Search

Classical models of memory in psychology and neuroscience rely on similarity-based retrieval of stored patterns, where similarity is a function of retrieval cues and the stored patterns. While parsimonious, these models do not allow distinct representations for storage and retrieval, despite their distinct computational demands. Key-value memory systems, in contrast, distinguish representations used for storage (values) and those used for retrieval (keys). This allows key-value memory systems to optimize simultaneously for fidelity in storage and discriminability in retrieval. We review the computational foundations of key-value memory, its role in modern machine learning systems, related ideas from psychology and neuroscience, applications to a number of empirical puzzles, and possible biological implementations.

Magnetic reconnection is the key to explosive phenomena in the universe. The flux rope is crucial in three-dimensional magnetic reconnection theory and are commonly considered to be generated by secondary tearing mode instability. Here we show that the parallel electron flow moving toward the reconnection diffusion region can spontaneously form flux ropes. The electron flows form parallel current tubes in the separatrix region where the observational parameters suggest the tearing and Kelvin-Helmholtz instabilities are suppressed. The spontaneously formed flux ropes could indicate the importance of electron dynamics in a three-dimensional reconnection region., Comment: 21 pages, 7 figures

Electrons are accelerated to high, non-thermal energies during explosive energy-release events in space, such as magnetic reconnection. However, the properties and acceleration mechanisms of relativistic electrons directly associated with reconnection X-line are not well understood. This study utilizes Magnetospheric Multiscale (MMS) measurements to analyze the flux and spectral features of sub-relativistic to relativistic (~ 80 to 560 keV) electrons during a magnetic reconnection event in Earth's magnetotail. This event provided a unique opportunity to measure the electrons directly energized by X-line as MMS stayed in the separatrix layer, where the magnetic field directly connects to the X-line, for approximately half of the observation period. Our analysis revealed that the fluxes of relativistic electrons were clearly enhanced within the separatrix layer, and the highest flux was directed away from the X-line, which suggested that these electrons originated directly from the X-line. Spectral analysis showed that these relativistic electrons deviated from the main plasma sheet population and exhibited an "ankle" feature similar to that observed in galactic cosmic rays. The contribution of "ankle" electrons to the total electron energy density increased from 0.1% to 1% in the separatrix layer, though the spectral slopes did not exhibit clear variations. Further analysis indicated that while these relativistic electrons originated from the X-line, they experienced a non-negligible degree of scattering during transport. These findings provide clear evidence that magnetic reconnection in Earth's magnetotail can efficiently energize relativistic electrons directly at the X-line, providing new insights into the complex processes governing electron dynamics during magnetic reconnection., Comment: 23 pages, 5 figures

Does learning of task-relevant representations stop when behavior stops changing? Motivated by recent theoretical advances in machine learning and the intuitive observation that human experts continue to learn from practice even after mastery, we hypothesize that task-specific representation learning can continue, even when behavior plateaus. In a novel reanalysis of recently published neural data, we find evidence for such learning in posterior piriform cortex of mice following continued training on a task, long after behavior saturates at near-ceiling performance ("overtraining"). This learning is marked by an increase in decoding accuracy from piriform neural populations and improved performance on held-out generalization tests. We demonstrate that class representations in cortex continue to separate during overtraining, so that examples that were incorrectly classified at the beginning of overtraining can abruptly be correctly classified later on, despite no changes in behavior during that time. We hypothesize this hidden yet rich learning takes the form of approximate margin maximization; we validate this and other predictions in the neural data, as well as build and interpret a simple synthetic model that recapitulates these phenomena. We conclude by showing how this model of late-time feature learning implies an explanation for the empirical puzzle of overtraining reversal in animal learning, where task-specific representations are more robust to particular task changes because the learned features can be reused.

The magnetic cloud within an interplanetary coronal mass ejection (ICME) is characterized by high magnetic field intensities. In this study, we investigate the interaction of a magnetic cloud carrying a density structure with the Earth's bow shock during the ICME event on 24 April 2023. Elevated abundances of cold protons and heavier ions, namely alpha particles and singly charged helium ions, associated with the prominence plasma are observed within this structure. The plasma downstream of the bow shock exhibits an irregular compression pattern which could be due to the presence of heavy ions. Heavy ions carry a significant fraction of the upstream flow energy; however, due to their different charge per mass ratio and rigidity, they are less scattered by the electromagnetic and electrostatic waves at the shock. We find that downstream of the shock, while the thermal ion energy is only a small fraction of the background magnetic energy density, nevertheless increased ion fluxes reduce the characteristic wave speeds in the that region. As such, we observe a transition state of an unstable bow shock layer across which the plasma flow is super Alfv\'enic in both upstream and downstream regions. Our findings help with understanding the intense space weather impacts of such events.

