Search

Large language models (LLMs) are expensive to deploy. Parameter sharing offers a possible path towards reducing their size and cost, but its effectiveness in modern LLMs remains fairly limited. In this work, we revisit "layer tying" as form of parameter sharing in Transformers, and introduce novel methods for converting existing LLMs into smaller "Recursive Transformers" that share parameters across layers, with minimal loss of performance. Here, our Recursive Transformers are efficiently initialized from standard pretrained Transformers, but only use a single block of unique layers that is then repeated multiple times in a loop. We further improve performance by introducing Relaxed Recursive Transformers that add flexibility to the layer tying constraint via depth-wise low-rank adaptation (LoRA) modules, yet still preserve the compactness of the overall model. We show that our recursive models (e.g., recursive Gemma 1B) outperform both similar-sized vanilla pretrained models (such as TinyLlama 1.1B and Pythia 1B) and knowledge distillation baselines -- and can even recover most of the performance of the original "full-size" model (e.g., Gemma 2B with no shared parameters). Finally, we propose Continuous Depth-wise Batching, a promising new inference paradigm enabled by the Recursive Transformer when paired with early exiting. In a theoretical analysis, we show that this has the potential to lead to significant (2-3x) gains in inference throughput., Comment: 48 pages, 17 figures, 17 tables

We describe and analyze a new class of direct energy conversion schemes based on the adiabatic magnetic drift of charged particles in axisymmetric magnetic fields. The efficiency of conversion as well as the geometrical and dynamical limitations of the recoverable power are calculated. The geometries of these axisymmetric field configurations are suited for direct energy conversion in radiating advanced aneutronic reactors and in advanced divertors of deuterium-tritium tokamak reactors. The E cross B configurations considered here do not suffer from the classical drawbacks and limitations of thermionic and magnetohydrodynamic high temperature direct energy conversion devices., Comment: 15 pages, 14 figures

Theoretical studies have suggested that nonlinear effects can lead to "radio frequency condensation", which coalesces RF power deposition and driven current near the center of a magnetic island. It is predicted that an initially broad current profile can coalesce in islands when they reach sufficient width, providing automatic stabilization. Experimental validation of the theory has thus far been lacking. This paper proposes experiments on DIII-D for testing and refining the theory of the nonlinear effects.

A promising approach for improving reasoning in large language models is to use process reward models (PRMs). PRMs provide feedback at each step of a multi-step reasoning trace, potentially improving credit assignment over outcome reward models (ORMs) that only provide feedback at the final step. However, collecting dense, per-step human labels is not scalable, and training PRMs from automatically-labeled data has thus far led to limited gains. To improve a base policy by running search against a PRM or using it as dense rewards for reinforcement learning (RL), we ask: "How should we design process rewards?". Our key insight is that, to be effective, the process reward for a step should measure progress: a change in the likelihood of producing a correct response in the future, before and after taking the step, corresponding to the notion of step-level advantages in RL. Crucially, this progress should be measured under a prover policy distinct from the base policy. We theoretically characterize the set of good provers and our results show that optimizing process rewards from such provers improves exploration during test-time search and online RL. In fact, our characterization shows that weak prover policies can substantially improve a stronger base policy, which we also observe empirically. We validate our claims by training process advantage verifiers (PAVs) to predict progress under such provers, and show that compared to ORMs, test-time search against PAVs is $>8\%$ more accurate, and $1.5-5\times$ more compute-efficient. Online RL with dense rewards from PAVs enables one of the first results with $5-6\times$ gain in sample efficiency, and $>6\%$ gain in accuracy, over ORMs.

