Searcharxiv⌕ Search

arXiv subjects

Edward Lee

Publications and source records attributed to Edward Lee.

28 records · Page 2Linked to original sources
arXiv2019

Dataset Culling: Towards Efficient Training Of Distillation-Based Domain Specific Models

Real-time CNN-based object detection models for applications like surveillance can achieve high accuracy but are computationally expensive. Recent works have shown 10 to 100x reduction in computation cost for inference by using domain-specific networks. However, prior works have focused on inference only. If the domain model requires frequent retraining, training costs can pose a significant bottleneck. To address this, we propose Dataset Culling: a pipeline to reduce the size of the dataset for training, based on the prediction difficulty. Images that are easy to classify are filtered out since they contribute little to improving the accuracy. The difficulty is measured using our proposed confidence loss metric with little computational overhead. Dataset Culling is extended to optimize the image resolution to further improve training and inference costs. We develop fixed-angle, long-duration video datasets across several domains, and we show that the dataset size can be culled by a factor of 300x to reduce the total training time by 47x with no accuracy loss or even with slight improvement. Codes are available: https://github.com/kentaroy47/DatasetCulling

cs.CV↗
arXiv2018

Deep Learning with Attention to Predict Gestational Age of the Fetal Brain

Fetal brain imaging is a cornerstone of prenatal screening and early diagnosis of congenital anomalies. Knowledge of fetal gestational age is the key to the accurate assessment of brain development. This study develops an attention-based deep learning model to predict gestational age of the fetal brain. The proposed model is an end-to-end framework that combines key insights from multi-view MRI including axial, coronal, and sagittal views. The model also uses age-activated weakly-supervised attention maps to enable rotation-invariant localization of the fetal brain among background noise. We evaluate our methods on the collected fetal brain MRI cohort with a large age distribution from 125 to 273 days. Our extensive experiments show age prediction performance with R2 = 0.94 using multi-view MRI and attention.

cs.CV↗
arXiv2018

Training Domain Specific Models for Energy-Efficient Object Detection

We propose an end-to-end framework for training domain specific models (DSMs) to obtain both high accuracy and computational efficiency for object detection tasks. DSMs are trained with distillation \cite{hinton2015distilling} and focus on achieving high accuracy at a limited domain (e.g. fixed view of an intersection). We argue that DSMs can capture essential features well even with a small model size, enabling higher accuracy and efficiency than traditional techniques. In addition, we improve the training efficiency by reducing the dataset size by culling easy to classify images from the training set. For the limited domain, we observed that compact DSMs significantly surpass the accuracy of COCO trained models of the same size. By training on a compact dataset, we show that with an accuracy drop of only 3.6\%, the training time can be reduced by 93\%. The codes are uploaded in https://github.com/kentaroy47/training-domain-specific-models.

cs.CV↗
arXiv2018

The Temporary Exchange Problem

We formalize an allocation model under ordinal preferences that is more general than the well-studied Shapley-Scarf housing market. In our model, the agents do not just care which house or resource they get but also care about who gets their own resource. This assumption is especially important when considering temporary exchanges in which each resource is eventually returned to the owner. We show that several positive axiomatic and computational results that hold for housing markets do not extend to the more general setting. We then identify natural restrictions on the preferences of agents for which several positive results do hold. One of our central results is a general class of algorithms that return any allocation that is individually rational and Pareto optimal with respect to the responsive set extension.

