of the DDIM inversion technique by localizing latent representations manifold between the initial noise and generated samples. Then, we demonstrate the peculiar relation between initial Gaussian noise and its corresponding generations during diffusion training, showing that the high-level features of generated images stabilize rapidly, keeping the spatial distance relationship between noises and generations consistent throughout the training.

Recent personalization methods for diffusion models, such as Dreambooth, allow fine-tuning pre-trained models to generate new concepts. However, applying these techniques across multiple tasks in order to include, e.g., several new objects or styles, leads to mutual interference between their adapters. While recent studies attempt to mitigate this issue by combining trained adapters across tasks after fine-tuning, we adopt a more rigorous regime and investigate the personalization of large diffusion models under a continual learning scenario, where such interference leads to catastrophic forgetting of previous knowledge. To that end, we evaluate the na\"ive continual fine-tuning of customized models and compare this approach with three methods for consecutive adapters' training: sequentially merging new adapters, merging orthogonally initialized adapters, and updating only relevant parameters according to the task. In our experiments, we show that the proposed approaches mitigate forgetting when compared to the na\"ive approach.

Esports are a mostly sedentary activity. There is a growing need for investigation into how biomechanical and physical abilities can be optimized for esports through training. One such research avenue concerns the ability of esports players to perform balance tasks due to the prolonged sedentary states that are required to reach the top echelon of performance. Our aim for this work is to describe and compare physical abilities (balance, grip strength, and self-reported training habits) of top Polish StarCraft~2 tournament players. Esports players differed significantly from the reference group in their ability to balance on one leg. Additionally, in a grip strength test, the esports group fared worse than the reference group in all consecutive attempts. Despite self-reported physical activity in the esports group, player fitness requires further research. Training optimization could offset the issues arising from sedentary activity, and intensifying esports training so it could take less time overall., Comment: 9 pages, 3 figures, 3 tables

Most language model pre-training frameworks concatenate multiple documents into fixed-length sequences and use causal masking to compute the likelihood of each token given its context; this strategy is widely adopted due to its simplicity and efficiency. However, to this day, the influence of the pre-training sequence composition strategy on the generalisation properties of the model remains under-explored. In this work, we find that applying causal masking can lead to the inclusion of distracting information from previous documents during pre-training, which negatively impacts the performance of the models on language modelling and downstream tasks. In intra-document causal masking, the likelihood of each token is only conditioned on the previous tokens in the same document, eliminating potential distracting information from previous documents and significantly improving performance. Furthermore, we find that concatenating related documents can reduce some potential distractions during pre-training, and our proposed efficient retrieval-based sequence construction method, BM25Chunk, can improve in-context learning (+11.6\%), knowledge memorisation (+9.8\%), and context utilisation (+7.2\%) abilities of language models without sacrificing efficiency.

Recent advancements in long-context large language models have attracted significant attention, yet their practical applications often suffer from suboptimal context utilization. This study investigates structuring training data to enhance semantic interdependence, demonstrating that this approach effectively improves context utilization. To this end, we introduce the Structured Packing for Long Context (SPLiCe) method, which utilizes retrieval to collate mutually relevant documents into long and coherent training examples. We validate SPLiCe empirically across models of varying sizes -- 3B, 7B, and 13B -- achieving improved performance in long-context tasks, such as Qasper and HotpotQA. Remarkably, even brief fine-tuning with SPLiCe is sufficient to realize these benefits. Additionally, SPLiCe effectively mitigates the lost-in-middle phenomenon often observed in large models. Our comprehensive analysis of SPLiCe explores its design choices and reveals intriguing transfer effects; for instance, training on programming code enhances performance on natural language tasks., Comment: new experiments with a 13B model

Dispatching strategies for gas turbines (GTs) are changing in modern electricity grids. A growing incorporation of intermittent renewable energy requires GTs to operate more but shorter cycles and more frequently on partial loads. Deep reinforcement learning (DRL) has recently emerged as a tool that can cope with this development and dispatch GTs economically. The key advantages of DRL are a model-free optimization and the ability to handle uncertainties, such as those introduced by varying loads or renewable energy production. In this study, three popular DRL algorithms are implemented for an economic GT dispatch problem on a case study in Alberta, Canada. We highlight the benefits of DRL by incorporating an existing thermodynamic software provided by Siemens Energy into the environment model and by simulating uncertainty via varying electricity prices, loads, and ambient conditions. Among the tested algorithms and baseline methods, Deep Q-Networks (DQN) obtained the highest rewards while Proximal Policy Optimization (PPO) was the most sample efficient. We further propose and implement a method to assign GT operation and maintenance cost dynamically based on operating hours and cycles. Compared to existing methods, our approach better approximates the true cost of modern GT dispatch and hence leads to more realistic policies., Comment: This work has been accepted to IFAC for publication under a Creative Commons Licence CC-BY-NC-ND