Sheared flow increases the reactivity of fusion plasma. In unmagnetized DT plasma with flow gradients comparable to the mean free path of reacting ions, reactivity can be more than doubled. Neutron spectra are modified, helping to explain anomalous observations. The reactivity enhancement is particularly relevant in inertial confinement fusion (ICF), where it allows turbulent kinetic energy to contribute to the fusion burn even before thermalizing. In high-yield ICF experiments, the effect is most pronounced before bang time, suggesting a new mechanism for fast ignition., Comment: 6 pages, 4 figures

Voxel-based Morphometry (VBM) has emerged as a powerful approach in neuroimaging research, utilized in over 7,000 studies since the year 2000. Using Magnetic Resonance Imaging (MRI) data, VBM assesses variations in the local density of brain tissue and examines its associations with biological and psychometric variables. Here, we present deepmriprep, a neural network-based pipeline that performs all necessary preprocessing steps for VBM analysis of T1-weighted MR images using deep neural networks. Utilizing the Graphics Processing Unit (GPU), deepmriprep is 37 times faster than CAT12, the leading VBM preprocessing toolbox. The proposed method matches CAT12 in accuracy for tissue segmentation and image registration across more than 100 datasets and shows strong correlations in VBM results. Tissue segmentation maps from deepmriprep have over 95% agreement with ground truth maps, and its non-linear registration, using supervised SYMNet, predicts smooth deformation fields comparable to CAT12. The high processing speed of deepmriprep enables rapid preprocessing of extensive datasets and thereby fosters the application of VBM analysis to large-scale neuroimaging studies and opens the door to real-time applications. Finally, deepmripreps straightforward, modular design enables researchers to easily understand, reuse, and advance the underlying methods, fostering further advancements in neuroimaging research. deepmriprep can be conveniently installed as a Python package and is publicly accessible at https://github.com/wwu-mmll/deepmriprep.

Occupational outcomes like entrepreneurship are generally considered personal information that individuals should have the autonomy to disclose. With the advancing capability of artificial intelligence (AI) to infer private details from widely available human-centric data, such as social media, it is crucial to investigate whether AI can accurately extract private occupational information from such data. In this study, we demonstrate that deep neural networks can classify individuals as entrepreneurs based on a single facial image with high accuracy in data sourced from Crunchbase, a premier source for entrepreneurship data. Utilizing a dataset comprising facial images of 40,728 individuals, including both entrepreneurs and non-entrepreneurs, we trained a Convolutional Neural Network (CNN) and evaluated its classification performance. While human experts (n=650) and trained participants (n=133) were unable to classify entrepreneurs with accuracy above chance levels (>50%), the AI model achieved a classification accuracy of 79.51%. Several robustness tests show that this high level of accuracy is maintained under various conditions., Comment: 46 pages, 2 tables, 11 figures

It is possible to produce a ponderomotive effect in a plasma system without time-varying fields, if the plasma flows over spatially oscillations in the field. This can be achieved by superimposing a spatially oscillatory perturbation on a guide field, then setting up an electric field perpendicular to the guide field to drive flow over the perturbation. However, subtle distinctions in the structure of the resulting electric field can entirely change the behavior of the resulting ponderomotive force. Previous work has shown that, in slab models, these distinctions can be explained in terms of the polarization of the effective wave that appears in the co-moving frame. Here we consider what happens to this picture in a cylindrical system, where the transformation to the co-moving (rotating) frame is not inertial. It turns out that the non-inertial nature of this frame transformation can lead to counterintuitive behavior, partly due to the appearance of parallel (magnetic-field-aligned) electric fields in the rotating frame even in cases where none existed in the laboratory frame. Apart from the academic interest of this study, the practical impact lies in being better able to anticipate the antenna configuration on the plasma periphery of a cylindrical plasma that will lead to optimal ponderomotive barrier formation in the interior plasma., Comment: 9 pages, 3 figures

Plasma is capable of mediating the conversion of two pump photons into two different photons through a relativistic four-wave mixing nonlinearity. Spontaneously created photon pairs are emitted at symmetric angles with respect to the colinear pump direction, and the emission rate is largest if they have identical frequency. Thus, two orthogonally polarized pumps can produce polarization-entangled photon pairs through a mm-long homogeneous plasma. The noise from Raman scattering can be avoided if the pump detuning differs from twice the plasma frequency. On the other hand, pump detuning exactly equal to twice the plasma frequency can significantly enhance the interaction rate, which allows for the production of two-mode squeezed states. Remarkably, the amplified noise from Raman scattering are correlated and hence can be suppressed in one of the output quadratures, thereby maintaining the squeezing magnitude., Comment: 11 pages, 4 figures

