text stringlengths 4.64k 21.5k | arxiv_id stringlengths 10 10 | url stringlengths 32 32 | timestamp stringdate 2022-05-25 02:06:08 2023-03-06 02:13:47 | yymm stringdate 2205-01-01 00:00:00 2302-01-01 00:00:00 | language stringclasses 1
value | Qwen3_token_count int64 4.1k 4.1k | start_index int64 0 301k | uuid stringlengths 36 36 |
|---|---|---|---|---|---|---|---|---|
channels $M$} & 80 \\
\multicolumn{1}{c|}{Mel-spectrogram time window $T$} & 27 \\ \hline \hline
\end{tabular}
\label{tab:BaselineDataAttributes}
\end{table}
\noindent
Finally, all the pre-processed data resulted in 2202106 frames of faces, together with 33000 mel-spectro... | 2212.13810 | https://arxiv.org/abs/2212.13810 | 2022-12-29 02:15:30 | 2212 | en | 4,096 | 1,936 | e8158d72-955b-4e30-a832-8d9d0e9424de |
{equation}
\label{eq:functional-recursion}
Q_n(\theta) = \max \left\{0, \ Q_{n - 1} (\theta) + g(x_n, \theta) \right\}
\end{equation}
for $n \geq 1$.
The value of our test statistic will be $\mathcal{Q}_n = \max_\theta Q_n(\theta)$.
At each iteration the functional cost $Q_n(\theta)$ will be the maxi... | 2302.02718 | https://arxiv.org/abs/2302.02718 | 2023-02-07 02:26:13 | 2302 | en | 4,096 | 5,264 | b8996405-f293-4582-abbb-d4311e69a98e |
that one or two features have on the predicted outcome. In practice, more features could be considered but we limit to one or two for representation purposes, i.e. producing 2D or 3D plots. The plot can show whether the relationship between the target and a feature is linear, monotonic, or more complex.\\
The \textit{... | 2209.02057 | https://arxiv.org/abs/2209.02057 | 2022-09-29 02:00:59 | 2209 | en | 4,096 | 43,557 | 117b95b2-f21a-4a76-bb23-182d3c025cf7 |
$ spiked Wigner matrix. Recall the transformation defined in \eqref{eq:alpha_transform} and the transformed matrix $\widetilde M$ in \eqref{eq:transformed-wigner}. We consider a test statistic
\begin{equation} \begin{split} \label{eq:wt L_lambda}
\widetilde L_{\omega} &:= - \log \det \left( (1+\SNRF_g)I - \sqrt{\SNR... | 2301.05331 | https://arxiv.org/abs/2301.05331 | 2023-01-18 02:19:26 | 2301 | en | 4,096 | 17,941 | 1ea1feda-d449-40a2-94e7-364a257f5964 |
^2\log(1/\delta)}}\|M_i\|_2\right),
\end{align*}
and
\begin{align*}
\big|\big< M_i, \mu_2\delta^{\top} \big>\big|\le O\left({\sqrt{k_2d\sigma^2\log(1/\delta)}}\|M_i\|_2\right).
\end{align*}
Finally, for the last term in the right hand side, we have
\begin{align*}
\big|\big< M_i, Q \big>\big| = \Tr(QM_i) \le \Tr... | 2211.15762 | https://arxiv.org/abs/2211.15762 | 2022-11-30 02:02:06 | 2211 | en | 4,096 | 39,060 | 759bb636-f863-4bde-b004-09cf4ecfe49d |
lower. It can be for example a curve or a plane in the $7D$ parameter space. This definition of degeneracy is not rigorous. More precisely, a degenerate distribution is rigorously defined as a distribution whose support has Lebesgue measure equal to zero. However the notion of degeneracy is considered in this work as ... | 2211.02465 | https://arxiv.org/abs/2211.02465 | 2022-11-17 02:09:55 | 2211 | en | 4,096 | 17,815 | 08aae39e-e061-4932-93bb-2d4c806b4a06 |
in parallel with the extension of spectral clustering to the sparse regime by removing nodes of unusually large degrees \cite{chin2015stochastic,le2017concentration}. However, in contrast to that approach, our algorithm does not require any pre-processing step for the observed network.
Similar to existing work on... | 2301.04771 | https://arxiv.org/abs/2301.04771 | 2023-01-13 02:03:29 | 2301 | en | 4,096 | 1,431 | 5d6bd4b0-a893-4c26-a215-12d537db1ff3 |
C_T=\CLS+T$. Let $g\equiv 1$, $T\ge \CLS d$, and $h=O\prc{\E_{p_0}\ve{x}^2 + C_Td(\smf\vee \sms)^2}$. Then
\begin{align*}
\chi^2(q_{(k+1)h}||p_{(k+1)h}) &\leq
\chi^2(q_{kh}||p_{kh}) e^{(-\rc{8C_{T-kh}}+8\ep_1^2)h} + O(\ep_1^2 h + (\sms^2 + \smf^2d)h^2)
\end{align*}
and letting $t=T-Nh$, if $\ep_1<\fc{1}{128C... | 2206.06227 | https://arxiv.org/abs/2206.06227 | 2022-06-14 02:35:21 | 2206 | en | 4,096 | 26,696 | 5a0e6e6d-934b-448a-873f-8349aedb413b |
fty]^d} f_R(x) \frac{\nabla f_R(x + \epsilon)}{f_R(x + \epsilon)} dx\\
&= \int \nabla f_R(x) \left(\frac{ f_R(x - \epsilon) - f_R(x)}{f_R(x)} \right)dx
\end{align*}
since $$\int \nabla f_R(x)dx = 0$$
Now, by the definition of $\mathbbm{1}_R$
\begin{align*}
\mathbbm{1}_R &= \Exp_{x \sim f_R}[s_R(x) s_R(x)^T] = \int_{[-\... | 2302.02497 | https://arxiv.org/abs/2302.02497 | 2023-02-07 02:18:43 | 2302 | en | 4,096 | 10,170 | 3989f5d3-1eac-48fd-9867-ebc03fe26a6e |
designs \citep{tran2019constrained}.
We consider the scenario where querying the objective function is expensive and sample-efficiency is crucial.
In the case of designing the oil sorbent, evaluating the objective function requires manufacturing the material and measuring its properties in a laboratory, requiring si... | 2210.10199 | https://arxiv.org/abs/2210.10199 | 2022-10-20 02:04:41 | 2210 | en | 4,096 | 194 | ba0d99a3-5e00-4a6f-86d7-d40474a3eeac |
\figref{fig:benz}. At ambient conditions, one polymorph is stable (\ccdc{PHENAN08}), while the other is unstable (\ccdc{PHENAN14}, also shown in \figref{fig:benz}(c)). Looking at the same motifs in the unstable and stable phases, we see a shift leftwards as the motifs go from resisting crystallization to weakly bindin... | 2209.10709 | https://arxiv.org/abs/2209.10709 | 2022-09-23 02:05:10 | 2209 | en | 4,096 | 8,946 | c88aa546-5c9e-4ebe-8ec2-2cba370c5729 |
0.01 & \bfseries 0.70$\pm$0.01
& 0.67$\pm$0.00
\\
\bottomrule
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vskip -0.1in
\end{table}
\section{Introduction}
Despite the rapidly advancing field of deep learning ... | 2208.14919 | https://arxiv.org/abs/2208.14919 | 2022-09-01 02:17:58 | 2208 | en | 4,096 | 6,339 | 44070994-4a61-497c-95eb-b5e174afeea7 |
^{K{-}1}\sum\limits_{q=0}^{Q{-}1} \tilde{\nabla} f_{i_{t,k}}\left(w_{i_{t,k},q}^{\tau_{i_{t,k}}^t}\right).