Large language models have an exceptional capability to incorporate new information in a contextual manner. However, the full potential of such an approach is often restrained due to a limitation in the effective context length. One solution to this issue is to endow an attention layer with access to an external memory, which comprises of (key, value) pairs. Yet, as the number of documents increases, the proportion of relevant keys to irrelevant ones decreases, leading the model to focus more on the irrelevant keys. We identify a significant challenge, dubbed the distraction issue, where keys linked to different semantic values might overlap, making them hard to distinguish. To tackle this problem, we introduce the Focused Transformer (FoT), a technique that employs a training process inspired by contrastive learning. This novel approach enhances the structure of the (key, value) space, enabling an extension of the context length. Our method allows for fine-tuning pre-existing, large-scale models to lengthen their effective context. This is demonstrated by our fine-tuning of $3B$ and $7B$ OpenLLaMA checkpoints. The resulting models, which we name LongLLaMA, exhibit advancements in tasks requiring a long context. We further illustrate that our LongLLaMA models adeptly manage a $256 k$ context length for passkey retrieval.

Generally, crowd datasets can be collected or generated from real or synthetic sources. Real data is generated by using infrastructure-based sensors (such as static cameras or other sensors). The use of simulation tools can significantly reduce the time required to generate scenario-specific crowd datasets, facilitate data-driven research, and next build functional machine learning models. The main goal of this work was to develop an extension of crowd simulation (named CrowdSim2) and prove its usability in the application of people-tracking algorithms. The simulator is developed using the very popular Unity 3D engine with particular emphasis on the aspects of realism in the environment, weather conditions, traffic, and the movement and models of individual agents. Finally, three methods of tracking were used to validate generated dataset: IOU-Tracker, Deep-Sort, and Deep-TAMA.

Data scarcity has become one of the main obstacles to developing supervised models based on Artificial Intelligence in Computer Vision. Indeed, Deep Learning-based models systematically struggle when applied in new scenarios never seen during training and may not be adequately tested in non-ordinary yet crucial real-world situations. This paper presents and publicly releases CrowdSim2, a new synthetic collection of images suitable for people and vehicle detection gathered from a simulator based on the Unity graphical engine. It consists of thousands of images gathered from various synthetic scenarios resembling the real world, where we varied some factors of interest, such as the weather conditions and the number of objects in the scenes. The labels are automatically collected and consist of bounding boxes that precisely localize objects belonging to the two object classes, leaving out humans from the annotation pipeline. We exploited this new benchmark as a testing ground for some state-of-the-art detectors, showing that our simulated scenarios can be a valuable tool for measuring their performances in a controlled environment., Comment: Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, 2023

The exact complexity of solving parity games is a major open problem. Several authors have searched for efficient algorithms over specific classes of graphs. In particular, Obdr\v{z}\'{a}lek showed that for graphs of bounded tree-width or clique-width, the problem is in $\mathrm{P}$, which was later improved by Ganardi, who showed that it is even in $\mathrm{LOGCFL}$ (with an additional assumption for clique-width case). Here we extend this line of research by showing that for graphs of bounded tree-depth the problem of solving parity games is in logspace uniform $\text{AC}^0$. We achieve this by first considering a parameter that we obtain from a modification of clique-width, which we call shallow clique-width. We subsequently provide a suitable reduction., Comment: This is the full version of the paper that has been accepted at CSL 2023 and is going to be published in Leibniz International Proceedings in Informatics (LIPIcs)

We cross-correlate positions of galaxies measured in data from the first three years of the Dark Energy Survey with Compton-$y$-maps generated using data from the South Pole Telescope (SPT) and the {\it Planck} mission. We model this cross-correlation measurement together with the galaxy auto-correlation to constrain the distribution of gas in the Universe. We measure the hydrostatic mass bias or, equivalently, the mean halo bias-weighted electron pressure $\langle b_{h}P_{e}\rangle$, using large-scale information. We find $\langle b_{h}P_{e}\rangle$ to be $[0.16^{+0.03}_{-0.04},0.28^{+0.04}_{-0.05},0.45^{+0.06}_{-0.10},0.54^{+0.08}_{-0.07},0.61^{+0.08}_{-0.06},0.63^{+0.07}_{-0.08}]$ meV cm$^{-3}$ at redshifts $z \sim [0.30, 0.46, 0.62,0.77, 0.89, 0.97]$. These values are consistent with previous work where measurements exist in the redshift range. We also constrain the mean gas profile using small-scale information, enabled by the high-resolution of the SPT data. We compare our measurements to different parametrized profiles based on the cosmo-OWLS hydrodynamical simulations. We find that our data are consistent with the simulation that assumes an AGN heating temperature of $10^{8.5}$K but are incompatible with the model that assumes an AGN heating temperature of $10^{8.0}$K. These comparisons indicate that the data prefer a higher value of electron pressure than the simulations within $r_{500c}$ of the galaxies' halos., Comment: 20 pages, 9 figures. Submitted to MNRAS