The magnetic cloud (MC) of the Coronal Mass Ejection on April 24, 2023, contains sub-Alfv\'{e}nic solar wind, transforming Earth's magnetosphere from conventional bow-shock magnetotail configuration to Alfv\'{e}n wings. Utilizing measurements from the Magnetosphere Multiscale (MMS) mission, we present for the first time electron distribution signatures as the spacecraft traverses through various magnetic topologies during this transformation. Specifically, we characterize electrons inside the sub-Alfv\'{e}nic MC, on the dawn-dusk wing field lines and on the closed field lines. The signatures include strahl electrons in MC regions and energetic keV electrons streaming along the dawn and dusk wing field lines. We demonstrate the distribution signatures of dual wing reconnection, defined as reconnection between dawn-dusk Alfv\'{e}n wing field lines and the IMF. These signatures include four electron populations comprised of partially-depleted MC electrons and bi-directional energetic electrons with variations in energy and pitch-angle. The distributions reveal evidence of bursty magnetic reconnection under northward IMF., Comment: 11pages 4 figures

A suite of impressive scientific discoveries have been driven by recent advances in artificial intelligence. These almost all result from training flexible algorithms to solve difficult optimization problems specified in advance by teams of domain scientists and engineers with access to large amounts of data. Although extremely useful, this kind of problem solving only corresponds to one part of science - the "easy problem." The other part of scientific research is coming up with the problem itself - the "hard problem." Solving the hard problem is beyond the capacities of current algorithms for scientific discovery because it requires continual conceptual revision based on poorly defined constraints. We can make progress on understanding how humans solve the hard problem by studying the cognitive science of scientists, and then use the results to design new computational agents that automatically infer and update their scientific paradigms., Comment: 16 pages, 3 boxes, 5 figures

This short article highlights the unsolved problems of magnetic reconnection in collisionless plasma. The advanced in-situ plasma measurements and simulations enabled scientists to gain a novel understanding of magnetic reconnection. Still, outstanding questions remain on the complex dynamics and structures in the diffusion region, on the cross-scale and regional couplings, on the onset of magnetic reconnection, and on the details of energetics. Future directions of the magnetic reconnection research in terms of new observations, new simulations and interdisciplinary approaches are discussed., Comment: Submitted to Space Science Reviews. This is a review paper and is an outcome of the International Space Science Institute (ISSI) Workshop on Magnetic Reconnection: Explosive Energy Conversion in Space Plasmas held at June 27 - July 1, 2022

Dayside magnetic reconnection allows for the transfer of the solar wind's energy into Earth's magnetosphere. This process takes place in electron diffusion regions (EDRs) embedded in ion diffusion regions (IDRs), which form in the magnetopause boundary's current sheet. A significant out-of-plane parallel current contribution in the diffusion regions was reported in Beedle et al. 2023. In order to understand the underlying structure of this parallel current, we compared EDR statistical results with a 2.5D Particle-In Cell (PIC) simulation. From this comparison, we identified out-of-plane parallel current signatures as defining features of the outer EDR and IDR. This significant out-of-plane parallel current indicates the interaction of the IDR and EDR systems, and provides implications for not only understanding energy dissipation in the diffusion regions, but also determining the location of the outer EDR.