Thermally bistable fluid tends to self-organize into clouds of hot and cold material, which are internally uniform and separated by thin conduction fronts. The evolution of these clouds has been studied for isobaric systems, but when pressure is instead treated as a dynamical quantity and allowed to evolve self-consistently, fundamentally different dynamics appear. Such a treatment is necessary in some laboratory plasmas, whose volume is constrained but whose pressure can vary. Solutions are derived for the evolution of clouds, accounting for pressure variation and interactions between conduction fronts. Additional stable configurations and secondary instabilities are derived, which may be relevant to fusion plasmas and to the study of photoionized plasma in the laboratory., Comment: 17 pages, 6 figures

Synchrotron radiation losses are a significant cause of concern for high-temperature aneutronic fusion reactions such as proton-Boron 11. The fact that radiation losses occur primarily in the high-energy tail, where the radiation itself has a substantial impact on the electron distribution, necessitates a self-consistent approach to modeling the diffusion and drag induced by synchrotron absorption and emission. Furthermore, an accurate model must account for the fact that the radiation emission spectrum is momentum-dependent, and the plasma opacity is frequency-dependent. Here, we present a simple Fokker-Planck operator, built on a newly-solved-for blackbody synchrotron diffusion operator, which captures all relevant features of the synchrotron radiation. Focusing on magnetic mirror fusion plasmas, we show that significant suppression of the electron distribution occurs for relativistic values of the perpendicular electron momentum, which therefore emit much less radiation than predicted under the assumption of a Maxwell-Juttner distribution., Comment: 14 pages, 6 figures

Spacecraft detumbling with magnetic torque coils is an inherently underactuated control problem. Contemporary and classical magnetorquer detumbling methods do not adequately consider this underactuation, and suffer from poor performance as a result. These controllers can get stuck on an uncontrollable manifold, resulting in long detumbling times and high power consumption. This work presents a novel detumble controller based on a non-monotonic Lyapunov function that predicts the future magnetic field along the satellite's orbit and avoids uncontrollable configurations. In comparison to other controllers in the literature, our controller detumbles a satellite in significantly less time while also converging to lower overall angular momentum. We provide a derivation and proof of convergence for our controller as well as Monte-Carlo simulation results demonstrating its performance in representative use cases.

Prediction-powered inference (PPI) is a method that improves statistical estimates based on limited human-labeled data. PPI achieves this by combining small amounts of human-labeled data with larger amounts of data labeled by a reasonably accurate -- but potentially biased -- automatic system, in a way that results in tighter confidence intervals for certain parameters of interest (e.g., the mean performance of a language model). In this paper, we propose a method called Stratified Prediction-Powered Inference (StratPPI), in which we show that the basic PPI estimates can be considerably improved by employing simple data stratification strategies. Without making any assumptions on the underlying automatic labeling system or data distribution, we derive an algorithm for computing provably valid confidence intervals for population parameters (such as averages) that is based on stratified sampling. In particular, we show both theoretically and empirically that, with appropriate choices of stratification and sample allocation, our approach can provide substantially tighter confidence intervals than unstratified approaches. Specifically, StratPPI is expected to improve in cases where the performance of the autorater varies across different conditional distributions of the target data.

Charged particles interacting with electromagnetic waves have a portion of their energy tied up in wave-driven oscillations. When these waves are localized to the exhaust of linear magnetic confinement systems this ponderomotive effect can be utilized to enhance particle confinement. The same effect can be derived for particles moving via an $\mathbf{E} \times \mathbf{B}$ drift into a region of a static perturbation to the electromagnetic fields which has a large wave vector component in the direction of the motion. In this work we use a simplified slab model to self-consistently solve for the electromagnetic fields within the fluid flowing plasma of a static flute-like ($k_\parallel = 0)$ perturbation, and evaluate the resulting ponderomotive potential. We find that two types of perturbations can exist within the flowing plasma, which are an O wave and an X wave in the frame moving with the fluid. In the case of tenuous plasma, these perturbations are magnetostatic or electrostatic multipole-analog perpendicular to the guiding magnetic field in the lab frame, respectfully. For denser plasmas, the O wave-like perturbation is screened at the electron skin depth scale, and the X wave-like perturbation is a combination of a similar perpendicular electric perturbation and parallel magnetic perturbation. The ponderomotive potential generated in the X wave-like case is gyrofrequency-dependent, and can be used as either potential barriers or potential wells, depending on the direction of the flow velocity., Comment: 14 pages, 10 figures