We present cosmological constraints from the analysis of two-point correlation functions between galaxy positions and galaxy lensing measured in Dark Energy Survey (DES) Year 3 data and measurements of cosmic microwave background (CMB) lensing from the South Pole Telescope (SPT) and Planck. When jointly analyzing the DES-only two-point functions and the DES cross-correlations with SPT+Planck CMB lensing, we find $\Omega_{\rm m} = 0.344\pm 0.030$ and $S_8 \equiv \sigma_8 (\Omega_{\rm m}/0.3)^{0.5} = 0.773\pm 0.016$, assuming $\Lambda$CDM. When additionally combining with measurements of the CMB lensing autospectrum, we find $\Omega_{\rm m} = 0.306^{+0.018}_{-0.021}$ and $S_8 = 0.792\pm 0.012$. The high signal-to-noise of the CMB lensing cross-correlations enables several powerful consistency tests of these results, including comparisons with constraints derived from cross-correlations only, and comparisons designed to test the robustness of the galaxy lensing and clustering measurements from DES. Applying these tests to our measurements, we find no evidence of significant biases in the baseline cosmological constraints from the DES-only analyses or from the joint analyses with CMB lensing cross-correlations. However, the CMB lensing cross-correlations suggest possible problems with the correlation function measurements using alternative lens galaxy samples, in particular the redMaGiC galaxies and high-redshift MagLim galaxies, consistent with the findings of previous studies. We use the CMB lensing cross-correlations to identify directions for further investigating these problems., Comment: 20 pages, 15 figures

Cross-correlations of galaxy positions and galaxy shears with maps of gravitational lensing of the cosmic microwave background (CMB) are sensitive to the distribution of large-scale structure in the Universe. Such cross-correlations are also expected to be immune to some of the systematic effects that complicate correlation measurements internal to galaxy surveys. We present measurements and modeling of the cross-correlations between galaxy positions and galaxy lensing measured in the first three years of data from the Dark Energy Survey with CMB lensing maps derived from a combination of data from the 2500 deg$^2$ SPT-SZ survey conducted with the South Pole Telescope and full-sky data from the Planck satellite. The CMB lensing maps used in this analysis have been constructed in a way that minimizes biases from the thermal Sunyaev Zel'dovich effect, making them well suited for cross-correlation studies. The total signal-to-noise of the cross-correlation measurements is 23.9 (25.7) when using a choice of angular scales optimized for a linear (nonlinear) galaxy bias model. We use the cross-correlation measurements to obtain constraints on cosmological parameters. For our fiducial galaxy sample, which consist of four bins of magnitude-selected galaxies, we find constraints of $\Omega_{m} = 0.272^{+0.032}_{-0.052}$ and $S_{8} \equiv \sigma_8 \sqrt{\Omega_{m}/0.3}= 0.736^{+0.032}_{-0.028}$ ($\Omega_{m} = 0.245^{+0.026}_{-0.044}$ and $S_{8} = 0.734^{+0.035}_{-0.028}$) when assuming linear (nonlinear) galaxy bias in our modeling. Considering only the cross-correlation of galaxy shear with CMB lensing, we find $\Omega_{m} = 0.270^{+0.043}_{-0.061}$ and $S_{8} = 0.740^{+0.034}_{-0.029}$. Our constraints on $S_8$ are consistent with recent cosmic shear measurements, but lower than the values preferred by primary CMB measurements from Planck., Comment: 25 pages, 19 figures, submitted to PRD

