summaryrefslogtreecommitdiff
path: root/ictp-saifr_colloquium.tex
diff options
context:
space:
mode:
authorJaron Kent-Dobias <jaron@kent-dobias.com>2025-02-16 16:12:58 -0300
committerJaron Kent-Dobias <jaron@kent-dobias.com>2025-02-16 16:12:58 -0300
commit38ed0e4f23c4967925ced0c306450b0462e6ceff (patch)
tree723bb4a0d1c7791a43e880f09bb6310a96b6409e /ictp-saifr_colloquium.tex
parente4c01a637e4cb782ff9c67d9a42c54bb62a135d6 (diff)
downloadictp-saifr_colloquium-38ed0e4f23c4967925ced0c306450b0462e6ceff.tar.gz
ictp-saifr_colloquium-38ed0e4f23c4967925ced0c306450b0462e6ceff.tar.bz2
ictp-saifr_colloquium-38ed0e4f23c4967925ced0c306450b0462e6ceff.zip
Some tweaks after first practice.
Diffstat (limited to 'ictp-saifr_colloquium.tex')
-rw-r--r--ictp-saifr_colloquium.tex18
1 files changed, 15 insertions, 3 deletions
diff --git a/ictp-saifr_colloquium.tex b/ictp-saifr_colloquium.tex
index d883ed8..1eca601 100644
--- a/ictp-saifr_colloquium.tex
+++ b/ictp-saifr_colloquium.tex
@@ -389,7 +389,7 @@
\[
0=y_i-\hat f(x_i\mid\pmb a)\qquad\text{for all $1\leq i\leq M$}
\]
- with $N$ unknowns $\pmb a=[a_1,\ldots, a_N]$ gives a manifold of $N-M$ dimensions
+ with $N$ unknowns $\pmb a=[a_1,\ldots, a_N]$ gives a manifold of $DlN-M$ dimensions
\medskip
@@ -848,8 +848,8 @@
\begin{tikzpicture}
\draw (0,0) node[align=center] {Overparameterization works\\\includegraphics[width=3cm]{figs/fit_overparamfit_abs2.pdf}};
\draw (4,2) node[align=center] {Gradient descent\\implicitly regularizes\\\includegraphics[height=2cm]{figs/fit_gradient_5.pdf}};
- \draw (-4,2) node[align=center] {Neural networks\\are good bases\\\includegraphics[height=2cm]{figs/fit_basis_abs.pdf}};
- \draw (-4,-2) node[align=center] {Data is sparse\\and high-dimensional\\\includegraphics[height=2cm]{figs/fit_data_abs2.pdf}};
+ \draw (-4,2) node[align=center] {Neural networks\\are good models\\\includegraphics[height=2cm]{figs/fit_basis_abs.pdf}};
+ \draw (-4,-2) node[align=center] {Data is sparse, low-noise,\\and high-dimensional\\\includegraphics[height=2cm]{figs/fit_data_abs2.pdf}};
\draw (4,-2) node[align=center] {SGD finds\\high-entropy solutions\\\includegraphics[height=2cm]{figs/gradient_vs_sgd_4.png}};
\end{tikzpicture}
\end{frame}
@@ -1301,6 +1301,11 @@
\framesubtitle{Conclusions and future directions}
\begin{columns}
\begin{column}{0.5\textwidth}
+
+ Many outstanding questions:
+
+ \bigskip
+
\textbf{How do these structures interact with dynamics?}
\bigskip
@@ -1312,6 +1317,13 @@
\textbf{How do topological properties of solutions correlate with their quality?}
\end{column}
\begin{column}{0.5\textwidth}
+ \begin{tikzpicture}[scale=0.6, every node/.style={scale=0.6}]
+ \draw (0,0) node[align=center] {Overparameterization works\\\includegraphics[width=3cm]{figs/fit_overparamfit_abs2.pdf}};
+ \draw (4,2) node[align=center] {Gradient descent\\implicitly regularizes\\\includegraphics[height=2cm]{figs/fit_gradient_5.pdf}};
+ \draw (-4,2) node[align=center] {Neural networks\\are good models\\\includegraphics[height=2cm]{figs/fit_basis_abs.pdf}};
+ \draw (-4,-2) node[align=center] {Data is sparse, low-noise,\\and high-dimensional\\\includegraphics[height=2cm]{figs/fit_data_abs2.pdf}};
+ \draw (4,-2) node[align=center] {SGD finds\\high-entropy solutions\\\includegraphics[height=2cm]{figs/gradient_vs_sgd_4.png}};
+ \end{tikzpicture}
\end{column}
\end{columns}
\end{frame}