We introduce the Block Transformer which adopts hierarchical global-to-local modeling to autoregressive transformers to mitigate the inference bottlenecks associated with self-attention. Self-attention requires the key-value (KV) cache of all previous sequences to be retrieved from memory at every decoding step to retrieve context information, leading to two primary bottlenecks during batch inference. First, there is a significant delay in obtaining the first token, as the information of the entire prompt must first be processed to prefill the KV cache. Second, computation of subsequent tokens is bottlenecked by the high memory I/O demand of fetching the entire KV cache, which grows linearly with sequence length, incurring quadratic memory reads overall. We design the Block Transformer to strategically mitigate these costs, by incorporating coarsity and locality into an integrated global-to-local architecture. At the lower layers, we aggregate tokens into fixed size blocks to apply attention across the entire sequence at coarse-grained detail, to capture the global context while minimizing KV cache overhead. At upper layers, we apply attention within each block to decode individual tokens, to model fine-grained details with a lightweight local KV cache. We pretrain vanilla and Block Transformers from scratch and demonstrate that Block Transformers reach 10--20x inference throughput compared to vanilla transformers with equivalent perplexity and zero-shot task performance. Code is available at https://github.com/itsnamgyu/block-transformer., Comment: 37 pages, 24 figures, 7 tables

Language model (LM) post-training (or alignment) involves maximizing a reward function that is derived from preference annotations. Direct Preference Optimization (DPO) is a popular offline alignment method that trains a policy directly on preference data without the need to train a reward model or apply reinforcement learning. However, typical preference datasets have only a single, or at most a few, annotation per preference pair, which causes DPO to overconfidently assign rewards that trend towards infinite magnitude. This frequently leads to degenerate policies, sometimes causing even the probabilities of the preferred generations to go to zero. In this work, we analyze this phenomenon and propose distillation to get a better proxy for the true preference distribution over generation pairs: we train the LM to produce probabilities that match the distribution induced by a reward model trained on the preference data. Moreover, to account for uncertainty in the reward model we are distilling from, we optimize against a family of reward models that, as a whole, is likely to include at least one reasonable proxy for the preference distribution. Our results show that distilling from such a family of reward models leads to improved robustness to distribution shift in preference annotations, while preserving the simple supervised nature of DPO.

Broad classes of plasma phenomena can be understood in terms of phase-space rearrangements. For example, the net effect of a wave-particle interaction may consist of moving populations of particles from one region of phase space to another. Different phenomena drive rearrangements that obey different rules. When those rules can be specified, it is possible to calculate bounds that limit the possible effects the rearrangement could have (such as limits on how much energy can be extracted from the particles). This leads to two problems. The first is to understand the mapping between the allowed class of rearrangements and the possible outcomes that these rearrangements can have on the overall distribution. The second is to understand which rules are appropriate for which physical systems. There has been recent progress on both fronts, but a variety of interesting questions remain., Comment: 10 pages, 5 figures. Part of Special Collection: Papers from the 65th Annual Meeting of the APS Division of Plasma Physics

Fast rotation can improve the stability and confinement of fusion plasmas. However, to maintain a rapidly rotating fusion plasma in steady state, significant energy must be invested in spinning up each incoming fuel ion. We show here that, under the right circumstances, collisional cross-field radial fueling can directly transfer drift energy between outgoing and incoming ions without the need for external power recirculation, thereby reducing the energy costs of maintaining the rotation., Comment: 5 pages, 2 figures

Prediction-powered inference (PPI) is a method that improves statistical estimates based on limited human-labeled data. Specifically, PPI methods provide tighter confidence intervals by combining small amounts of human-labeled data with larger amounts of data labeled by a reasonably accurate, but potentially biased, automatic system. We propose a framework for PPI based on Bayesian inference that allows researchers to develop new task-appropriate PPI methods easily. Exploiting the ease with which we can design new metrics, we propose improved PPI methods for several importantcases, such as autoraters that give discrete responses (e.g., prompted LLM ``judges'') and autoraters with scores that have a non-linear relationship to human scores.