\end{align}
Due to Assumption~\ref{assump:smoothness}, we can infer that $f$ is $L$-smooth, thus
\begin{align}\label{eq:main-0}
f(w^{t{+}1}&) \overset{\eqref{eq:smoothness}}{\leq} f(w^{t}) + \frac{L\eta^2\beta... | 2210.01161 | https://arxiv.org/abs/2210.01161 | 2022-10-05 02:00:51 | 2210 | en | 4,096 | 7,179 | 63bfaef6-ad48-4632-9afd-98fea13bd4f5 |
cal{R} \} }, \dots, \bm{1}_{\{ t_{n} \in \mathcal{R} \} }, \bm{1}_{\{ t_1 \in \mathcal{S} \} }, \dots, \bm{1}_{\{ t_{n} \in \mathcal{S} \} } \right),
\end{equation*}
where $\mathcal{R}:=\{ r_1, \dotsc, r_{k_U} \}$ and $\mathcal{S} := \{ s_1, \dotsc, s_{k_V}\} \cup \{ t_k \}$, as well as $\tilde \Upsilon$ as the su... | 2206.14284 | https://arxiv.org/abs/2206.14284 | 2022-06-30 02:03:23 | 2206 | en | 4,096 | 38,868 | 00477142-6498-4d60-9db7-95fb5b630d04 |
start and once at the end of the window.
Additionally, features from the two preceding and two following 30s-windows are utilized for the epoch of interest.
In total, the feature vector contains 1048 features.
}
\label{fig:feat_extract_windows}
\end{figure}
Figure~\ref{fig:feat_extract_windows} s... | 2207.07753 | https://arxiv.org/abs/2207.07753 | 2022-07-20 02:17:33 | 2207 | en | 4,096 | 3,071 | 823b8f90-a751-4298-ae18-77c1b6c69037 |
22.png} &
\includegraphics[align=c,width=1.4cm]{TBIOM/imgs/RFW/greedy-LMMAES/magface/s_1_train_3.037_test_2.373.png} &
\includegraphics[align=c,width=1.4cm]{TBIOM/imgs/RFW/greedy-LMMAES/magface/s_7_train_1.719_test_2.053.png} &
\includegraphics[align=c,width=1.4cm]{TBIOM/imgs/RFW/greedy-LMMAES/magface/s_2_train_2.828_t... | 2211.13964 | https://arxiv.org/abs/2211.13964 | 2022-11-29 02:18:27 | 2211 | en | 4,096 | 21,601 | 229d6d8b-b314-4e74-8028-8268ee80806f |
for the case of $K = 2$ arms. Finally, we evaluate \ensuremath{\tt BayesElim}\xspace on several synthetic bandit instances and demonstrate the benefit of using the prior in BAI.
One surprising property of our upper and lower bounds is that they are proportional to $1 / \sqrt{n}$, where $n$ is the budget. At a first s... | 2211.08572 | https://arxiv.org/abs/2211.08572 | 2022-11-17 02:05:09 | 2211 | en | 4,096 | 28,882 | 106c4522-24dd-4286-aa7b-9ab6e0c07e4b |
$\Psi$, provide the tightest prediction bands, not only when numerous observations are available, but also in small sample sizes scenario.
Moreover, one can notice that FAR(1)-EK+ do not significantly improves FAR(1)-EK neither in terms of coverage, nor in terms of band size.
We acknowledge that, when $T=... | 2207.13656 | https://arxiv.org/abs/2207.13656 | 2022-07-28 02:19:37 | 2207 | en | 4,096 | 13,012 | 6d3a0cb2-6664-4439-a6c3-f6308e643ae2 |
). Other examples of group synchronization include phase synchronization ($\mathcal{G}=SO(2)$) with applications to cryo-electron microscopy imaging, permutation synchronization ($\mathcal{G}=S_n$) with applications to multi-object matching, and $\mathbb Z_2$-synchronization ($\mathcal{G}=\mathbb Z_2$) with applicatio... | 2206.08994 | https://arxiv.org/abs/2206.08994 | 2022-06-22 02:02:38 | 2206 | en | 4,096 | 281 | 11656b17-0f9e-4e53-b783-ed4b108bae90 |
cal{C}_k}\prt{\ve{x}} \widehat{\mathcal{M}}_k \prt{\ve{x}},
\end{equation}
where $\mathbbm{1}_{\mathcal{C}_k}\prt{\ve{x}}$ is equal to $1$ if $\ve{x}$ is predicted to belong to the cluster $\mathcal{C}_k$, \emph{i.e.}, $\mathcal{M}^{\text{SVC}}\prt{\ve{x}} = \ell_k$ and $0$ otherwise;
\item \emph{Soft recombinatio... | 2211.16909 | https://arxiv.org/abs/2211.16909 | 2022-12-01 02:12:56 | 2211 | en | 4,096 | 3,192 | 096a5d08-5ee8-43aa-ab7f-badd80d98b61 |
(Y)}), we can further determine the specific reconstruct error of each input data point.
\begin{theorem}
\label{Result3}
The reconstruction error of each input data point measured by MSE can be formulated as
\begin{equation}
\label{equation:reconstruct error}
\begin{split}
W(Y) &=\big(Y-\hat{Y}\big)^{\top}\... | 2211.03054 | https://arxiv.org/abs/2211.03054 | 2022-11-08 02:12:37 | 2211 | en | 4,096 | 4,600 | 20e1848c-267f-4619-8319-61d314290a07 |
8}\cdot \frac{\log(2/\delta)}{\epsilon^2} \cdot \mathrm{Id}}\,.$
\item\label[step]{step:clustering-histograms} Pick $\bm \Phi\sim N\Paren{0, \frac{1}{d^*}}^{d^*\times d}\,, \mathbf{q}\overset{u.a.r.}{\sim}\brac{0,b}$ and run the histogram learner of \cref{lemma:private-histogram-learner-high-dimension} with input $\b... | 2301.04822 | https://arxiv.org/abs/2301.04822 | 2023-01-13 02:05:24 | 2301 | en | 4,096 | 4,155 | 8fe348d7-3cee-468b-8301-863dc60edef7 |
temporal discretization schemes are widely used to simulate time-dependent PDEs \citep{NAbook}. However, these methods can be prohibitively expensive, particularly in several spatial dimensions and for long-time integration.
Recently, machine learning based algorithms are increasingly being used for the fast and acc... | 2302.03580 | https://arxiv.org/abs/2302.03580 | 2023-02-08 02:19:35 | 2302 | en | 4,096 | 123 | 5af3ea31-103c-4aba-90b0-3b526c63e88e |
(1-f) \times \lambda_B^k \exp(-\lambda_B)
} \\\\
%
&= \frac{
f \times (k-\lambda_A)\lambda_A^{k-1} \exp(-\lambda_A)
}
{
f \times \lambda_A^k \exp(-\lambda_A)
+ (1-f) \times \lambda_B^k \exp(-\lambda_B)
}
\label{eq:diff_D_lambda}
\end{a... | 2302.04081 | https://arxiv.org/abs/2302.04081 | 2023-02-09 02:13:19 | 2302 | en | 4,096 | 3,134 | 963f81be-5c13-4665-bdeb-50035bdd7c65 |
system given in \eqref{eq:system}. Suppose there is a class of controllers $g_\theta(\cdot)$, parameterized by $\theta \in \Theta$. The goal is to design a controller $g_\theta(\cdot)$ that minimizes a control cost,
\begin{subequations}\label{eq:policy_opt}
\begin{align}
\min_{\theta} J(\theta) = &\sum\nolimits_{... | 2206.01704 | https://arxiv.org/abs/2206.01704 | 2022-06-06 02:18:27 | 2206 | en | 4,096 | 122 | e86339f8-c074-4f24-a4b1-12815fe6c40c |
} is the most popular measure of the similarity between two data clusterings. Nevertheless, the RI can't ensure a result that is close to 0 when the clustering is utterly arbitrary. Accordingly, we also use the ARI \cite{hubert1985comparing}, which corrects the flaws of RI in some way. For both RI and ARI, they take va... | 2212.03122 | https://arxiv.org/abs/2212.03122 | 2022-12-12 02:13:32 | 2212 | en | 4,096 | 6,884 | 249a46e1-304b-475c-8d37-3a705df7c2c1 |
known to suffer from the curse of dimensionality \citep{scott1991feasibility}. Thus, inspired by \citep{wang2021projected, chen2020projected}, we consider estimating the density in a low-dimensional subspace in which the likelihood of data changes significantly. To find such a subspace, we use the gradient information... | 2206.00944 | https://arxiv.org/abs/2206.00944 | 2022-06-03 02:12:47 | 2206 | en | 4,096 | 5,489 | bf44d839-59c5-495a-a77a-b6f50eee751b |
)= \Theta(\rho)$ and (ii) inequality~\eqref{eq:HellTight} holds.