Joint analyses of cross-correlations between measurements of galaxy positions, galaxy lensing, and lensing of the cosmic microwave background (CMB) offer powerful constraints on the large-scale structure of the Universe. In a forthcoming analysis, we will present cosmological constraints from the analysis of such cross-correlations measured using Year 3 data from the Dark Energy Survey (DES), and CMB data from the South Pole Telescope (SPT) and Planck. Here we present two key ingredients of this analysis: (1) an improved CMB lensing map in the SPT-SZ survey footprint, and (2) the analysis methodology that will be used to extract cosmological information from the cross-correlation measurements. Relative to previous lensing maps made from the same CMB observations, we have implemented techniques to remove contamination from the thermal Sunyaev Zel'dovich effect, enabling the extraction of cosmological information from smaller angular scales of the cross-correlation measurements than in previous analyses with DES Year 1 data. We describe our model for the cross-correlations between these maps and DES data, and validate our modeling choices to demonstrate the robustness of our analysis. We then forecast the expected cosmological constraints from the galaxy survey-CMB lensing auto and cross-correlations. We find that the galaxy-CMB lensing and galaxy shear-CMB lensing correlations will on their own provide a constraint on $S_8=\sigma_8 \sqrt{\Omega_{\rm m}/0.3}$ at the few percent level, providing a powerful consistency check for the DES-only constraints. We explore scenarios where external priors on shear calibration are removed, finding that the joint analysis of CMB lensing cross-correlations can provide constraints on the shear calibration amplitude at the 5 to 10% level., Comment: 30 pages, 20 figures, To be submitted to PRD

We provide the first combined cosmological analysis of South Pole Telescope (SPT) and Planck cluster catalogs. The aim is to provide an independent calibration for Planck scaling relations, exploiting the cosmological constraining power of the SPT-SZ cluster catalog and its dedicated weak lensing (WL) and X-ray follow-up observations. We build a new version of the Planck cluster likelihood. In the $\nu \Lambda$CDM scenario, focusing on the mass slope and mass bias of Planck scaling relations, we find $\alpha_{\text{SZ}} = 1.49_{-0.10}^{+0.07}$ and $(1-b)_{\text{SZ}} = 0.69_{-0.14}^{+0.07}$ respectively. The results for the mass slope show a $\sim 4 \, \sigma$ departure from the self-similar evolution, $\alpha_{\text{SZ}} \sim 1.8$. This shift is mainly driven by the matter density value preferred by SPT data, $\Omega_m = 0.30 \pm 0.03$, lower than the one obtained by Planck data alone, $\Omega_m = 0.37_{-0.06}^{+0.02}$. The mass bias constraints are consistent both with outcomes of hydrodynamical simulations and external WL calibrations, $(1-b) \sim 0.8$, and with results required by the Planck cosmic microwave background cosmology, $(1-b) \sim 0.6$. From this analysis, we obtain a new catalog of Planck cluster masses $M_{500}$. We estimate the ratio between the published Planck $M_{\text{SZ}}$ masses and our derived masses $M_{500}$, as a "measured mass bias", $(1-b)_M$. We analyse the mass, redshift and detection noise dependence of $(1-b)_M$, finding an increasing trend towards high redshift and low mass. These results mimic the effect of departure from self-similarity in cluster evolution, showing different dependencies for the low-mass high-mass, low-z high-z regimes., Comment: 22 pages, 10 figures, the catalogs can be found at http://pole.uchicago.edu/public/data/sptplanck_cluster

We show the improvement to cosmological constraints from galaxy cluster surveys with the addition of CMB-cluster lensing data. We explore the cosmological implications of adding mass information from the 3.1$\sigma$ detection of gravitational lensing of the cosmic microwave background (CMB) by galaxy clusters to the Sunyaev-Zel'dovich (SZ) selected galaxy cluster sample from the 2500 deg$^2$ SPT-SZ survey and targeted optical and X-ray followup data. In the $\Lambda$CDM model, the combination of the cluster sample with the Planck power spectrum measurements prefers $\sigma_8 \left(\Omega_m/0.3 \right)^{0.5} = 0.831 \pm 0.020$. Adding the cluster data reduces the uncertainty on this quantity by a factor of 1.4, which is unchanged whether or not the 3.1$\sigma$ CMB-cluster lensing measurement is included. We then forecast the impact of CMB-cluster lensing measurements with future cluster catalogs. Adding CMB-cluster lensing measurements to the SZ cluster catalog of the on-going SPT-3G survey is expected to improve the expected constraint on the dark energy equation of state $w$ by a factor of 1.3 to $\sigma(w) = 0.19$. We find the largest improvements from CMB-cluster lensing measurements to be for $\sigma_8$, where adding CMB-cluster lensing data to the cluster number counts reduces the expected uncertainty on $\sigma_8$ by factors of 2.4 and 3.6 for SPT-3G and CMB-S4 respectively., Comment: Comments are welcome

The signal resulting from magnetic resonance spectroscopy is occupied by noises and irregularities so in the further analysis preprocessing techniques have to be introduced. The main idea of the paper is to develop a model of a signal as a sum of harmonics and to find its parameters. Such an approach is based on singular value decomposition applied to the data arranged in the Hankel matrix (HSVD) and can be used in each step of preprocessing techniques. For that purpose a method has was tested on real phantom data.