We statistically investigate high-frequency whistler waves (with frequencies higher than $\sim 10$ % of the local elect ron cyclotron frequency) at Earth's bow shock using Magnetospheric Multi-Scale (MMS) spacecraft observations. We focus specifically on the wave power within the shock transition layer, where we expect electron acceleration via stochastic sh ock drift acceleration (SSDA) to occur associated with efficient pitch-angle scattering by whistler waves. We find that the wave power is positively correlated with both the Alfv\'en Mach number in the normal incidence frame $M_{\rm A}$ and in the de Hoffmann-Teller frame $M_{\rm A}/\cos \theta_{Bn}$. The empirical relation with $M_{\rm A}/\cos \theta_{Bn}$ is compared with the theory of SSDA that predicts a threshold wave power proportional to $(M_{\rm A}/\cos \theta_{Bn})^{-2}$. The result suggests that the wave power exceeds the theoretical threshold for $M_{\rm A} / \cos \theta_{Bn} \gtrsim 30-60$, beyond which efficient electron acceleration is expected. This aligns very well with previous statistical analysis of electron acceleration at Earth's bow shock (M. Oka, G eophys.~Res.~Lett., 33, 5, 2006). Therefore, we consider that this study provides further support for SSDA as the mechanism of electron acceleration at Earth's bow shock. At higher-Mach-number astrophysical shocks, SSDA will be able to inject electrons into the diffusive shock acceleration process for subsequent acceleration to cosmic-ray energies., Comment: 15 pages, 12 figures. Accepted for publication in Physics of Plasmas

We present a comprehensive study of Magnetospheric Multiscale (MMS) spacecraft encounter with KHI during a geomagnetic storm, focusing on elucidating key turbulence properties and reconnection signatures observed at the edges of KH vortices. The spectral slope for electric field stays approximately constant for frequencies below the ion cyclotron frequency and exhibits a break around the lower hybrid frequency, indicating wave activity. Furthermore, MMS observes a current sheet accompanied by intense electron jets and features consistent with strong guide-field asymmetric reconnection across the magnetopause. Substantial agyrotropy (by a factor of 10) in electron distribution functions is observed in the reconnecting current sheet and at the edges of KH. Our observation presents a multi-scale view into KH turbulence under strongly driven conditions and into the dynamics occurring at electron dissipation scales., Comment: 10 pages, including 4 figures

On April 24th, 2023, a CME event caused the solar wind to become sub-Alfvenic, leading to the development of an Alfven Wing configuration in the Earth's Magnetosphere. Alfven Wings have previously been observed as cavities of low flow in Jupiter's magnetosphere, but the observing satellites did not have the ability to directly measure the Alfven Wings' current structures. Through in situ measurements made by the Magnetospheric Multiscale (MMS) spacecraft, the April 24th event provides us with the first direct measurements of current structures during an Alfven Wing configuration. We have found two distinct types of current structures associated with the Alfven Wing transformation as well as the magnetosphere recovery. These structures are observed to be significantly more anti-field-aligned and electron-driven than typical magnetopause currents, indicating the disruptions caused to the magnetosphere current system by the Alfven Wing formation.

We report a rare regime of Earth's magnetosphere interaction with sub-Alfv\'enic solar wind in which the windsock-like magnetosphere transforms into one with Alfv\'en wings. In the magnetic cloud of a Coronal Mass Ejection (CME) on April 24, 2023, NASA's Magnetospheric Multiscale mission distinguishes the following features: (1) unshocked and accelerated cold CME plasma coming directly against Earth's dayside magnetosphere; (2) dynamical wing filaments representing new channels of magnetic connection between the magnetosphere and foot points of the Sun's erupted flux rope; (3) cold CME ions observed with energized counter-streaming electrons, evidence of CME plasma captured due to reconnection between magnetic-cloud and Alfv\'en-wing field lines. The reported measurements advance our knowledge of CME interaction with planetary magnetospheres, and open new opportunities to understand how sub-Alfv\'enic plasma flows impact astrophysical bodies such as Mercury, moons of Jupiter, and exoplanets close to their host stars., Comment: 14 pages, including 4 figures, Under review in Geophys. Res. Lett

Adaptive behavior often requires predicting future events. The theory of reinforcement learning prescribes what kinds of predictive representations are useful and how to compute them. This paper integrates these theoretical ideas with work on cognition and neuroscience. We pay special attention to the successor representation (SR) and its generalizations, which have been widely applied both as engineering tools and models of brain function. This convergence suggests that particular kinds of predictive representations may function as versatile building blocks of intelligence., Comment: accepted to Neural Computation