Using \Cref{fact:testing}, we have the following:
\begin{align}
\label{eq:testinglowerbd}
\mathrm{n}^{\!\ast}(p,q) = \Theta \left( \frac{1}{d_{\mathrm{h}}^2(p,q)} \right) = \Theta \left(\frac{1}{\rho} \right) = \Theta\left( n_0 \right).
\end{align}
T... | 2206.02765 | https://arxiv.org/abs/2206.02765 | 2022-06-07 02:35:40 | 2206 | en | 4,096 | 16,035 | 438ac410-a535-4df7-b112-c6f6486d1cc0 |
(w^*, Z) \big) - \frac{\sigma}8 \|w - w^*\|_\H^2 \right] \right\}
\\&
\leq \mathbb E_\varepsilon \exp\left\{ \lambda \sup\limits_{w \in \mathcal W[2^k r, 2^{k+1} r]} \left[ P_n \varepsilon \big( f(w, Z) - f(w^*, Z) \big)\right] - 4^{k-1}\sigma \lambda r^2 / 2\right\}
\\&
\leq \mathbb E_\varepsilon \exp\... | 2302.10726 | https://arxiv.org/abs/2302.10726 | 2023-02-22 02:16:59 | 2302 | en | 4,096 | 9,252 | 853bfb90-27ef-49bd-9c4d-cd2ee778942e |
{p}}$ of our methods to that of the FoS.
\textbf{Design 1 (nonlinear scenario):} We generate 20 random predictors ($K=20$), where $X_{k}$ are i.i.d. uniform random variables from $[a, b]$ with $a \in \{-4,-3, -2, -1, 0\}, b \in \{3,4,5,6,7\}$ for all $k$. We apply the polynomial transformation to 50\% of the continuo... | 2208.05776 | https://arxiv.org/abs/2208.05776 | 2022-08-12 02:11:26 | 2208 | en | 4,096 | 17,329 | 40611a90-69ef-49cd-b61c-65ad8c88f1d4 |
ldots,m\}$ and $j,j'\sim {\rm Unif}\{1,\ldots,n\}$, then
\begin{align*}
&{1\over m}\sum_{i=1}^m \mathbb{E}_{i',j,j'}\left(W_{i,j,i',j'}\|\Theta_\beta(X_{i,j})-\Theta_\beta(X_{i',j'})\|^2\right)\\
=&{1\over n^2m^2}\sum_{i_1,i_2=1}^m\sum_{j_1,j_2=1}^nW_{i_1,j_1,i_2,j_2}\|\Theta_\beta(X_{i_1,j_1})-\Theta_\beta(X_{i_... | 2211.00460 | https://arxiv.org/abs/2211.00460 | 2022-11-02 01:14:27 | 2211 | en | 4,096 | 12,209 | 82bdecf1-b009-41d1-a950-cab5c26d5e0f |
decomposition}
\label{alg:ikd}
\begin{algorithmic}[1]
\Function{ikd}{$\bm X\in \mathbb{R}^{T\times N}, f$}
\State $\bm S \gets \frac{1}{N-1} \rbr{\bm X - \bar{\bm X}\bm{1}^\mathsf{T}} \rbr{\bm X - \bar{\bm X}\bm{1}^\mathsf{T}}^\mathsf{T}$ \Comment $\bm S$ serves as an estimator of the covarianc... | 2211.05961 | https://arxiv.org/abs/2211.05961 | 2022-11-14 02:04:34 | 2211 | en | 4,096 | 4,650 | aaf69519-6045-4128-8b26-371aa3791392 |
where $\bm{q}_k$ denotes the $k$-th row of $\bm{Q}$, which is a $d_b \times d_r$ matrix such that $\bm{Q}\bm{Q}^\top = \bm{W} + \bm{D}$, and $\bm{D}$ is an arbitrary diagonal matrix that ensures $\bm{W} + \bm{D}$ is positive semi-definite. The above is a continuous relaxation of the Boltzmann Machine distribution on $\... | 2205.09559 | https://arxiv.org/abs/2205.09559 | 2022-05-31 02:21:29 | 2205 | en | 4,096 | 7,984 | 56901025-a8e5-4d8a-aaf8-9e3ef3a4f58e |
right. ,
$$
$$
m_i(t)=\left\{ \begin{array}{rcl} x_i(t) & \mbox{if} & a(t)=i \\ 0 & \mbox{if} & a(t) \neq i
\end{array}\right. ,
$$
$$
\xi_i(t)=\sum_{s=1}^{t}m_i(t)\eta_i(t)=\sum_{s \in T_i(t)} x_i(s)(r_i(s)-x_i(t)^{\top}\mu_i).
$$
Note that $\operatorname{det}\left({A}_{i}(t)\right) \leq (T_i(t))^d$. For $d \geq 2$, ... | 2206.12463 | https://arxiv.org/abs/2206.12463 | 2022-06-28 02:00:58 | 2206 | en | 4,096 | 12,650 | a3750387-410b-45d7-a578-25cf39a0fd17 |
ls-fs} requires a certain level of continuity of the distribution of covariate $X$. These extra conditions can be easily satisfied, e.g., when $X$ obeys a non-degenerate normal distribution. To study the Bahadur representation, we need a generalization of the coupling we defined in Equation~\eqref{eq:coupling}. Now let... | 2212.14883 | https://arxiv.org/abs/2212.14883 | 2023-01-02 02:21:14 | 2212 | en | 4,096 | 16,732 | decb0dc6-b17d-4ddd-92b0-16a75f868a4e |
\subsubsection{Targeting for disparate impact}
Table \ref{tab:dnn-table} presents the three $2\times 2$ tables
comparing the results of the unconstrained DNN classifier ($\hat{Y}^\star$) and three DNN classifiers ($\hat{Y}$) trained on the dataset \textit{Adult}:
(i) only with the DI constraint, (ii) with the ... | 2301.08375 | https://arxiv.org/abs/2301.08375 | 2023-01-23 02:04:06 | 2301 | en | 4,096 | 7,774 | 6e1948ee-c691-4f44-9577-1bed0f68430d |
}'(v_t)) \|^2 | \mathcal{F}_t \right] } \right] \\
&\leq 2 \eta \sigma_2 \sum_{t=0}^T \sqrt{ \mathbb{E} \left[ \| \nabla f( v_t - \eta \epsilon_{t+1}'(v_t)) \|^2 \right] } \\
&\leq 2 \eta \sigma_2 \sqrt{(T+1)\sum_{t=0}^T \mathbb{E} \left[ \| \nabla f( v_t - \eta \epsilon_{t+1}'(v_t)) \|^2 \right] } \\
&\leq 2 \eta \s... | 2302.09376 | https://arxiv.org/abs/2302.09376 | 2023-02-21 02:10:02 | 2302 | en | 4,096 | 16,089 | db4cb9b3-4837-4ba5-97b6-5328f19d69e8 |
have geometric properties of
length $q^{l}:=N_l^{-1}\left(h^{l}\right)^Th^{l}$ and the pairwise covariance $q^{l}_{12}:=N_l^{-1}\left(h_1^{l}\right)^T h_2^{l}$ of two inputs $x^{0,1}$ and $x^{0,2}$ which propagate through the network according to functions of the network entries' variances and nonlinear activation $(... | 2301.13710 | https://arxiv.org/abs/2301.13710 | 2023-02-01 02:18:38 | 2301 | en | 4,096 | 465 | 1214a8b6-3a8a-42ad-bca8-1039216715ec |
}^K\widehat{p}\{x,Y=k\}\textrm{exp}(\theta_k^\top T(x) + \alpha_k)\right)\\
& \text{subject to }\textstyle \int_{\mathcal{X}} \sum_{k=1}^K\widehat{p}\{x,Y=k\}\textrm{exp}(\theta_k^\top T(x) + \alpha_k)dx = 1
\end{aligned}
\right.