We present component-separated maps of the primary cosmic microwave background/kinematic Sunyaev-Zel'dovich (SZ) amplitude and the thermal SZ Compton-$y$ parameter, created using data from the South Pole Telescope (SPT) and the Planck satellite. These maps, which cover the $\sim$2500 square degrees of the Southern sky imaged by the SPT-SZ survey, represent a significant improvement over previous such products available in this region by virtue of their higher angular resolution (1.25 arcminutes for our highest resolution Compton-$y$ maps) and lower noise at small angular scales. In this work we detail the construction of these maps using linear combination techniques, including our method for limiting the correlation of our lowest-noise Compton-$y$ map products with the cosmic infrared background. We perform a range of validation tests on these data products to test our sky modeling and combination algorithms, and we find good performance in all of these tests. Recognizing the potential utility of these data products for a wide range of astrophysical and cosmological analyses, including studies of the gas properties of galaxies, groups, and clusters, we make these products publicly available at http://pole.uchicago.edu/public/data/sptsz_ymap and on the NASA/LAMBDA website., Comment: Minor changes to match version accepted in ApJS

We present a catalog of emissive point sources detected in the SPT-SZ survey, a contiguous 2530-square-degree area surveyed with the South Pole Telescope (SPT) from 2008 - 2011 in three bands centered at 95, 150, and 220 GHz. The catalog contains 4845 sources measured at a significance of 4.5 sigma or greater in at least one band, corresponding to detections above approximately 9.8, 5.8, and 20.4 mJy in 95, 150, and 220 GHz, respectively. Spectral behavior in the SPT bands is used for source classification into two populations based on the underlying physical mechanisms of compact, emissive sources that are bright at millimeter wavelengths: synchrotron radiation from active galactic nuclei and thermal emission from dust. The latter population includes a component of high-redshift sources often referred to as submillimeter galaxies (SMGs). In the relatively bright flux ranges probed by the survey, these sources are expected to be magnified by strong gravitational lensing. The survey also contains sources consistent with protoclusters, groups of dusty galaxies at high redshift undergoing collapse. We cross-match the SPT-SZ catalog with external catalogs at radio, infrared, and X-ray wavelengths and identify available redshift information. The catalog splits into 3980 synchrotron-dominated and 865 dust-dominated sources and we determine a list of 506 SMGs. Ten sources in the catalog are identified as stars. We calculate number counts for the full catalog, and synchrotron and dusty components, using a bootstrap method and compare our measured counts with models. This paper represents the third and final catalog of point sources in the SPT-SZ survey., Comment: 43 pages, 19 figures, 7 tables. Submitted to AAS Journals 03/05/2020. 03/23/2020 co-author affiliation and acknowledgments updated

We report new measurements of millimeter-wave power spectra in the angular multipole range $2000 \le \ell \le 11,000$ (angular scales $5^\prime \gtrsim \theta \gtrsim 1^\prime$). By adding 95 and 150\,GHz data from the low-noise 500 deg$^2$ SPTpol survey to the SPT-SZ three-frequency 2540 deg$^2$ survey, we substantially reduce the uncertainties in these bands. These power spectra include contributions from the primary cosmic microwave background, cosmic infrared background, radio galaxies, and thermal and kinematic Sunyaev-Zel'dovich (SZ) effects. The data favor a thermal SZ (tSZ) power at 143\,GHz of $D^{\rm tSZ}_{3000} = 3.42 \pm 0.54~ \mu {\rm K}^2$ and a kinematic SZ (kSZ) power of $D^{\rm kSZ}_{3000} = 3.0 \pm 1.0~ \mu {\rm K}^2$. This is the first measurement of kSZ power at $\ge 3\,\sigma$. We study the implications of the measured kSZ power for the epoch of reionization, finding the duration of reionization to be $\Delta z_{re} = 1.0^{+1.6}_{-0.7}$ ($\Delta z_{re}< 4.1$ at 95% confidence), when combined with our previously published tSZ bispectrum measurement., Comment: Submitted to ApJ, 16 pages. (revised portions of the introduction and description of bandpower estimation)