arXiv Science⌕ Search

arXiv subjects

Jack Ruder

Publications and source records attributed to Jack Ruder.

2 recordsLinked to original sources

An Omnidirectional WECT: Closed-Form Computation of the Directionally Integrated Weighted Euler Characteristic Transform

The weighted Euler characteristic transform (WECT) and related directional descriptors have become popular, efficiently computable tools for shape analysis across a variety of domains. In practice, however, they are approximated over a finite set of filtration directions, and thus the vectorized descriptor depends on this choice of directions and the rotational orientation of the data. In this manuscript, we formally introduce the directionally integrated weighted Euler characteristic transform (DIWECT), which averages the WECT over all directions $s \in \mathbb{S}^n$ for a weighted simplicial complex immersed in $\mathbb{R}^{n+1}$. A framework for computing the DIWECT exactly, without sampling directions, is outlined: for each simplex, we partition $\mathbb{S}^n$ into the sets of directions in which a given vertex enters the filtration last, and then the DIWECT at height $t$ reduces to measuring intersections of these regions with a set of directions in which the simplex has entered the filtration by height~$t$. We provide implementations for complexes immersed in $\mathbb{R}^2$ and $\mathbb{R}^3$, in which case the problem reduces to computing arc lengths and spherical areas. Experiments on synthetic and real-world data show that the DIWECT, while resulting in more compact vectorizations, also successfully encodes shape information and discards noise from data orientation. On a barley seed classification task, it outperforms the WECT in every pose except a hand-curated frame, whose advantage disappears under a two-degree rotation error. Compared with averaging the WECT over sampled directions, the DIWECT is often faster at matched accuracy. In one instance, this is upwards of a $50 \times$ speedup on a one-million-cell image with an error of $10^{-3}$.

cs.CG↗

Diffusion-Driven State Space Models

In many domains, practitioners seek models that produce accurate forecasts while faithfully capturing latent system dynamics. Existing approaches typically sacrifice one of these goals: deep state space models often assume Gaussian latent transitions, limiting fit and forecasting, while diffusion models are highly expressive but lack principled inference for the underlying dynamics. To combine the strengths of both, we introduce the Diffusion-Driven State Space Model (DDSSM), which replaces the conventional Gaussian transition distribution with a diffusion model. Our DDSSM resolves the open problem of how to jointly train an autoencoder and a diffusion model on sequential data, thereby extending the literature on latent diffusion models for time series. Moreover, we find that the DDSSM empirically outperforms a state-of-the-art deep SSM at fitting and forecasting a simulated time series with multimodal transitions.

stat.ML↗