\label{eq:robust-distribution-matching}
\end{equation}
where $D$ is a discrepancy between ... | 2205.13577 | https://arxiv.org/abs/2205.13577 | 2022-05-30 02:00:36 | 2205 | en | 4,096 | 32,787 | e79165eb-dfa5-439c-9280-1a738c6aeafe |
use the rate-double robustness property of the target and \Cref{lemma-rates} to infer the required rate conditions on $(\epsilon, \lambda)$ that would retain the beneficial statistical properties of the canonical AIPW estimator. The next result combines our previous characterizations with the standard analysis of AIPW... | 2208.13701 | https://arxiv.org/abs/2208.13701 | 2022-09-01 02:07:31 | 2208 | en | 4,096 | 9,328 | 1a75f2bb-d4c1-417a-b7c7-b1c33a0ddfbf |
is not asymptotically negligible. We remind the reader that once the aggregation cost is negligible, SATL is still rate-optimal based on Theorem \ref{lowerbound theorem}.
Turning back to the excess risk of SATL, we can see that it is sharper than the excess risk of fitting OFLR with the target dataset only for a cert... | 2206.04277 | https://arxiv.org/abs/2206.04277 | 2022-06-10 02:08:41 | 2206 | en | 4,096 | 21,160 | 62792649-e5cb-401e-ae80-22b2255c0193 |
{\Large$\color{black}\thicksim$} & \text{\Large$\color{black}\thicksim$} & \text{\Large$\color{black}\thicksim$} \\
Forward MultiPass PI & G & \text{\Large$\color{black}\checkmark$} & \text{\Large$\color{black}\checkmark$} & $\times$ & \text{\Large$\color{black}\checkmark$} & \te... | 2211.08943 | https://arxiv.org/abs/2211.08943 | 2022-11-17 02:14:57 | 2211 | en | 4,096 | 5,846 | ba9219ad-9f24-4dbb-8a0d-35bf8178238c |
are sorted following the same order at the lower panel.
\textbf{Bottom:} 2D plot of the available ISs. The primary IS cost is set to 1 and the auxiliary ISs cost to 0.2.}
\label{fig:braninmultiple}
\end{figure*}
\subsection{XGBoost hyperparameter tuning}
We now assess the performance of rMFBO on a real-world hyperp... | 2210.13937 | https://arxiv.org/abs/2210.13937 | 2022-10-26 02:15:01 | 2210 | en | 4,096 | 9,876 | 032d71da-d299-403d-9ab3-77e011539d13 |
,r,s)$ in practice; see Section \ref{sec:empirical} for empirical evidence.
Furthermore, in \eqref{eq:model-scalar}, since $\ell_{h,k}(\boldsymbol{\omega})$'s are scalars, they are applied to $\boldsymbol{y}_{t-h}$'s coordinate-wisely and hence do not change the cross-sectional information. More concretely, this c... | 2209.01172 | https://arxiv.org/abs/2209.01172 | 2022-09-29 02:17:25 | 2209 | en | 4,096 | 10,156 | b2c46fc8-92a8-472c-bd35-6cfb73b697ea |
Each stage will consist of training the network for $\lfloor N / T \rfloor$ epochs. Between any two stages, we apply re-initialization as described earlier. This ensures that the computational cost in terms of the total number of gradient steps is approximately equal to the cost of standard training for $N$ epochs. Th... | 2206.10011 | https://arxiv.org/abs/2206.10011 | 2022-06-22 02:37:53 | 2206 | en | 4,096 | 2,330 | c782c36b-1d43-4cfe-bc52-85c7730b7d53 |
}^{\textsf{A},\ddagger}_{h+1/2}) - \hat L^{\textsf{A},j,\pi}_h(\bm{\theta}^{\textsf{A},j}_h, \hat \bm{f}^{\textsf{A},j}_h; \bm{\theta}^{\textsf{A},\ddagger}_{h+1/2}) \right |\\
& \qquad \qquad + \left | \hat L^{\textsf{A},j,\pi}_h(\bm{\theta}^{\textsf{A},j}_h, \hat \bm{f}^{\textsf{A},j}_h; \bm{\theta}^{\textsf{A},\dda... | 2212.12167 | https://arxiv.org/abs/2212.12167 | 2022-12-26 02:06:26 | 2212 | en | 4,096 | 32,971 | 60215dde-9e22-4dc3-abee-bc57831caddb |
(\X_{\indexI}\right)}}{\argmin } \mathbb{E}[Y \mid \DO{\X_{\indexI}}{\x_{I}}, \mathcal{G}] \text{,}
\end{align}%
where $D\left(\X_{\indexI}\right)$ denotes the interventional domain of $\X_{\indexI}$ and the causal effect depends on the true graph $\mathcal{G}$. Solving the problem in \cref{eq_objective} is challenging... | 2208.10981 | https://arxiv.org/abs/2208.10981 | 2022-08-24 02:19:32 | 2208 | en | 4,096 | 2,926 | a8be89da-4490-4202-883f-a37bb9669a0c |
both (\ref{eqn:Definition1}) and (\ref{eqn:Definition2}) are equivalent to the Kronecker delta
function, denoted by $\delta(\cdot)$. Notice that $\delta_{=}(x_k, x_\ell)$ and $\delta_{+}(x_k, x_\ell)$ are both
equal to $\delta(x_k - x_\ell)$ when $\mathcal{X} = \mathbb{Z}/2\mathbb{Z}$ (i.e., the binary field).
In th... | 2208.05333 | https://arxiv.org/abs/2208.05333 | 2022-08-11 02:13:38 | 2208 | en | 4,096 | 2,268 | 1a18a3e9-8d69-4c6d-b58e-9433601df083 |
star}_{x_ga_g, x_{h+1}} \Big) \Big\}.
\end{align}
then we have $-\grad F^{{\sf Tr}}_{{\sf bal}}(M)=\phi(\lambda, m)$, where $\phi$ is as defined in Eq. (\ref{equation:phi-lambda-m}).
\end{lemma}
\begin{algorithm}[t]
\caption{Balanced EFCE-OMD (OMD form; equivalent FTRL form in Algorithm~\ref{algorithm:balanced-ef... | 2205.15294 | https://arxiv.org/abs/2205.15294 | 2022-06-03 02:21:04 | 2205 | en | 4,096 | 14,891 | 0adefcf6-a3c3-4a3c-80d5-7c1402358842 |
$ and $X'$, $\mathcal{A}(X)$ and $\mathcal{A}(X')$ are $({\varepsilon}, \delta)$-indistinguishable.
