Affinity Posters
Tiny Papers Poster Session 3
Krystal Maughan · Thomas F Burns
Halle B
Live content is unavailable. Log in and register to view live content
Schedule
Wed 1:45 a.m. - 3:45 a.m.
|
How Many OptiFaces? A New Evaluation Metric For 3D Face Reconstruction
(
Poster
#316
)
>
link
Poster Location: Halle B #316 Three dimensional face reconstruction is a challenging problem, so much so that the mean face is highly competitive with recent learning-based approaches for 3D face reconstruction from 2D images. No other universal baselines for this task exist. We propose a novel baseline that selects a subset of face meshes, called OptiFaces, that minimise overall 3D reconstruction error. This is a universal approach to calculate dataset-specific metrics for 3D face reconstruction, offering intuitive new baselines for the interpretation of 3D reconstruction error. |
Will Rowan · Patrik Huber · Nick Pears · Andrew Keeling 🔗 |
Wed 1:45 a.m. - 3:45 a.m.
|
Weighted Branch Aggregation Based Deep Learning Model for Track Detection in Autonomous Racing
(
Poster
#317
)
>
link
Poster Location: Halle B #317 Intelligent track detection is a vital component of autonomous racing cars. We develop a novel Weighted Branch Aggregation based Convolutional Neural Network (WeBACNN) model that can accurately detect the track while being robust against image blurring due to high speed, and can work independently of lane markings. The code and dataset for this work is available at (anonymous). |
Shreya Ghosh · Yi-Huan Chen · Ching-Hsiang Huang · Abu Shafin Mohammad Mahdee Jameel · Aly El Gamal · Samuel Labi 🔗 |
Wed 1:45 a.m. - 3:45 a.m.
|
Nonlinear model reduction for operator learning
(
Poster
#318
)
>
link
Poster Location: Halle B #318 Operator learning provides methods to approximate mappings between infinite-dimensional function spaces. Deep operator networks (DeepONets) are a notable architecture in this field. Recently, an extension of DeepONet based on model reduction and neural networks, proper orthogonal decomposition (POD)-DeepONet, has been able to outperform other architectures in terms of accuracy for several benchmark tests. We extend this idea towards nonlinear model order reduction by proposing an efficient framework that combines neural networks with kernel principal component analysis (KPCA) for operator learning. We conduct experiments on three test cases, including the Navier–Stokes equation. Our results demonstrate the superior performance of KPCA-DeepONet over POD-DeepONet. |
Hamidreza Eivazi · Stefan Wittek · Andreas Rausch 🔗 |
Wed 1:45 a.m. - 3:45 a.m.
|
Parameter and Data Efficient Spectral Style-DCGAN
(
Poster
#319
)
>
link
Poster Location: Halle B #319 We present a simple, highly parameter, and data-efficient adversarial network for unconditional face generation. Our method: Spectral Style-DCGAN or SSD utilizes only 6.574 million parameters and 4739 dog faces from the Animal Faces HQ (AFHQ) dataset as training samples while preserving fidelity at low resolutions up to 64x64. Code available at Anonymous-repo. |
Aryan Garg 🔗 |
Wed 1:45 a.m. - 3:45 a.m.
|
VoltaVision: A Transfer Learning model for electronic component classification
(
Poster
#320
)
>
link
Poster Location: Halle B #320 In this paper, we analyze the effectiveness of transfer learning on classifying electronic components. Transfer learning reuses pre-trained models to save time and resources in building a robust classifier rather than learning from scratch. Our work introduces a lightweight CNN, coined as VoltaVision, and compares its performance against more complex models. We test the hypothesis that transferring knowledge from a similar task to our target domain yields better results than state-of-the-art models trained on general datasets. Our dataset and code for this work are available at https://anonymous.4open.science/r/VoltaVision-E4A5. |
Anas Mohammad Ishfaqul Muktadir Osmani · Taimur Rahman · Salekul Islam 🔗 |
Wed 1:45 a.m. - 3:45 a.m.
|
Affinity-based Homophily: Can we measure homophily of a graph without using node labels?
(
Poster
#321
)
>
link
Poster Location: Halle B #321 The homophily (heterophily) ratio in a graph represents the proportion of edges connecting nodes with similar (dissimilar) class labels. Existing methods for estimating the homophily ratio typically rely on knowing the class labels of each node in the graph. While several algorithms address both homophilic and heterophilic graphs, they necessitate prior knowledge of the homophily ratio to choose the appropriate one. To address this limitation, we propose a novel metric for measuring homophily ratio without information about node labels. In our approach, we define learnable affinity vectors for each node, characterizing the expected feature relationships with its neighbors. Our method, Affinity-based Homophily, derives the homophily ratio using these affinity vectors, eliminating the need for prior node label information. We conducted experiments on various benchmark homophilic and heterophilic graphs, demonstrating the commendable performance of our homophily measure. |
Indranil Ojha · Kushal Bose · Swagatam Das 🔗 |
Wed 1:45 a.m. - 3:45 a.m.
