Combining Semantic Guidance and Deep Reinforcement Learning For Generating Human Level Paintings

Recent research efforts have been focused on teaching machines “howto paint”, in a manner similar to a human painter . Previous methods have been limited to datasets with little variation in position, scale and saliency of the foreground object . Semantic Guidancepipeline proposes bi-level painting procedure for learning the distinction between foreground and background brush strokes at training time .…

DeRF Decomposed Radiance Fields

Spatially decompose a scene and dedicate smaller networks for each decomposed part . This allows us near-constantinference time regardless of the number of decomposed parts . Voronoi spatial decomposition is preferable for this purpose, as it is compatible with the Painter’s Algorithm for efficient and GPU-friendlyrendering .…

World Model as a Graph Learning Latent Landmarks for Planning

Planning – the ability to analyze the structure of a problem in the large – is a hallmark of humanintelligence . Deep reinforcement learning (RL) has shown great promise for solving relatively straightforward control tasks, but it remains an openproblem how to best incorporate planning into existing deep RL paradigms to handle increasingly complex environments .…

Quantum algorithms for matrix scaling and matrix balancing

Matrix scaling and matrix balancing are two basic linear-algebraic problems . We study the power and limitations of quantum algorithms for these problems . Our quantum algorithms run in time for scaling or balancing an $n \times n$ matrix (given by an oracle) with $m$non-zero entries to within $ell_1$-error $varepsilon$ We achieve a polynomial speed-up in terms of $n, at the expense of a worse polynompositional dependence on the obtained $ell-1$ error .…

Unsupervised Object Detection with LiDAR Clues

In this paper, we present the first practical method for unsupervised objectdetection with the aid of LiDAR clues . By exploiting the 3D scenestructure, the issue of localization can be mitigated . We further identify another major issue, seldom noticed by the community, that the long-tailed and open-ended (sub-)category distribution should be accommodated .…

Deep Physics aware Inference of Cloth Deformation for Monocular Human Performance Capture

existing methods either do not estimate clothing at all or model clothdeformations with simple geometric priors . This leads to noticeable artifacts in their constructions, such as baked-in wrinkles and implausible deformations that defy gravity . We propose a person-specific, learning-based method thatintegrates a finite element-based simulation layer into the training process .…

Causal inference using deep neural networks

Causal inference from observation data is a core problem in many scientific fields . We present a general supervised deep learning framework that transforms input vectors to an image-like representation for every pair of inputs . We train a convolutional neural network on a normalized empirical probability density distribution (NEPDF)matrix .…

Bounds for Algorithmic Mutual Information and a Unifilar Order Estimator

Inspired by Hilberg’s hypothesis, which states that mutual information between blocks for natural language grows like a power law, we seek for links between power-law growth rate of algorithmic mutual information and someestimator of the unifilar order . This order estimator is intractable and follows the ideas byMerhav, Gutman, and Ziv (1989) and by Ziv and Merhav (1992) We show that all sources of a finite order exhibitsub-power-law .…

Evaluation of quality measures for color quantization

Visual quality evaluation is one of the challenging basic problems in imageprocessing . It also plays a central role in the shaping, implementation,optimization, and testing of many methods . The existing image qualityassessment methods focused on images corrupted by common degradation types while little attention was paid to color quantization .…

Learning to Expand Reinforced Pseudo relevance Feedback Selection for Information seeking Conversations

Thepseudo-relevance feedback (PRF) has demonstrated its effectiveness in incorporating relevance signals from external documents . We proposed a reinforcedselector to extract useful PRF terms to enhance response candidates and a BERTbased response ranker to rank the PRF-enhanced responses . We have also deployed our method on online production in ane-commerce company, which shows a significant improvement over the existing online ranking system .…

Modeling the Evolution of Retina Neural Network

Retinal circuitry shows many similar structures across a very broad array of species, both vertebrate andnon-vertebrate . This surprisingly conservative pattern raises a question of how evolution leadsto it, and whether there is any alternative that can also prompt helpfulpreprocessing .…

The Unreasonable Effectiveness of Encoder Decoder Networks for Retinal Vessel Segmentation

We propose an encoder-decoder framework for segmentation of blood vessels in retinal images that relies on extraction of large-scale patches at multiple image-scales during training . We show that this framework – called VLight – avoids overfitting to specific training images and generalizes well across different datasets, which makes it highly suitable for real-world applications where robustness, accuracy and low inferencetime on high-resolution fundus images is required .…

Genome assembly a universal theoretical framework unifying and generalizing the safe and complete algorithms

Genome assembly is a fundamental problem in Bioinformatics, requiring toreconstructing a source genome from an assembly graph built from a set of reads . The goal is to find what is definitely present in all solutions, orwhat is safe . The long-standing open problem offinding all the safe parts of the solutions was recently solved by a majortheoretical result [RECOMB’16].…