Zero-knowledge proofs (ZKPs) are computationally demanding to generate. Their importance for applications like ZK-Rollups has prompted some to outsource ZKP generation to a market of specialized provers. However, existing market designs either do not fit the ZKP setting or lack formal description and analysis. In this work, we propose a formal ZKP market model that captures the interactions between users submitting ZKP tasks and provers competing to generate proofs. Building on this model, we introduce $Proo\varphi$, an auction-based ZKP market mechanism. We prove that $Proo\varphi$ is incentive compatible for users and provers, and budget balanced. We augment $Proo\varphi$ with system-level designs to address the practical challenges of our setting, such as Sybil attacks, misreporting of prover capacity, and collusion. We analyze our system-level designs and show how they can mitigate the various security concerns.

We develop a principled procedure for determining when a large language model (LLM) should abstain from responding (e.g., by saying "I don't know") in a general domain, instead of resorting to possibly "hallucinating" a non-sensical or incorrect answer. Building on earlier approaches that use self-consistency as a more reliable measure of model confidence, we propose using the LLM itself to self-evaluate the similarity between each of its sampled responses for a given query. We then further leverage conformal prediction techniques to develop an abstention procedure that benefits from rigorous theoretical guarantees on the hallucination rate (error rate). Experimentally, our resulting conformal abstention method reliably bounds the hallucination rate on various closed-book, open-domain generative question answering datasets, while also maintaining a significantly less conservative abstention rate on a dataset with long responses (Temporal Sequences) compared to baselines using log-probability scores to quantify uncertainty, while achieveing comparable performance on a dataset with short answers (TriviaQA). To evaluate the experiments automatically, one needs to determine if two responses are equivalent given a question. Following standard practice, we use a thresholded similarity function to determine if two responses match, but also provide a method for calibrating the threshold based on conformal prediction, with theoretical guarantees on the accuracy of the match prediction, which might be of independent interest.

Although existing technology cannot yet directly produce fields at the Schwinger level, experimental facilities can already explore strong-field QED phenomena by taking advantage of the Lorentz boost of energetic electron beams. Recent studies show that QED cascades can create electron-positron pairs at sufficiently high density to exhibit collective plasma effects. Signatures of the collective pair plasma effects can appear in exquisite detail through plasma-induced frequency upshifts and chirps in the laser spectrum. Maximizing the magnitude of the QED plasma signature demands high pair density and low pair energy, which suits the configuration of colliding an over $10^{18}{Jm^{-3}}$ energy-density electron beam with a $10^{22}\mathrm{-}10^{23}{Wcm^{-2}}$ intensity laser pulse. The collision creates pairs that have a large plasma frequency, made even larger as they slow down or reverse direction due to both the radiation reaction and laser pressure. This paper explains at a tutorial level the key properties of the QED cascades and laser frequency upshift, and at the same time finds the minimum parameters that can be used to produce observable QED plasma., Comment: 10 pages, 3 figures

In the limit of sufficiently fast rotation, rotating mirror traps are known to be stable against the loss-cone modes associated with conventional (non-rotating) mirrors. This paper calculates how quickly a mirror configuration must rotate in order for several of these modes to be stabilized (in particular, the high-frequency convective loss cone, drift cyclotron loss cone, and Dory-Guest-Harris modes). Commonalities in the stabilization conditions for these modes then motivate a modified formulation of the Gardner free energy and diffusively accessible free energy to be used for systems in which the important modes have wave vectors that are orthogonal or nearly orthogonal to the magnetic field, as well as a modification to include the effects of a loss region in phase space., Comment: 26 pages, 6 figures