cs.GT↗
arXiv2017

Exact Algorithms via Multivariate Subroutines

We consider the family of $Φ$-Subset problems, where the input consists of an instance $I$ of size $N$ over a universe $U_I$ of size $n$ and the task is to check whether the universe contains a subset with property $Φ$ (e.g., $Φ$ could be the property of being a feedback vertex set for the input graph of size at most $k$). Our main tool is a simple randomized algorithm which solves $Φ$-Subset in time $(1+b-\frac{1}{c})^n N^{O(1)}$, provided that there is an algorithm for the $Φ$-Extension problem with running time $b^{n-|X|} c^k N^{O(1)}$. Here, the input for $Φ$-Extension is an instance $I$ of size $N$ over a universe $U_I$ of size $n$, a subset $X\subseteq U_I$, and an integer $k$, and the task is to check whether there is a set $Y$ with $X\subseteq Y \subseteq U_I$ and $|Y\setminus X|\le k$ with property $Φ$. We derandomize this algorithm at the cost of increasing the running time by a subexponential factor in $n$, and we adapt it to the enumeration setting where we need to enumerate all subsets of the universe with property $Φ$. This generalizes the results of Fomin et al. [STOC 2016] who proved the case where $b=1$. As case studies, we use these results to design faster deterministic algorithms for: - checking whether a graph has a feedback vertex set of size at most $k$ - enumerating all minimal feedback vertex sets - enumerating all minimal vertex covers of size at most $k$, and - enumerating all minimal 3-hitting sets. We obtain these results by deriving new $b^{n-|X|} c^k N^{O(1)}$-time algorithms for the corresponding $Φ$-Extension problems (or enumeration variant). In some cases, this is done by adapting the analysis of an existing algorithm, or in other cases by designing a new algorithm. Our analyses are based on Measure and Conquer, but the value to minimize, $1+b-\frac{1}{c}$, is unconventional and requires non-convex optimization.

cs.DS↗
arXiv2016

Faster Graph Coloring in Polynomial Space

We present a polynomial-space algorithm that computes the number independent sets of any input graph in time $O(1.1387^n)$ for graphs with maximum degree 3 and in time $O(1.2355^n)$ for general graphs, where n is the number of vertices. Together with the inclusion-exclusion approach of Björklund, Husfeldt, and Koivisto [SIAM J. Comput. 2009], this leads to a faster polynomial-space algorithm for the graph coloring problem with running time $O(2.2355^n)$. As a byproduct, we also obtain an exponential-space $O(1.2330^n)$ time algorithm for counting independent sets. Our main algorithm counts independent sets in graphs with maximum degree 3 and no vertex with three neighbors of degree 3. This polynomial-space algorithm is analyzed using the recently introduced Separate, Measure and Conquer approach [Gaspers & Sorkin, ICALP 2015]. Using Wahlström's compound measure approach, this improvement in running time for small degree graphs is then bootstrapped to larger degrees, giving the improvement for general graphs. Combining both approaches leads to some inflexibility in choosing vertices to branch on for the small-degree cases, which we counter by structural graph properties.

cs.DS↗
arXiv2014

Capturing collective conflict dynamics with sparse social circuits

We discuss a set of computational techniques, called Inductive Game Theory, for extracting strategic decision-making rules from time series data and constructing probabilistic social circuits. We construct these circuits by connecting component individuals and groups with strategies in a game and propose an inductive approach to reconstructing the edges. We demonstrate this approach with conflict behavior in a society of pigtailed macaques by identifying significant patterns in decision-making by individuals. With the constructed circuit, we then capture macroscopic features of the system that were not specified in the construction of the initial circuit, providing a mapping between individual level behaviors to collective behaviors over the scale of the group. We extend on previous work in Inductive Game Theory by more efficiently searching the space of possible strategies by grouping individuals into socially relevant sets to produce a more efficient, parsimonious specification of the underlying interactions between components. We discuss how we reduce the dimensionality of these circuits using coarse-graining or compression to build cognitive effective theories for collective behavior.

cs.SI↗
arXiv2009

A modular quintic Calabi-Yau threefold of level 55

In this note we search the parameter space of Horrocks-Mumford quintic threefolds and locate a Calabi-Yau threefold which is modular, in the sense that the L-function of its middle-dimensional cohomology is associated to a classical modular form of weight 4 and level 55.

math.AG↗
arXiv2005

A Modular Non-Rigid Calabi-Yau Threefold

We construct an algebraic variety by resolving singularities of a quintic Calabi-Yau threefold. The middle cohomology of the threefold is shown to contain a piece coming from a pair of elliptic surfaces. The resulting quotient is a two-dimensional Galois representation. By using the Lefschetz fixed-point theorem in étale cohomology and counting points on the variety over finite fields, this Galois representation is shown to be modular.

math.AG↗