\end{definition}
In local DP, users interact with the server and send outputs of randomized algorithm. In the fully adaptive case, they can communicate with the server in an arbitrary order with adaptive interaction. F... | 2208.04591 | https://arxiv.org/abs/2208.04591 | 2022-08-10 02:09:36 | 2208 | en | 4,096 | 7,860 | 9b6168ba-709b-4f2e-b926-26d1a85195af |
bm{Q}_k^{(1)}[1:m+1])=m+1 \mid \{\bm{x}_i\}_{i \in M}) = 1.\]
This is true for any $\{\bm{x}_i\}_{i \in M}$ satisfying $\rank(\bm{Q}_k^{(1)}[1:m])=m$, which by inductive assumption is true with probability 1 and so marginalizing over this event we have $\mathbb{P}(\rank(\bm{Q}_k^{(1)}[1:m+1])=m+1) = 1$. Taking $m = d -... | 2207.03609 | https://arxiv.org/abs/2207.03609 | 2022-07-11 02:04:16 | 2207 | en | 4,096 | 29,945 | 666a1fc7-cc80-467f-931e-71e35c3277f7 |
$ and $\hat{\theta}_n^{\textrm{ATET}}$.
The algorithm for estimating the ATE and ATET are conceptually similar to Algorithm~1. We delegate the detailed outline to Appendix \ref{app:algorithms}. Mean and median aggregation over cross-fitting repetitions are implemented as outlined in Remark~2.
\section{DDML with Ins... | 2301.09397 | https://arxiv.org/abs/2301.09397 | 2023-01-24 02:20:45 | 2301 | en | 4,096 | 32,080 | 5e8b2416-b2d7-43e6-b822-85508cf87b36 |
of=x, xshift=1.5cm] (filter1)
{${\gfilter}(\mA{1})$};
\node [coordinate, right of=filter1,xshift=0.3cm](test5){};
\node[draw,below of=x, yshift=-0.3cm, left of=x, xshift=1.5cm]
(filterC) {${\gfilter}(\mA{\mulG})$};
\node[ above of=filterC, yshift=0.4cm]
(filters2) {$\vdots$};
\node [co... | 2207.14019 | https://arxiv.org/abs/2207.14019 | 2022-07-29 02:17:19 | 2207 | en | 4,096 | 1,638 | c22f9488-e125-47e7-b4f6-dd316412df03 |
}_{(-)}$ for $\alpha\ge0$. We reparametrize the contrast parameter for a more direct comparison with our method.}
To gain intuition, consider the special case where positive samples correspond to noisy signal and negative samples are pure noise; that is,
\begin{align}\label{eq:signal}
{\bf x}_t=\delta_t\times\text... | 2211.07723 | https://arxiv.org/abs/2211.07723 | 2022-11-16 02:01:32 | 2211 | en | 4,096 | 1,100 | 04b930a1-c7a2-48dc-bcd2-8e543c96a9ca |
properties of anytime optimal online learners}
The following lemma gives a characterization of the optimal post-$S$ mistake bound of anytime optimal online learning in terms of the Littlestone dimension of the version space. The proof is implicit in the proof of Theorem 3 from \cite{littlestone1988learning}.
\begin{... | 2302.04357 | https://arxiv.org/abs/2302.04357 | 2023-02-10 02:02:44 | 2302 | en | 4,096 | 6,763 | 03f39549-3a52-4640-9e87-bb2171e29138 |
manifold $M$. Let $f: U \rightarrow M$ be a generative function that is a smooth submersion from an open set $U \subseteq \mathbb{R}^n$ onto the data manifold. Then, for each datum $p$ with latent code $x_p$ (ie. $f(x_p) = p$), there exists a local smooth reparametrization $\phi_p$ such that the first $n$ reparametriz... | 2210.07857 | https://arxiv.org/abs/2210.07857 | 2022-10-19 02:12:05 | 2210 | en | 4,096 | 23,532 | 80dd228f-140f-4b6e-aa0c-54cb2c390d23 |
a certain distance, $\delta$, away from the origin.
The small interval around the
origin is defined by $[-\delta,+\delta]$. Inside this region, the L2 loss
function is used since it is continuous. Outside this region, the L1 loss
is used but great care is taken to match the derivatives
at the interface between the tw... | 2208.04564 | https://arxiv.org/abs/2208.04564 | 2022-08-15 02:04:40 | 2208 | en | 4,096 | 3,052 | 5d12a68c-136c-4ea1-8711-ecc23c73c409 |
one wants to split the batch for reasons of efficiency. In our case, the separation concerned elements belonging to different classes, so as to efficiently collect the gradient at associated with each of them. Group normalization needs an additional parameter to fix the number of blocks within which to group features ... | 2207.00391 | https://arxiv.org/abs/2207.00391 | 2022-07-04 02:14:20 | 2207 | en | 4,096 | 12,045 | f0376495-5cdb-407f-9410-65ed95902ae2 |
} = $]
{
\tiny
\setstretch{0.25}
charact, book, music, role, featur, stori, star, titl, direct, version,
\par
}
\item[$\mathcal{V}_{107} = $]
{
\tiny
\setstretch{0.25}
larger, less, below, greater, higher, smaller, abov, reduc, lower, low,
\par
}
\item[$\mathcal{V}_{108} = $]
{
... | 2210.01679 | https://arxiv.org/abs/2210.01679 | 2022-10-05 02:17:52 | 2210 | en | 4,096 | 95,982 | bb351258-a475-41d8-8697-7631b64818a4 |
}_a(\scal{\mu^*}{X_{t,a}})$, that is the probability that a reward sampled from the distribution of arm $a$ is lower than the reward of $X_{t,a}$. We argue that the relative rank, allows to have a fair way of comparing individuals from different groups and thus introduce the following fairness definition.
\begin{defi... | 2206.03150 | https://arxiv.org/abs/2206.03150 | 2022-06-08 02:13:51 | 2206 | en | 4,096 | 2,926 | 5d2f257d-4fa6-4b3b-98fd-0b03d97a4647 |
features for the 4D model and found that it still performs fairly well.
\section{Architecture}
\section{Experimental Evaluation}
\label{sec:experiments}
Evaluating the effectiveness of our methods is difficult because there are very few tensors with known rank so we can solely rely on real-world data. As such, w... | 2206.09316 | https://arxiv.org/abs/2206.09316 | 2022-06-22 02:14:42 | 2206 | en | 4,096 | 5,432 | 11f6e564-9086-4689-9461-ff512c8a56dc |
end{equation}
where $x'_{i,j}\in\B_{sem}(x_i^u)$ is the $j$-th augmented data for the $i$-th unlabeled sample obtained by class-invariant data augmentation.
On the other hand, term (3) depends on the Rademacher complexity of the hypothesis class and the size of the training data. A number of theoretical analysis on th... | 2209.12611 | https://arxiv.org/abs/2209.12611 | 2022-09-27 02:29:42 | 2209 | en | 4,096 | 6,952 | 42a436a0-69ca-44dc-a8c2-11de37265dcb |
thing_repulsive}, we show that three main \emph{linearized} continuous GNN models are either \emph{smoothing} or $\mathrm{LFD}$ as per Definition \ref{defn: lfd_hfd}.