We perform experiments in the Large Plasma Device (LAPD) at the University of California, Los Angeles, studying how different end-electrode biasing schemes modify the radial potential profile in the machine. We impose biasing profiles of different polarities and gradient signs on a set of five concentric electrodes placed 12 m downstream from the plasma source. We find that imposing concave-down profiles (negative potential radial gradient) on the electrodes create radial potential profiles halfway up the plasma column that are comparable to those imposed on the electrodes, regardless of the biasing polarity. On the other hand, imposing concave-up profiles (positive potential radial gradient) leads to non-monotonic radial potential profiles. This observation can be explained by the current drawn through the electrodes and the parallel plasma resistivity, highlighting their important role in controlling the rotation of plasma. Concave-down plasma potential profiles, obtained by drawing electrons on the axis, are predicted to drive azimuthal drift velocities that can approach significant fractions of the ion sound speed in the central region of the plasma column., Comment: 24 pages, 12 figures, 3 tables

Reward models play a key role in aligning language model applications towards human preferences. However, this setup creates an incentive for the language model to exploit errors in the reward model to achieve high estimated reward, a phenomenon often termed \emph{reward hacking}. A natural mitigation is to train an ensemble of reward models, aggregating over model outputs to obtain a more robust reward estimate. We explore the application of reward ensembles to alignment at both training time (through reinforcement learning) and inference time (through reranking). First, we show that reward models are \emph{underspecified}: reward models that perform similarly in-distribution can yield very different rewards when used in alignment, due to distribution shift. Second, underspecification results in overoptimization, where alignment to one reward model does not improve reward as measured by another reward model trained on the same data. Third, overoptimization is mitigated by the use of reward ensembles, and ensembles that vary by their \emph{pretraining} seeds lead to better generalization than ensembles that differ only by their \emph{fine-tuning} seeds, with both outperforming individual reward models. However, even pretrain reward ensembles do not eliminate reward hacking: we show several qualitative reward hacking phenomena that are not mitigated by ensembling because all reward models in the ensemble exhibit similar error patterns., Comment: Published at the 2024 Conference on Language Modeling (CoLM)

Flow cytometry is widely used to identify cell populations in patient-derived fluids such as peripheral blood (PB) or cerebrospinal fluid (CSF). While ubiquitous in research and clinical practice, flow cytometry requires gating, i.e. cell type identification which requires labor-intensive and error-prone manual adjustments. To facilitate this process, we designed GateNet, the first neural network architecture enabling full end-to-end automated gating without the need to correct for batch effects. We train GateNet with over 8,000,000 events based on N=127 PB and CSF samples which were manually labeled independently by four experts. We show that for novel, unseen samples, GateNet achieves human-level performance (F1 score ranging from 0.910 to 0.997). In addition we apply GateNet to a publicly available dataset confirming generalization with an F1 score of 0.936. As our implementation utilizes graphics processing units (GPU), gating only needs 15 microseconds per event. Importantly, we also show that GateNet only requires ~10 samples to reach human-level performance, rendering it widely applicable in all domains of flow cytometry.

Self-focusing instability is a well-known phenomenon of nonlinear optics, which is of great importance in the field of laser-plasma interactions. Self-focusing instability leads to beam focusing and, consequently, breakup into multiple laser filaments. The majority of applications tend to avoid the laser filamentation regime due to its detrimental role on laser spot profile and peak intensity. In our work, using nonlinear Schr\"{o}dinger equation solver and particle-in-cell simulations, we address the problem of interaction of multiple parallel beams in plasmas. We consider both non-relativistic and moderately relativistic regimes, and demonstrate how the physics of parallel beam interaction transitions from the familiar self- and mutual-focusing instabilities in the non-relativistic regime to moderately relativistic regime, where an analytical description of filament interaction is impenetrable., Comment: 9 pages, 7 figures

Observing collective effects originating from the interplay between quantum electrodynamics and plasma physics might be achieved in upcoming experiments. In particular, the generation of electron-positron pairs and the observation of their collective dynamics could be simultaneously achieved in a collision between an intense laser and a highly relativistic electron beam through a laser frequency shift driven by an increase in the plasma density increase. In this collision, the radiation of high energy photons will serve a dual purpose: first, in seeding the cascade of pair generation; and, second, in decelerating the created pairs for detection. The deceleration results in a detectable shift in the plasma frequency. This deceleration was previously studied considering only a small sample of individual pair particles. However, the highly stochastic nature of the quantum radiation reaction in the strong field regime limits the descriptive power of the average behavior to the dynamics of pair particles. Here, we examine the full kinetic evolution of generated pairs in order to more accurately model the relativistically adjusted plasma density. As we show, the most effective pair energy for creating observable signatures occurs at a local minimum, obtained at finite laser field strength due to the tradeoff between pair deceleration and the relativistic particle oscillation at increasing laser intensity. For a small number of laser cycles, the quantum radiation reaction may re-arrange the generated pairs into anisotropic distributions in momentum space, although, in the one dimensional simulations considered here, this anisotropy quickly decreases.

