tikzposter

\documentclass[25pt,a0paper,portrait,margin=0mm,innermargin=15mm,blockverticalspace=15mm,colspace=15mm,subcolspace=8mm]{tikzposter}

\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{amsmath,amssymb}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{hyperref}

\title{\parbox{\linewidth}{\centering Contrastive Pretraining for Multivariate Time Series: Scaling, Robustness, and Transfer}}
\author{First A. Author $^{1}$ \quad Second B. Author $^{1}$ \quad Third C. Author $^{2}$}
\institute{$^{1}$University of Example \quad $^{2}$Example Research Labs}

\usetheme{Default}
\usecolorstyle{Russia}
\usetitlestyle{Default}
\useblockstyle{Default}

\begin{document}
\maketitle

\block{Introduction}{
Multivariate time series arise across finance, health, energy, and industrial
monitoring. Hand-crafted features and supervised training from scratch have
dominated the literature, but recent self-supervised methods suggest that
pretraining on large unlabeled corpora can transfer broadly. We investigate
contrastive objectives tailored to the temporal structure of these signals.
\textbf{Our three contributions:} (i) a temporal view-generation strategy that
outperforms naive augmentation; (ii) scaling laws for pretraining data and
model size; (iii) systematic transfer experiments across eight downstream tasks.
}

\begin{columns}
\column{0.5}

\block{Method}{
Given a window $x \in \mathbb{R}^{T \times d}$, we generate two views $x^{(1)}, x^{(2)}$ via
random cropping, jitter, and frequency-masking. An encoder $f_\theta$ maps each view to
a latent $z^{(i)}$. We train with InfoNCE:
\begin{equation*}
\mathcal{L}_\text{NCE} = - \log
\frac{\exp(\text{sim}(z^{(1)}, z^{(2)})/\tau)}{\sum_{k}\exp(\text{sim}(z^{(1)}, z^{(k)})/\tau)}.
\end{equation*}
The encoder is a stack of dilated temporal convolutions followed by a light
Transformer. Key design: shared parameters across channels, plus a learned
channel-embedding that breaks permutation symmetry when needed.
}

\block{Scaling Laws}{
We pretrain at five data scales (1M to 1B windows) and four model sizes (5M to 400M
parameters). Validation loss follows a power law in both axes with
exponents $-0.13$ (data) and $-0.09$ (parameters). Compute-optimal scaling
trades data and parameters at roughly 2:1 tokens-per-parameter.
}

\column{0.5}

\block{Downstream Results}{
\centering
\begin{tabular}{lcccc}
\toprule
Task & Scratch & Supervised & Prior SSL & \textbf{Ours} \\
\midrule
Anomaly (F1)     & 0.71 & 0.78 & 0.82 & \textbf{0.89} \\
Forecast (CRPS)  & 0.92 & 0.83 & 0.79 & \textbf{0.71} \\
Classification (Acc) & 0.64 & 0.73 & 0.76 & \textbf{0.81} \\
Segmentation (IoU)   & 0.52 & 0.61 & 0.65 & \textbf{0.72} \\
\bottomrule
\end{tabular}
\vspace{0.5em}
Averaged over eight public benchmarks. Full table in the paper.
}

\block{Robustness}{
We perturb the test inputs with Gaussian noise, missing windows, and
sampling-rate changes. Our pretrained model degrades gracefully
(--4.2 F1 at 20\% missing) compared to supervised baselines (--14.8 F1),
confirming that the learned representations capture redundant, useful structure.
}

\end{columns}

\block{Ablation and Analysis}{
Removing the frequency-masking augmentation costs 3.7 F1 points on the anomaly
benchmark. Removing the channel embedding costs 1.8 points but hurts more on
heterogeneous datasets. Surprisingly, using only cropping (without jitter) works
well on smooth signals but collapses on noisy medical data, suggesting augmentations
should be tuned to the modality. Finally, pretraining crosses datasets are a free
lunch: joint pretraining on finance and health corpora never hurts either domain.
}

\block{Conclusion}{
Contrastive pretraining on large unlabeled time-series corpora transfers broadly and
robustly. Scaling laws suggest that data, not parameters, is currently the bottleneck.
We release the pretrained checkpoints, the augmentation library, and the benchmark
harness at \texttt{example.com/tsc-pretrain}.
}

\block{References}{
\footnotesize
[1] A. Author, B. Author. Self-Supervised Time Series. \emph{ICML}, 2024.\\
[2] C. Author. Contrastive Representations. \emph{NeurIPS}, 2023.\\
[3] D. Author, E. Author. Scaling Laws for Representations. \emph{TMLR}, 2024.
}

\end{document}