The linearized {PDE-GCN$_{D}$} model \cite{eliasof2021pde} corresponds to choosing $\beta = 0$ and $\boldsymbol{\Omega} = \mathbf{W} = \mathbf{K}(t)^\to... | 2206.10991 | https://arxiv.org/abs/2206.10991 | 2022-08-16 02:30:10 | 2206 | en | 4,096 | 12,657 | 9a38ce36-b945-4b18-bfea-2f22314c8d63 |
on an unprecedented proteome-wide scale by using AlphaFold-predicted structures~\cite{jumper2021highly} to predict IDRs, in contrast to previous work which considered far fewer experimentally derived structures.
\begin{figure}[ht]
\centering
\includegraphics[width=\textwidth]{figures/alphafold/widths.pdf}
... | 2301.09633 | https://arxiv.org/abs/2301.09633 | 2023-02-03 02:17:40 | 2301 | en | 4,096 | 12,209 | 7d652726-ed0e-4725-9660-e0788ed751dc |
of 64 and a learning rate of $1e^{-3}$. The number of dimensions of the latent $z$ is $10$ for the DSprites dataset and $6$D for the 3D Shapes dataset. We prefer dimension $2$ to analyze the reconstruction and generative quality of the model. In all cases, the Adam optimizer is used for training the models.\\
\textbf{... | 2206.04225 | https://arxiv.org/abs/2206.04225 | 2022-06-10 02:06:47 | 2206 | en | 4,096 | 9,142 | f4c3d4d1-7096-472e-b116-24e94beb3cfd |
psi}(s_i,a_i) - \log \pi_\psi(a'| s'_i) \Big)^2 \nonumber
\end{align*}
for a replay buffer $\mathcal{D}=\{(s_i,a_i,r_i,s'_i)|i=1,\ldots,N\}$ comprising a non-i.i.d. set of observations collected from the target control environment. We choose the hypothesis space in a parameterized fashion as $\mathcal{H}_{\psi} := \{ ... | 2301.12776 | https://arxiv.org/abs/2301.12776 | 2023-01-31 02:26:26 | 2301 | en | 4,096 | 6,088 | f32e2e28-119d-4cd9-b296-10c8639b6eb5 |
icolumn{2}{c}{$49.57 \pm 0.44$} &
\multicolumn{2}{c}{$45.20 \pm 0.22$} \\ \cmidrule(l){2-14}
&
Focal &
\multicolumn{2}{c}{$\mathbf{36.11 \pm 0.20}$} &
\multicolumn{2}{c}{$\mathbf{37.64 \pm 0.18}$} &
\multicolumn{2}{c}{$\mathbf{39.77 \pm 0.10}$} &
\multicolumn{2}{c}{$\mathbf{42.37 \pm 0.16}$} &
\multicol... | 2206.02286 | https://arxiv.org/abs/2206.02286 | 2022-06-07 02:20:27 | 2206 | en | 4,096 | 12,985 | 1c8d0cca-d8bd-4596-8bd9-80580d613928 |
reconciliation framework by \cite{Panagiotelis2020}, in the Gaussian case with a quadratic loss function.
So far, there is however no approach to reconcile count time series in a principled way.
\section{Reconciliation of Count Variables: a Minimal Example}
\label{sec:toyExample}
Consider the reconciliation of the ... | 2207.09322 | https://arxiv.org/abs/2207.09322 | 2022-07-20 02:19:59 | 2207 | en | 4,096 | 3,577 | 7a63e195-3bec-4757-a9da-42daa845fb18 |
group memberships:
\begin{equation*}
DEO = |P(\hat{Y}|A=1, Y=1) - P(\hat{Y}|A=0, Y=1)|
\end{equation*}
Additional notions of fairness include \emph{calibration} (CAL) \citep{kleinberg2016inherent}, which ensures that predictions are calibrated between subgroups, i.e. $Y \perp \!\!\! \perp A\ |\ \hat{Y}$. For a com... | 2211.06138 | https://arxiv.org/abs/2211.06138 | 2022-11-14 02:09:58 | 2211 | en | 4,096 | 1,629 | 970edf1b-2f8d-4bf3-96f8-4a3c452c8788 |
through the majority votes of base learners trained from disjoint data. Given the number of partitions $k$ as a hyperparameter, DPA is a learner constructed using a deterministic base learner $f$
and a hash function $h: X \times Y \to [k]$ mapping labeled samples to integers between $0$ and $k-1$.
The construction is... | 2208.03309 | https://arxiv.org/abs/2208.03309 | 2022-08-08 02:16:38 | 2208 | en | 4,096 | 4,400 | af509fa8-ff3c-4a5b-9b5a-ea340bb0f1dd |
gp{hiv}$ status should guarantee a tailored improvement performance for the sensitive group, $\textgp{hiv}=\textgp{+}$. Otherwise, it would be better not to solicit this information in the first place as the information could inflicts harm when leaked~\citep[see e.g.,][]{bansal2010impact}.
\item {\em Self-Re... | 2206.02058 | https://arxiv.org/abs/2206.02058 | 2022-06-07 02:12:09 | 2206 | en | 4,096 | 10,523 | 23f1d89b-6778-461d-b55e-8d8ccaaa0b5b |
|c|c|c|c|c|c|c|c|}
\hline
Data & Metrics & Auto & WARIMA & ETS & SETAR & LightGBM & ARNN & RNN & TCN & DeepAR & Transformer & {\color{blue}{W-Trans.}} \\
& & ARIMA \cite{hyndman2008automatic} & \cite{aminghafari2007forecasting} & \cite{hyndman2008forecasting} & \cite{tong1990non} & \cite{k... | 2209.03945 | https://arxiv.org/abs/2209.03945 | 2022-09-09 02:18:48 | 2209 | en | 4,096 | 6,491 | a11f9e3f-1a3e-4cdc-a96b-e382ab6bfec6 |
||}{Logistic noise} \\
\hline
\hline
- & RankG & AbPNL & RESIT \\
\hline
100 & \textbf{1.05 $\pm$ 1.3} & 3.82 $\pm$ 1.15 & 2.5 $\pm$ 1.24\\
500 & \textbf{0.39 $\pm$ 0.79} & 2.84 $\pm$ 1.02 & 2.46 $\pm$ 1\\
1000 & \textbf{0.14 $\pm$ 0.47} & 3.01 $\pm$ 1.14 & 2.53 $\pm$ 1.11\\
1500 & \textbf{0 $\... | 2302.12341 | https://arxiv.org/abs/2302.12341 | 2023-02-27 02:02:27 | 2302 | en | 4,096 | 24,892 | f74d6f8f-7865-425c-844c-2cbc41f1351b |
the integrality assumption, the best known lower bound remains two hidden layers \cite{mukherjee2017lower}, which is already valid for the simple function $\max\{0,x_1,x_2\}$.
While the integrality assumption is rather implausible for practical neural network applications where weights are usually tuned by gradien... | 2302.12553 | https://arxiv.org/abs/2302.12553 | 2023-02-27 02:10:39 | 2302 | en | 4,096 | 1,508 | f7aa0dd7-3978-4d44-867c-655b23a1ca62 |
tab:runtimes} shows typical run times for each task.
\begin{table*}
\centering
\begin{tabular}{l|c|c|c|c}
& PCA & tSNE & UMAP & PED \\\hline
Gaussian & 0.09 & 901.94 & 180.46 & 294.09 \\
Bernoulli & 0.11 & 923.85 & 174.71 & 267.81 \\
Poisson & 0.10 & 698.34 & 176.33 & 189.87... | 2211.05943 | https://arxiv.org/abs/2211.05943 | 2022-11-14 02:04:02 | 2211 | en | 4,096 | 27,204 | 17777dbb-f8f6-4bf0-bef5-da7d0adc2ab2 |
\rm eff})}$, we have $\|
\Phi(w^{\sigma}_\ast) - \varphi_\ast \|_2^2 = O(\sigma^2)$
while
$\|
\Phi(w^{\sigma, (\rm eff)}_\ast) - \Phi(w^{\sigma}_\ast) \|_2^2 = O(\sigma^{3})$.