|
TOWARDS FAIRNESS CONSTRAINED RESTLESS MULTI-ARMED BANDITS: A CASE STUDY OF MATERNAL AND CHILD CARE DOMAIN
(
Poster
#322
)
>
link
Poster Location: Halle B #322 Restless multi-armed bandits (RMABs) are widely used for resource allocation in dynamic environments, but they typically do not consider fairness implications. This paper introduces a fairness-aware approach for offline RMABs. We propose a Kullback-Leibler (KL) divergence-based fairness metric to quantify the discrepancy between the selected and the overall population. This is incorporated as a regularizer into the soft whittle index optimization. We evaluate our fairness-aware algorithm on a real-world RMAB dataset where initial results suggest that our approach can potentially improve fairness while preserving solution quality. |
Gargi Singh · Milind Tambe · Aparna Taneja 🔗 |
Wed 1:45 a.m. - 3:45 a.m.
|
Evaluating Groups of Features via Consistency, Contiguity, and Stability
(
Poster
#323
)
>
link
Poster Location: Halle B #323 Feature attributions explain model predictions by assigning importance scores to input features. In high-dimensional data such as images, these scores are often assigned to groups of features at a time. There are a variety of strategies for creating these groups, ranging from simple patches to deep-learning-based segmentation algorithms. What makes certain groups better than others for explanations? We formally define three key criteria for interpretable groups of features: consistency, contiguity, and stability. Surprisingly, we find that patch-based groups outperform groups created via modern segmentation tools. |
Chaehyeon Kim · Weiqiu You · Shreya Havaldar · Eric Wong 🔗 |
Wed 1:45 a.m. - 3:45 a.m.
|
The IMO Small Challenge: First IMO Dataset for LLMs
(
Poster
#324
)
>
link
Poster Location: Halle B #324 We introduce the IMO Small Challenge: A curated collection of the easiest possible IMO problems and other competitive mathematical problems. The goal is to bridge the existing gap in the range of available dataset difficulties in terms of testing problem-solving skills: Currently, datasets are predominantly either too easy (MATH or GSM8K), excessively challenging (solving arbitrary IMO problems, such as required by the IMO Grand Challenge, and embodied by the miniF2F dataset) or focus too little on problem-solving (\emph{GHOSTS}). Our challenge interpolates this difficulty range and serves as a test bench for next-generation language models. We release a preliminary version of a dataset that accompanies this challenge. It is grounded in natural language, and problems are annotated with solutions and other metadata, such as the type of proof strategy used, in order to facilitate semi-automatic evaluation of LLMs' outputs beyond classical correct-incorrect keyword matching. |
Simon Frieder · Mirek Olšák · Julius Berner · Thomas Lukasiewicz 🔗 |
Wed 1:45 a.m. - 3:45 a.m.
|
Can Graph Neural Networks learn node-level structural features?
(
Poster
#325
)
>
link
Poster Location: Halle B #325 Graph Neural Networks (GNNs) have become one of the most widely adopted solutions for graph machine learning (GML) tasks. They perform feature learning on graphs using message passing on the network structure, avoiding the feature engineering step required for traditional tabular approaches for GML tasks. However, it is unclear which structural features GNNs can or cannot easily learn from data, especially for node- and edge-level properties. In this work, we propose a methodology to investigate which structural features GNNs can reconstruct from graph data. We conducted a first experimental analysis on one of the most used benchmarks for GML, considering some of the most well-known node-level features, such as centrality and transitivity measures. The results show that GNNs can easily reconstruct PageRank and in/out-degree centralities. But, surprisingly, GNNs can also learn centrality measures based on shortest path distances. Moreover, they reach quite good performance in learning the local clustering coefficient. |
Manuel Dileo · Matteo Zignani 🔗 |
Wed 1:45 a.m. - 3:45 a.m.
|
Chemical Language Models Have Problems with Chemistry: A Case Study on Molecule Captioning Task
(
poster
)
>
link
Poster Location: #326 Drug discovery has been greatly enhanced through the recent fusion of molecular sciences and natural language processing, leading these research fields to significant advancements. Considering the crucial role of molecule representation in chemical understanding within these models, we introduce novel probing tests designed to evaluate chemical knowledge of molecular structure in state-of-the-art language models (LMs), specifically MolT5 and Chem+Text T5. These probing tests are conducted on a molecule captioning task to gather evidence and insights into the language models' comprehension of chemical information. By applying rules to transform molecular SMILES into equivalent variants, we have observed significant differences in the natural language descriptions generated by the LM for a given molecule depending on the exact transformation used. |
Kuzma Khrabrov 🔗 |
Wed 1:45 a.m. - 3:45 a.m.
|
BACKTRACKING MATHEMATICAL REASONING OF LANGUAGE MODELS TO THE PRETRAINING DATA
(
poster
)
>
link
Poster Location: #327 In this study, we identify subsets of model pretraining data that contribute to the math reasoning ability of language models and evaluate it on several mathematical tasks (e.g., addition, multiplication). We find that training on math-only data improves simple arithmetic but doesn't fully account for complex reasoning abilities, such as chain-of-thought reasoning. We also find that code data contributes to chain-of-thought reasoning while reducing arithmetic performance. |
Yasaman Razeghi 🔗 |