In multimodal machine learning, multiple modalities of data (e.g., text and
images) are combined to facilitate the learning of a better machine learning
model, which remains applicable to a corresponding unimodal task (e.g., text
generation). Recently, multimodal machine learning has enjoyed huge empirical
success (e.g. GPT-4). Motivated to develop theoretical justification for this
empirical success, Lu (NeurIPS '23, ALT '24) introduces a theory of multimodal
learning, and considers possible separations between theoretical models of
multimodal and unimodal learning. In particular, Lu (ALT '24) shows a
computational separation, which is relevant to worst-case instances of the
learning task.
In this paper, we give a stronger average-case computational separation,
where for "typical" instances of the learning task, unimodal learning is
computationally hard, but multimodal learning is easy. We then question how
"organic" the average-case separation is. Would it be encountered in practice?
To this end, we prove that under natural conditions, any given computational
separation between average-case unimodal and multimodal learning tasks implies
a corresponding cryptographic key agreement protocol. We suggest to interpret
this as evidence that very strong computational advantages of multimodal
learning may arise infrequently in practice, since they exist only for the
"pathological" case of inherently cryptographic distributions. However, this
does not apply to possible (super-polynomial) statistical advantages.

多模式机器学习与单模式学习的理论模型、计算分离以及平均情况的计算分离是本篇论文的主要研究内容。

关于多模态与单模态机器学习之间更强的计算分离

On Stronger Computational Separations Between Multimodal and Unimodal  Machine Learning

The observed lensed fraction of high-redshift quasars $(\sim0.2\%)$ is
significantly lower than previous theoretical predictions $(\gtrsim4\%)$. We
revisit the lensed fraction of high-redshift quasars predicted by theoretical
models, where we adopt recent measurements of galaxy velocity dispersion
functions (VDFs) and explore a wide range of quasar luminosity function (QLF)
parameters. We use both analytical methods and mock catalogs which give
consistent results. For ordinary QLF parameters and the depth of current
high-redshift quasar surveys $(m_z\lesssim22)$, our model suggests a
multiply-imaged fraction of $F_\text{multi}\sim 0.4\%-0.8\%$. The predicted
lensed fraction is $\sim1\%-6\%$ for the brightest $z_s\sim6$ quasars
$(m_z\lesssim19)$, depending on the QLF. The systematic uncertainties of the
predicted lensed fraction in previous models can be as large as $2-4$ times and
are dominated by the VDF. Applying VDFs from recent measurements decreases the
predicted lensed fraction and relieves the tension between observations and
theoretical models. Given the depth of current imaging surveys, there are
$\sim15$ lensed quasars at $z_s>5.5$ detectable over the sky. Upcoming sky
surveys like the LSST survey and the {\em Euclid} survey will find several tens
of lensed quasars at this redshift range.

采用最新的星系速度分散函数和探究广泛的类星体亮度函数参数，我们重新审视了理论模型预测的高红移类星体的透镜分数，在当前高红移类星体调查的深度下，我们的模型建议具有两个或多个成像的量约为 0.4％-0.8％，对于最亮的 z≈6 的类星体，预测的透镜比例为约 1％-6％，在上一代模型的预测透镜比例的系统误差可以高达 2-4 倍，并且受星系速度分散函数的支配，无论是理论还是实际观测，该文献在证明高红移类星体的透镜比例时是具有参考意义的。

重新审视高红移类星体的透镜分数

Revisiting the Lensed Fraction of High-Redshift Quasars

Much of the progress made in time-domain astronomy is accomplished by
relating observational multi-wavelength time series data to models derived from
our understanding of physical laws. This goal is typically accomplished by
dividing the task in two: collecting data (observing), and constructing models
to represent that data (theorizing). Owing to the natural tendency for
specialization, a disconnect can develop between the best available theories
and the best available data, potentially delaying advances in our understanding
new classes of transients. We introduce MOSFiT: the Modular Open-Source Fitter
for Transients, a Python-based package that downloads transient datasets from
open online catalogs (e.g., the Open Supernova Catalog), generates Monte Carlo
ensembles of semi-analytical light curve fits to those datasets and their
associated Bayesian parameter posteriors, and optionally delivers the fitting
results back to those same catalogs to make them available to the rest of the
community. MOSFiT is designed to help bridge the gap between observations and
theory in time-domain astronomy; in addition to making the application of
existing models and creation of new models as simple as possible, MOSFiT yields
statistically robust predictions for transient characteristics, with a standard
output format that includes all the setup information necessary to reproduce a
given result. As large-scale surveys such as LSST discover entirely new classes
of transients, tools such as MOSFiT will be critical for enabling rapid
comparison of models against data in statistically consistent, reproducible,
and scientifically beneficial ways.

介绍了 MOSFiT：一个用于解决时间域天文学中观测数据与理论模型之间断裂的 Python 工具，可以自动下载数据，并生成半解析光变曲线拟合结果及业已确定的贝叶斯参数，可将拟合结果更新到数据源以分享给其他人

MOSFiT: 用于瞬变事件的模块化开源拟合器

MOSFiT: Modular Open-Source Fitter for Transients

We review the problem of dark energy, including a survey of theoretical
models and some aspects of numerical studies.

该研究评估了关于暗能量的理论模型和数字研究方面的问题。