\end{restatable}
Note that (a) we characterize the prediction function $\Phi$ taken at the various minimizers to characterize asymptotic equi... | 2206.04613 | https://arxiv.org/abs/2206.04613 | 2022-06-13 02:17:43 | 2206 | en | 4,096 | 6,339 | 1973c6c1-201e-4e72-9c56-b81703edc5e4 |
{flalign}
where $\mathbb{E}[\bullet]$ is the expectation,
${\rm KL}[\bullet\|\bullet]$ is the Kullback-Leibler (KL) divergence,
and $q({\bf W}_v)$ is a variational distribution
that approximates the true posterior $p({\bf W}_v \mid \{\bm{y}_v\})$.
On the basis of the mean-field approximation~\cite{bishop:pattern},
we t... | 2206.12141 | https://arxiv.org/abs/2206.12141 | 2022-06-27 02:08:47 | 2206 | en | 4,096 | 8,856 | 1029432e-a4c7-46cf-b260-2cec3e0bc8a5 |
_{T\to\infty} \mathbb{E}\left[ \mathcal{U}\left( \frac{1}{T} \sum_{t=0}^{T-1} \bbf\left(\bbH_t, \bbp^{\bbphi}\left(\bbH_t, \bbmu_{\lfloor t/T_0 \rfloor}; \bbphi^{\star}\right) \right) \right)\right] \nonumber \\
&\geq P^{\star} - \ccalO(c\eta_{\mu}) - \ccalO(\epsilon)
\label{eq:thm_optimality_state_augmented}
\end{alig... | 2210.16412 | https://arxiv.org/abs/2210.16412 | 2022-11-01 01:03:20 | 2210 | en | 4,096 | 2,926 | 9ae3cfe6-560f-451f-8f16-c8814da58b4a |
tilde{\psi}_{f_0}\|_L^2 = \int_0^1 a(t)^2 f_0(t) dt- \psi(f_0)^2$. We take as prior the exponentiated Gaussian process prior \eqref{def:gp_inducement} with $W$ a rescaled squared exponential process (Example \ref{ex:SE}) with length scale $k_n = n^{-\frac{1}{1+2\gamma}}$. Figure \ref{fig:gp_bvm_condition_illustration} ... | 2301.08158 | https://arxiv.org/abs/2301.08158 | 2023-01-20 02:15:34 | 2301 | en | 4,096 | 17,583 | 6b3e5845-cd05-4da3-83f2-440eab86a617 |
KCDE.
In \textbf{AR}, GARCH slightly outperforms
{FlexCodeTS} \ and both of them outperform NNKCDE.
The high performance of GARCH in
\textbf{AR} is expected, since
it is the only scenario in which
the GARCH parametric model is correctly specified.
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{fi... | 2301.09671 | https://arxiv.org/abs/2301.09671 | 2023-01-25 02:00:40 | 2301 | en | 4,096 | 6,446 | 2af45248-a59b-4733-943b-90ff4e1b8af5 |
^4.
\end{align*}
To bound the variance, one can proceed with
\[
\E\norm{\nabla \ell - \nabla{\cal L}}^2\leq 3\E\norm{a - \E[a]}^2 + 3\E\norm{b-\E[b]}^2 + 3\E\norm{c - \E[c]}^2.
\]
Let us begin with the part in $a$,
\begin{align*}
&\E\bracket{\norm{\frac{1}{m}\sum_{i\in[m]}\phi(\... | 2302.02774 | https://arxiv.org/abs/2302.02774 | 2023-02-07 02:28:17 | 2302 | en | 4,096 | 49,325 | 81589963-4b4b-4601-ad26-763c74014230 |
{williamson_unified_2020, jinglei2018, sapp2014targeted}. Retraining often acts as the best benchmark to evaluate other VI estimation methods \cite{hooker_benchmark_2019} due to its accuracy, yet it is computationally infeasible in high-dimensional settings.
Other methods for VI estimation include knockoff methods \cit... | 2207.09097 | https://arxiv.org/abs/2207.09097 | 2022-07-20 02:09:36 | 2207 | en | 4,096 | 348 | e16aa6c3-2bfa-4f1f-b163-d3f0647905be |
rightarrow s_{n})$, we define the path regularization via optimal transport as follows:
\begin{equation}\label{def: eq path regularization}
\begin{aligned}
\mathcal{L}_{\text{OT}}(\tau ) := \sum_{t=0}^{n-1}{ \text{OT}_{\mathbf{C}_{t;\theta}}\left ( P_F(\cdot| s_t;\theta),P_F(\cdot| s_{t+1};\theta) \right ) ... | 2209.15092 | https://arxiv.org/abs/2209.15092 | 2022-10-03 02:03:14 | 2209 | en | 4,096 | 5,545 | 3e021a65-ac43-4d9f-823e-3b6d509fed28 |
_ablation_fid_cnn_acc}). All results for MNIST and Fashion-MNIST on all three privacy settings ($\varepsilon{=}\{0.2, 1, 10\}$) can be found in~\Cref{tab:noise_augmentation_ablation}.
\input{tables/noise_augmentation_ablation}
\subsubsection{Diffusion Model Config} \label{sec:diffusion_model_backbone_ablation}
In the m... | 2210.09929 | https://arxiv.org/abs/2210.09929 | 2022-10-19 02:17:38 | 2210 | en | 4,096 | 7,954 | c85c06d7-4f6c-4592-8495-c626abb695b0 |
}, \boldsymbol{a}_{t}\right)\right]
$$
where $\boldsymbol{h}_{t}=f_{\xi}\left(\operatorname{aug}\left(\boldsymbol{s}_{t}\right)\right)$ is the encoded augmented visual observation, $\boldsymbol{a}_{t}=\pi_{\phi}\left(\boldsymbol{h}_{t}\right)+\epsilon$ is the action with clipped noise to smooth the targets $\epsilon \s... | 2206.04779 | https://arxiv.org/abs/2206.04779 | 2022-06-13 02:04:14 | 2206 | en | 4,096 | 16,594 | c4594df3-cc3c-4e98-a22b-5508f27e1376 |
mathbf{X}_i \sim \mu_i$.
The above inequality can be written as
\begin{equation}
\left(\lambda A_1 + (1-\lambda)A_0\right)^q \leq \lambda A_1^q + (1-\lambda)A_0^q.
\end{equation}
Since $q(q-1)u^{q-2} \geq 0$ for $q> 1$ and $u>0$ it follows that the function $u\rightarrow u^q$ is convex on $\mathbb{R}^+$, so this ineq... | 2209.04722 | https://arxiv.org/abs/2209.04722 | 2022-09-13 02:10:25 | 2209 | en | 4,096 | 5,159 | e0129beb-6a29-4ec7-97cd-c670020eb861 |
\theta}(i,D)=
\begin{bmatrix}
\textbf{B}_{1}(1, D/2^{i-1}), &\textbf{0} &... &\textbf{0}\\
\textbf{0} &\textbf{B}_{2}(1, D/2^{i-1}) &... &\textbf{0}\\
&...\\
\textbf{0} &\textbf{0} &... &\textbf{B}_{2^{i-1}}(1, D/2^{i-1})
\end{bmatrix}
\end{equation*}
where $\textbf{B}_{j}(1, D/2^{i-1})$ is a le... | 2209.13774 | https://arxiv.org/abs/2209.13774 | 2022-09-29 02:06:04 | 2209 | en | 4,096 | 14,459 | a0aa664d-b7a5-4a1a-8456-39a8b5196c09 |
{S}^*$ in comparison with all the other $A_j$'s}, i.e.,
$$
\mathrm{H}_0: \lvert \mathcal{S}^* \cap A_1 \rvert = \max_j \lvert \mathcal{S}^* \cap A_j \rvert \text{~versus~}
\mathrm{H}_1: \lvert \mathcal{S}^* \cap A_1 \rvert < \max_j \lvert \mathcal{S}^* \cap A_j \rvert.
$$
The corresponding property set is $\mathcal... | 2301.12254 | https://arxiv.org/abs/2301.12254 | 2023-02-03 02:16:51 | 2301 | en | 4,096 | 9,197 | b4b1a16f-e4d5-4fd1-a29e-f0a849cec42f |
{ and } F^{T}H = 0,
\end{equation}
where $L \in \mathbb{R}^{n \times n}$ is the graph Laplacian,
$F \in \mathbb{R}^{n \times (h-1)}$ is from~\eqref{eq:fairmath},
and $F^TH=0$ is the original~\eqref{eq:fairmath} right-multiplied
with $\widehat D^{-1}$ due to the scaling~\eqref{eq:normal-H} of $H$.