It has been proposed that, when processing a stream of events, humans divide their experiences in terms of inferred latent causes (LCs) to support context-dependent learning. However, when shared structure is present across contexts, it is still unclear how the "splitting" of LCs and learning of shared structure can be simultaneously achieved. Here, we present the Latent Cause Network (LCNet), a neural network model of LC inference. Through learning, it naturally stores structure that is shared across tasks in the network weights. Additionally, it represents context-specific structure using a context module, controlled by a Bayesian nonparametric inference algorithm, which assigns a unique context vector for each inferred LC. Across three simulations, we found that LCNet could 1) extract shared structure across LCs in a function learning task while avoiding catastrophic interference, 2) capture human data on curriculum effects in schema learning, and 3) infer the underlying event structure when processing naturalistic videos of daily events. Overall, these results demonstrate a computationally feasible approach to reconciling shared structure and context-specific structure in a model of LCs that is scalable from laboratory experiment settings to naturalistic settings.

Large language models (LLMs) are being increasingly incorporated into scientific workflows. However, we have yet to fully grasp the implications of this integration. How should the advent of large language models affect the practice of science? For this opinion piece, we have invited four diverse groups of scientists to reflect on this query, sharing their perspectives and engaging in debate. Schulz et al. make the argument that working with LLMs is not fundamentally different from working with human collaborators, while Bender et al. argue that LLMs are often misused and over-hyped, and that their limitations warrant a focus on more specialized, easily interpretable tools. Marelli et al. emphasize the importance of transparent attribution and responsible use of LLMs. Finally, Botvinick and Gershman advocate that humans should retain responsibility for determining the scientific roadmap. To facilitate the discussion, the four perspectives are complemented with a response from each group. By putting these different perspectives in conversation, we aim to bring attention to important considerations within the academic community regarding the adoption of LLMs and their impact on both current and future scientific practices.

We analyze the local dynamics of magnetotail reconnection onset using Magnetospheric Multiscale (MMS) data. In conjunction with MMS, the macroscopic dynamics of this event were captured by a number of other ground and space-based observatories, as is reported in a companion paper. We find that the local dynamics of the onset were characterized by the rapid thinning of the cross-tail current sheet below the ion inertial scale, accompanied by the growth of flapping waves and the subsequent onset of electron tearing. Multiple kinetic-scale magnetic islands were detected coincident with the growth of an initially sub-Alfv\'enic, demagnetized tailward ion exhaust. The onset and rapid enhancement of parallel electron inflow at the exhaust boundary was a remote signature of the intensification of reconnection Earthward of the spacecraft. Two secondary reconnection sites are found embedded within the exhaust from a primary X-line. The primary X-line was designated as such on the basis that (1) while multiple jet reversals were observed in the current sheet, only one reversal of the electron inflow was observed at the high-latitude exhaust boundary, (2) the reconnection electric field was roughly 5 times larger at the primary X-line than the secondary X-lines, and (3) energetic electron fluxes increased and transitioned from anti-field-aligned to isotropic during the primary X-line crossing, indicating a change in magnetic topology. The results are consistent with the idea that a primary X-line mediates the reconnection of lobe magnetic field lines and accelerates electrons more efficiently than its secondary X-line counterparts., Comment: In press, JGR Space Physics, JGRA58162

We analyze a magnetotail reconnection onset event on 3 July 2017 that was observed under otherwise quiescent magnetospheric conditions by a fortuitous conjunction of six space and ground-based observatories. The study investigates the large-scale coupling of the solar wind - magnetosphere system that precipitated the onset of the magnetotail reconnection, focusing on the processes that thinned and stretched the cross-tail current layer in the absence of significant flux loading during a two-hour-long preconditioning phase. It is demonstrated with data in the (1) upstream solar wind, (2) at the low-latitude magnetopause, (3) in the high-latitude polar cap, and (4) in the magnetotail that the typical picture of solar wind-driven current sheet thinning via flux loading does not appear relevant for this particular event. We find that the current sheet thinning was, instead, initiated by a transient solar wind pressure pulse and that the current sheet thinning continued even as the magnetotail and solar wind pressures decreased. We suggest that field line curvature induced scattering (observed by Magnetospheric Multiscale (MMS)) and precipitation (observed by Defense Meteorological Satellite Program (DMSP)) of high-energy thermal protons may have evacuated plasma sheet thermal energy, which may require a thinning of the plasma sheet to preserve pressure equilibrium with the solar wind., Comment: In press, JGR space physics, JGRA58161