Resonant upconversion through a sixth order relativistic nonlinearity resulting in a unique resonance was recently proposed [V. M. Malkin and N. J. Fisch, Physical Review E 108, 045208 (2023)]. The high order resonance is a unique non-integer multiple of a driving pump frequency resulting in a frequency upshift by a factor of $\approx 3.73$. We demonstrate the presence, unique requirements, and growth of this mode numerically. Through tuning waves to high amplitude, in a mildly underdense plasma, the six-photon process may grow more than other non-resonant, but lower order processes. The growth of the high frequency mode remains below the nonlinear growth regime. However, extending current numerical results to more strongly coupled resonances with longer pulse propagation distances suggests a pathway to significant upconversion.

We report the observation of longitudinal filamentation of an electron-positron pair plasma in a beam-driven QED cascade. The filaments are created in the "pair-reflection" regime, where the generated pairs are partially stopped and reflected in the strong laser field. The density filaments form near the center of the laser pulse and have diameters similar to the laser wavelength. They develop and saturate within a few laser cycles and do not induce sizable magnetostatic fields. We rule out the onset of two-stream instability or Weibel instability and attribute the origin of pair filamentation to laser ponderomotive forces. The small plasma filaments induce strong scattering of laser energy to large angles, serving as a signature of collective QED plasma dynamics., Comment: 9 pages, 9 figures

Adjustable hyperparameters of machine learning models typically impact various key trade-offs such as accuracy, fairness, robustness, or inference cost. Our goal in this paper is to find a configuration that adheres to user-specified limits on certain risks while being useful with respect to other conflicting metrics. We solve this by combining Bayesian Optimization (BO) with rigorous risk-controlling procedures, where our core idea is to steer BO towards an efficient testing strategy. Our BO method identifies a set of Pareto optimal configurations residing in a designated region of interest. The resulting candidates are statistically verified and the best-performing configuration is selected with guaranteed risk levels. We demonstrate the effectiveness of our approach on a range of tasks with multiple desiderata, including low error rates, equitable predictions, handling spurious correlations, managing rate and distortion in generative models, and reducing computational costs.

Major depressive disorder (MDD) is a complex psychiatric disorder that affects the lives of hundreds of millions of individuals around the globe. Even today, researchers debate if morphological alterations in the brain are linked to MDD, likely due to the heterogeneity of this disorder. The application of deep learning tools to neuroimaging data, capable of capturing complex non-linear patterns, has the potential to provide diagnostic and predictive biomarkers for MDD. However, previous attempts to demarcate MDD patients and healthy controls (HC) based on segmented cortical features via linear machine learning approaches have reported low accuracies. In this study, we used globally representative data from the ENIGMA-MDD working group containing an extensive sample of people with MDD (N=2,772) and HC (N=4,240), which allows a comprehensive analysis with generalizable results. Based on the hypothesis that integration of vertex-wise cortical features can improve classification performance, we evaluated the classification of a DenseNet and a Support Vector Machine (SVM), with the expectation that the former would outperform the latter. As we analyzed a multi-site sample, we additionally applied the ComBat harmonization tool to remove potential nuisance effects of site. We found that both classifiers exhibited close to chance performance (balanced accuracy DenseNet: 51%; SVM: 53%), when estimated on unseen sites. Slightly higher classification performance (balanced accuracy DenseNet: 58%; SVM: 55%) was found when the cross-validation folds contained subjects from all sites, indicating site effect. In conclusion, the integration of vertex-wise morphometric features and the use of the non-linear classifier did not lead to the differentiability between MDD and HC. Our results support the notion that MDD classification on this combination of features and classifiers is unfeasible.