Equivariant Conditional Neural Processes

Equivariant Conditional Neural Processes (EquivCNPs) is a newmember of the Neural Process family that models vector-valued data in anequivariant manner with respect to isometries of $mathbb{R}^n$ We find Equiv CNPs are more robust against overfitting to local conditions of the training data .…

Recursive Projection Aggregation Decoding of Reed Muller Codes

Reed-Muller (RM) codes are one of the oldest families of codes . The proposed approach consists of multiple sparse RPAs that are generated by performing only a selection of projections in each sparsified decoder . Simulation results show that our proposed approach reducesthe RPA decoder’s computations up to $80\%$ with negligible performance loss.…

Recalibration of Neural Networks for Point Cloud Analysis

Spatial and channel re-calibration have become powerful concepts in computervision . Their ability to capture long-range dependencies is especially useful for those networks that extract local features, such as CNNs . We propose a set ofre-Calibration blocks that extend Squeeze and Excitation blocks and that can beadded to any network for 3D point cloud analysis that builds a globaldescriptor by hierarchically combining features from multiple localneighborhoods .…

SurFree a fast surrogate free black box attack

SurFreeproposes to bypass these attacks by focusing on careful trials along diversedirections, guided by precise indications of geometrical properties of the classifier decision boundaries . We exhibit a faster distortion decay underlow query amounts (few hundreds to a thousand) while remaining competitive athigher query budgets .…

Stay Connected Leave no Trace Enhancing Security and Privacy in WiFi via Obfuscating Radiometric Fingerprints

Recent works propose practical fingerprinting solutions that can be readily implemented in commercial-off-the-shelf devices . RF-Veil is a radiometric fingerprinting solution that protects user privacy by obfuscating the fingerprint of the transmitter for non-legitimate recipients . The proposed solution allows communicating with other devices, which don’t use RFVeil, which do not use the fingerprinting method, to communicate with each other .…

TLeague A Framework for Competitive Self Play based Distributed Multi Agent Reinforcement Learning

Competitive Self-Play (CSP) based Multi-Agent Reinforcement Learning (MARL) has shown phenomenal breakthroughs recently . Strong AIs are achieved for several benchmarks, including Dota 2, Glory of Kings, Quake III, StarCraft II, to name a few . Despite the success, the MARL training is extremely datathirsty, requiring typically billions of (if not trillions of) frames be seen from the environment during training in order for learning a high performanceagent .…

All You Need is a Good Functional Prior for Bayesian Deep Learning

The Bayesian treatment of neural networks dictates that a prior distributionis specified over their weight and bias parameters . This poses a challenge because modern neural networks are characterized by a large number of parameters . The choice of these priors has an uncontrolled effect on the induced functional prior, which is the distribution of the functions obtained by sampling the parameters from their prior distribution .…

A combination of Residual Distribution and the Active Flux formulations or a new class of schemes that can combine several writings of the same hyperbolic problem application to the 1D Euler equations

We show how to combine conservative and non conservative formulations of an hyperbolic system thathas a conservative form . This is inspired from two different class of schemes: the Residual Distribution one, and the Active Flux one . This new classof scheme is proved to satisfy a Lax-Wendroff like theorem .…

Rapid Exploration of Optimization Strategies on Advanced Architectures using TestSNAP and LAMMPS

The Spectral Neighbor Analysis Potential (SNAP) is amachine-learned inter-atomic potential utilized in cutting-edge moleculardynamics simulations . Previous implementations of the SNAP calculation showed adownward trend in their performance relative to peak on newer-generation CPUs and low performance on GPUs . We find a $22x time-to-solution improvement relative to an existing implementation as measured on an NVIDIA Tesla V100-16GBfor an important benchmark .…

Interpreting U Nets via Task Driven Multiscale Dictionary Learning

U-Nets have been tremendously successful in many imaging inverse problems . We show that one can reduce a U-Net to a tractable, well-understood sparsity-driven dictionary model . This model can be trained in a task-drivendictionary learning framework and yield comparable results to standard U-Ns on a number of relevant tasks, including CT and MRI reconstruction .…

Wedge Lifted Codes

We define wedge-lifted codes, a variant of lifted codes, and study theirlocality properties . They give improved trade-offs between redundancy and locality among binary codes . We show that (taking the trace of) wedge-lift codesyields binary codes with the $t$-disjoint repair property ($t$)…

Neural Representations for Modeling Variation in English Speech

Transformer-based speech representations lead to significant performance gains over the use of phonetic transcriptions . We use these representations to computeword-based pronunciation differences between non-native and native speakers of English . We alsodemonstrate that these neural speech representations capture segmentaldifferences, but also intonational and durational differences that cannot be represented by a set of discrete symbols used in phonetic transcripts .…