We propose that the grokking phenomenon, where the train loss of a neural network decreases much earlier than its test loss, can arise due to a neural network transitioning from lazy training dynamics to a rich, feature learning regime. To illustrate this mechanism, we study the simple setting of vanilla gradient descent on a polynomial regression problem with a two layer neural network which exhibits grokking without regularization in a way that cannot be explained by existing theories. We identify sufficient statistics for the test loss of such a network, and tracking these over training reveals that grokking arises in this setting when the network first attempts to fit a kernel regression solution with its initial features, followed by late-time feature learning where a generalizing solution is identified after train loss is already low. We find that the key determinants of grokking are the rate of feature learning -- which can be controlled precisely by parameters that scale the network output -- and the alignment of the initial features with the target function $y(x)$. We argue this delayed generalization arises when (1) the top eigenvectors of the initial neural tangent kernel and the task labels $y(x)$ are misaligned, but (2) the dataset size is large enough so that it is possible for the network to generalize eventually, but not so large that train loss perfectly tracks test loss at all epochs, and (3) the network begins training in the lazy regime so does not learn features immediately. We conclude with evidence that this transition from lazy (linear model) to rich training (feature learning) can control grokking in more general settings, like on MNIST, one-layer Transformers, and student-teacher networks., Comment: Adding new experiments on higher degree Hermite polynomials, multi-index targets, removed DMFT analysis from this version

At Jupiter, part of the auroral radio emissions are induced by the Galilean moons Io, Europa and Ganymede. Until now, except for Ganymede, they have been only remotely detected, using ground-based radio-telescopes or electric antennas aboard spacecraft. The polar trajectory of the Juno orbiter allows the spacecraft to cross the range of magnetic flux tubes which sustain the various Jupiter-satellite interactions, and in turn to sample in situ the associated radio emission regions. In this study, we focus on the detection and the characterization of radio sources associated with Io, Europa and Ganymede. Using electric wave measurements or radio observations (Juno/Waves), in situ electron measurements (Juno/JADE-E), and magnetic field measurements (Juno/MAG) we demonstrate that the Cyclotron Maser Instability (CMI) driven by a loss-cone electron distribution function is responsible for the encountered radio sources. We confirmed that radio emissions are associated with Main (MAW) or Reflected Alfv\'en Wing (RAW), but also show that for Europa and Ganymede, induced radio emissions are associated with Transhemispheric Electron Beam (TEB). For each traversed radio source, we determine the latitudinal extension, the CMI-resonant electron energy, and the bandwidth of the emission. We show that the presence of Alfv\'en perturbations and downward field aligned currents are necessary for the radio emissions to be amplified.

The solar wind is a continuous outflow of charged particles from the Sun's atmosphere into the solar system. At Earth, the solar wind's outward pressure is balanced by the Earth's magnetic field in a boundary layer known as the magnetopause. Plasma density and temperature differences across the boundary layer generate the Chapman-Ferraro current which supports the magnetopause. Along the dayside magnetopause, magnetic reconnection can occur in electron diffusion regions (EDRs) embedded into the larger ion diffusion regions (IDRs). These diffusion regions form when opposing magnetic field lines in the solar wind and Earth's magnetic field merge, releasing magnetic energy into the surrounding plasma. While previous studies have given us a general understanding of the structure of the diffusion regions, we still do not have a good grasp of how they are statistically differentiated from the non-diffusion region magnetopause. By investigating 251 magnetopause crossings from NASA's Magnetospheric Multiscale (MMS) Mission, we demonstrate that EDR magnetopause crossings show current densities an order of magnitude higher than regular magnetopause crossings - crossings that either passed through the reconnection exhausts or through the non-reconnecting magnetopause, providing a baseline for the magnetopause current sheet under a wide range of driving conditions. Significant current signatures parallel to the local magnetic field in EDR crossings are also identified, which is in contrast to the dominantly perpendicular current found in the regular magnetopause. Additionally, we show that the ion velocity along the magnetopause is highly correlated with a crossing's location, indicating the presence of magnetosheath flows inside the magnetopause.