The idea of enfor... | 2210.16435 | https://arxiv.org/abs/2210.16435 | 2022-11-01 01:04:08 | 2210 | en | 4,096 | 5,087 | d2f609c8-15e5-4692-a820-41b3b1d26a52 |
1}
There exists $h_{\max}$ such that
$\left\Vert \boldsymbol{\Sigma}_{{\cal T}_h}-\boldsymbol{\Sigma}_0\right\Vert_{\max}\le h_{\max}.$
\end{assumption}
Assumptions \ref{C10} and \ref{C11} are imposed to guarantee the consistency of transfer learning
estimator of the asymptotic variance \citep{tian2022transfer} if $h_... | 2211.14578 | https://arxiv.org/abs/2211.14578 | 2022-11-29 02:07:23 | 2211 | en | 4,096 | 14,967 | ff3b8f7e-9c4c-4211-b785-85ca1a8c57e0 |
htt{NTK}}) \\
&\leq 2Cn^{-\frac{2}{3}}\log^{2}\frac{6}{\delta}.
\end{aligned}
\end{equation*}
\section{Proof of Theorem \ref{thm:bad_gen}}
\begin{proof}[Proof of Theorem \ref{thm:bad_gen} $i)$]
Theorem \ref{thm:bad_gen} $i)$ is a direct corollary of the following lemmas and Proposition \ref{... | 2302.05933 | https://arxiv.org/abs/2302.05933 | 2023-02-14 02:16:36 | 2302 | en | 4,096 | 80,349 | 5ad7dc6c-8c72-46b0-8efb-58030a7307d3 |
a bound on generalization gap of order $O(\frac{1}{n} (f^{-1}(\frac{1}{T}))^{2/k})$ provided $m=\Omega((f^{-1}(\frac{1}{T}))^{2/k})$ (see Theorem \ref{thm:gengapapp} in Appendix). Hence, the tail behavior of $f$ controls both the generalization gap and minimum width requirement. In particular, polynomial losses with t... | 2302.09235 | https://arxiv.org/abs/2302.09235 | 2023-02-21 02:05:55 | 2302 | en | 4,096 | 7,612 | 0bad16f4-b093-4033-8058-e1e13f4494e9 |
d}}\bm{w}, \quad \bm{\theta}(t=0)\sim\rho_0,
\label{NoisySGD}
\end{align}
where $\bm{w}$ is the standard Wiener process in $\mathbb{R}^{p}$, and the random initial condition $\bm{\theta}(t=0)$ follows the law of a suitable PDF $\rho_0$ supported over $\mathbb{R}^{p}$.
In this regularized case, \eqref{NN_PDE} resu... | 2210.13879 | https://arxiv.org/abs/2210.13879 | 2022-10-26 02:13:58 | 2210 | en | 4,096 | 3,799 | 3ee3912d-001d-4812-b037-3138c0cff1ba |
this convex hull of randomized procedures and the convex geometry of the \textsc{taf}\xspace curve:
\begin{restatable}{theorem}{convexfauc} \label{thm:convexfauc}
For a fixed collection of models $\mathcal{H}$,
$\textsc{tafi}\xspace_{\mathcal{H}} = \overline{\conv}(\textsc{taf}\xspace_{\mathcal{H}}) = \textsc{taf}\x... | 2206.00074 | https://arxiv.org/abs/2206.00074 | 2022-06-02 02:02:32 | 2206 | en | 4,096 | 6,431 | c0c0dbbd-229a-4ab3-8a9e-607f024d6fb0 |
-random, time-invariant, measurable, and Lipschitz in $\mathbb{X}$.
We now consider a simplified version of Barnsley's fern. Let $\mathbb{X}=\mathbb{R}^2$, $x_0=0$, and $K=2$ and consider the affine transformations,
\begin{align}
G_1(x) \,=\, \begin{bmatrix} 0.40& -0.3733\\ 0.060& 0.60\end{bmatrix}x+{\begin{bmatrix} ... | 2211.08018 | https://arxiv.org/abs/2211.08018 | 2022-11-16 02:12:05 | 2211 | en | 4,096 | 13,468 | 5f820fe1-4d5d-4b05-89f2-cf05619a6ccb |
e-carlo-proof}
Suppose $\mathbb{E}_{\bar{\mathbb{P}}} h(\bm x, \rv \xi) < \infty$. The inner maximization of \eqref{eq:dro-method-dist-ball} under the Wasserstein distance, i.e.,
\begin{equation}\label{eq:wasserstein-DRO}
\displaystyle \max_{\P: W_p(\P, \bar{\mathbb{P}}) \le \epsilon} \int_{\Xi} h(\bm x, \rv \x... | 2301.13565 | https://arxiv.org/abs/2301.13565 | 2023-02-01 02:14:05 | 2301 | en | 4,096 | 11,593 | 897c64b1-0f87-4f2f-b101-11c32e48f01c |
sc SBoxCI}{} intervals. This simple experiment demonstrates that the {\sc SBox}{} and {\sc SBoxCI}{} methods are able to exploit the correlations among the dimensions much more effectively than the {\sc SBox}(Bonferroni) method.
\subsection{Studies of {\sc SQBox}{} and {\sc SQBoxCI}{}}
In the following experiments, w... | 2206.04860 | https://arxiv.org/abs/2206.04860 | 2022-06-23 02:04:36 | 2206 | en | 4,096 | 13,583 | a0610ac2-7e6f-4a91-8d8d-d9f30b939a20 |
x$ and $\xcf$ by the Euclidean distance, but more refined
distance measures have also been proposed
\cite{wachter2017counterfactual, karimi2020model,
poyiadzi2020face, joshi2019towards, arvanitidis2020geometrically}.
For simplicity, we
restrict attention to the Euclidean distance in our results, but we
expect that t... | 2205.15834 | https://arxiv.org/abs/2205.15834 | 2022-06-01 02:21:27 | 2205 | en | 4,096 | 1,936 | d85683bb-6736-4b4e-a7f8-5f5c76485332 |
_d$ and another dimension $d^{\prime}$ satisfying $a_{d^{\prime}} < b_{d^{\prime}}$. We say that vector $a$ is non-dominated by $b$ if there is such a dimension $d$ that $a_d > b_d$.
The Pareto order relationship can determine the Pareto optimality of arms as follows. Formally, in stochastic settings, an arm $j$ is s... | 2212.00884 | https://arxiv.org/abs/2212.00884 | 2022-12-05 02:02:28 | 2212 | en | 4,096 | 4,272 | 3a80d53b-e6b4-4d96-a345-1946b355fd8d |
among different treatment groups (for example, distribution invariance, domain adaptation, local similarity, domain overlap, mutual information, and so on)?
\end{itemize}
As shown in Figure \ref{fig:framework}, different from the previous surveys based on the taxonomy of the methodologies for treatment effect estima... | 2302.00848 | https://arxiv.org/abs/2302.00848 | 2023-02-03 02:06:03 | 2302 | en | 4,096 | 602 | b49e029b-0312-4796-a399-394f167946c0 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.