Exploration is essential in reinforcement learning, particularly in environments where external rewards are sparse. Here we focus on exploration with intrinsic rewards, where the agent transiently augments the external rewards with self-generated intrinsic rewards. Although the study of intrinsic rewards has a long history, existing methods focus on composing the intrinsic reward based on measures of future prospects of states, ignoring the information contained in the retrospective structure of transition sequences. Here we argue that the agent can utilise retrospective information to generate explorative behaviour with structure-awareness, facilitating efficient exploration based on global instead of local information. We propose Successor-Predecessor Intrinsic Exploration (SPIE), an exploration algorithm based on a novel intrinsic reward combining prospective and retrospective information. We show that SPIE yields more efficient and ethologically plausible exploratory behaviour in environments with sparse rewards and bottleneck states than competing methods. We also implement SPIE in deep reinforcement learning agents, and show that the resulting agent achieves stronger empirical performance than existing methods on sparse-reward Atari games.

A novel multispacecraft technique applied to Magnetospheric Multiscale (MMS) mission data collected in the Earth's magnetosheath enables evaluation of the energy cascade rate solving the full Yaglom's equation in a turbulent space plasma. The method differs from existing approaches in that (i) it is inherently three-dimensional; (ii) it provides a statistically significant number of estimates from a single data stream; and (iii) it allows for a direct visualization of energy flux in turbulent plasmas. This new technique will ultimately provide a realistic, comprehensive picture of the turbulence process in plasmas.

In turbulence, nonlinear terms drive energy transfer from large-scale eddies into small scales through the so-called energy cascade. Turbulence often relaxes toward states that minimize energy; typically these states are considered globally. However, turbulence can also relax toward local quasi-equilibrium states, creating patches or cells where the magnitude of nonlinearity is reduced and energy cascade is impaired. We show, for the first time, compelling observational evidence that this ``cellularization'' of turbulence can occur due to local relaxation in a strongly turbulent natural environment such as the Earth's magnetosheath.

We present a statistical analysis of ion distributions in magnetic reconnection jets using data from the Magnetospheric Multiscale spacecraft. Compared with the quiet plasma in which the jet propagates, we often find anisotropic and non-Maxwellian ion distributions in the plasma jets. We observe magnetic field fluctuations associated with unstable ion distributions, but the wave amplitudes are not large enough to scatter ions during the observed travel time of the jet. We estimate that the phase-space diffusion due to chaotic and quasi-adiabatic ion motion in the current sheet is sufficiently fast to be the primary process leading to isotropization.

We motivate and introduce CHARD: Clinical Health-Aware Reasoning across Dimensions, to investigate the capability of text generation models to act as implicit clinical knowledge bases and generate free-flow textual explanations about various health-related conditions across several dimensions. We collect and present an associated dataset, CHARDat, consisting of explanations about 52 health conditions across three clinical dimensions. We conduct extensive experiments using BART and T5 along with data augmentation, and perform automatic, human, and qualitative analyses. We show that while our models can perform decently, CHARD is very challenging with strong potential for further exploration., Comment: EACL 2023. Code available at https://github.com/styfeng/CHARD

We present the first estimation of the energy cascade rate in Jupiter's magnetosheath (MS). We use in-situ observations from the Jovian Auroral Distributions Experiment (JADE) and the magnetometer investigation (MAG) instruments onboard the Juno spacecraft, in concert with two recent compressible models to investigate the cascade rate in the magnetohydrodynamic (MHD) scales. While a high level of compressible density fluctuations is observed in the Jovian MS, a constant energy flux exists in the MHD inertial range. The compressible isothermal and polytropic energy cascade rates increase in the MHD range when density fluctuations are present. We find that the energy cascade rate in Jupiter's magnetosheath is at least two orders of magnitude (100 times) smaller than the corresponding typical value in the Earth's magnetosheath., Comment: To be published in The Astrophysical Journal

The most widely accepted view of memory in the brain holds that synapses are the storage sites of memory, and that memories are formed through associative modification of synapses. This view has been challenged on conceptual and empirical grounds. As an alternative, it has been proposed that molecules within the cell body are the storage sites of memory, and that memories are formed through biochemical operations on these molecules. This paper proposes a synthesis of these two views, grounded in a computational theory of memory. Synapses are conceived as storage sites for the parameters of an approximate posterior probability distribution over latent causes. Intracellular molecules are conceived as storage sites for the parameters of a generative model. The theory stipulates how these two components work together as part of an integrated algorithm for learning and inference.