Sankeerth Durvasula
1 followers · 1 following · 1019 views
on the atlas · 36
- [2010.02502] Denoising Diffusion Implicit Models2 savers
- Boosting Dynamic Programming Performance Using NVIDIA Hopper GPU DPX Instructions | NVIDIA Technical Blog1 savers
- 2112.107691 savers
- Compact3D: Compressing Gaussian Splat Radiance Field Models with Vector Quantization2 savers
- DreamGaussian: Generative Gaussian Splatting for Efficient 3D Content Creation | PDF2 savers
- TOG Submission Forms1 savers
- Technical Papers - SIGGRAPH Asia 20241 savers
- lec05.pdf2 savers
- quantization2 - songhan lecture1 savers
- mls.pdf1 savers
- DistriFusion: Distributed Parallel Inference for High-Resolution Diffusion Models | PDF1 savers
- Laplacian Mesh Editing1 savers
- gaussianflow1 savers
- Omnimotion1 savers
- Fast Training of Diffusion Models with Masked Transformers | PDF1 savers
- Elucidating the Design Space of Diffusion-Based Generative Models | PDF1 savers
- DUSt3R: Geometric 3D Vision Made Easy1 savers
- VastGaussian: Vast 3D Gaussians for Large Scene Reconstruction | PDF1 savers
- mpm course1 savers
- How NeRFs and 3D Gaussian Splatting are Reshaping SLAM: a Survey1 savers
- GaussianHair: Hair Modeling and Rendering with Light-aware Gaussians | PDF1 savers
- A Software-defined Tensor Streaming Multiprocessor for Large-scale Machine Learning1 savers
- Think Fast: A Tensor Streaming Processor (TSP) for Accelerating Deep Learning Workloads1 savers
- Scalable Diffusion Models with Transformers1 savers
- PC-NeRF: Parent-Child Neural Radiance Fields Using Sparse LiDAR Frames in Autonomous Driving Environments | PDF1 savers
- Xcube : Large-Scale 3D Generative Modeling using Sparse Voxel Hierarchies1 savers
- D-NeRF: Neural Radiance Fields for Dynamic Scenes1 savers
- Im4D: High-Fidelity and Real-Time Novel View Synthesis for Dynamic Scenes | PDF1 savers
- HyperReel: High-Fidelity 6-DoF Video with Ray-Conditioned Sampling | PDF1 savers
- RAFT: Recurrent All-Pairs Field Transforms for Optical Flow | PDF1 savers
- Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction | PDF1 savers
- ParticleNeRF: A Particle-Based Encoding for Online Neural Radiance Fields1 savers
- Index1 savers
- Curius / Onboarding2704 savers
- matrixcookbook.pdf9 savers
- [2111.00396] Efficiently Modeling Long Sequences with Structured State Spaces2 savers
highlights · 17
flexible can be molded to fit structure in arbitrary data. For example, we can define models in terms of any (non-negative) function φ ( x ) yielding th
Deep Unsupervised Learning using Nonequilibrium Thermodynamicstractable can be analytically evaluated and easily fit to data (e.g. a Gaussian or Laplac
Deep Unsupervised Learning using Nonequilibrium Thermodynamicsflexible distribution p ( x ) = φ ( x )
Deep Unsupervised Learning using Nonequilibrium ThermodynamicsA variety of analytic approximations exist which amelio- rate, but do not remove, this tradeoff–for instance mean field theory and its expansions (T, 1982; Tanaka, 1998), variational Bayes (Jordan et al., 1999), contrastive diver- gence (Welling & Hinton, 2002; Hinton, 2002), minimum probability flow (Sohl-Dickstein et al., 2011b;a), minimum KL contraction (Lyu, 2011), proper scoring rules (Gneit- ing & Raftery, 2007; Parry et al., 2012), score matching (Hyv ̈ arinen, 2005), pseudolikelihood (Besag, 1975), loopy belief propagation (Murphy et al., 1999), and many, many more. Non-parametric meth…
Deep Unsupervised Learning using Nonequilibrium ThermodynamicsA Comprehensive Introduction to Differential Geometry by Spivak (supplement). For more hardcore differential
So You Want to Learn Physics... — Susan Fowler(joint) expected predictive information gain
Unifying Approaches in Active Learning and Active Sampling via Fisher Information and Information-Theoretic Quantities | PDFexpected information gain (EIG)
Unifying Approaches in Active Learning and Active Sampling via Fisher Information and Information-Theoretic Quantities | PDFTable 1: Taxonomy of Information Quantities for Data Subset Selection. In general, information quantities can be split into ones for active sampling or active learning, into non-transductive and transductive ones, and in the transductive case, into taking an expectation or the joint over (additional) evaluation samples. Here, we show the information quantities for individual acquisition. For batch acquisition, { Y acq i } , { y acq i } , { x acq i } can be substituted. Active Learning Active Sampling Non-Transductive EIG/BALD I[Ω; Y acq | x acq ] IG I[Ω; y acq | x acq ] Transductive (using D e…
Unifying Approaches in Active Learning and Active Sampling via Fisher Information and Information-Theoretic Quantities | PDFadditional data D : = { ( x i ,y i ) }
Unifying Approaches in Active Learning and Active Sampling via Fisher Information and Information-Theoretic Quantities | PDFp( { y i } ,ω |{ x i } ) = p( { y i }|{ x i } ,ω ) p( ω )
Unifying Approaches in Active Learning and Active Sampling via Fisher Information and Information-Theoretic Quantities | PDFactive sampling improves training efficiency by filtering the training set to focus on the samples that will be the most informative for the model.
Unifying Approaches in Active Learning and Active Sampling via Fisher Information and Information-Theoretic Quantities | PDFiven access to unlabeled data, active learning selects the most informative samples to label for a given model, thus decreasing the number of required annotations to reach a given level of performance.
Unifying Approaches in Active Learning and Active Sampling via Fisher Information and Information-Theoretic Quantities | PDFWhen the log-likelihood has the form of Eq. (4), in our case the rendering error, the difference of the entropies in the R.H.S. of Eq. (6) can be approximated as [14]
FisherRF: Active View Selection and Uncertainty Quantification for Radiance Fields using Fisher Information | PDFim is to select the next best view that maximizes the Information Gain
FisherRF: Active View Selection and Uncertainty Quantification for Radiance Fields using Fisher Information | PDF( w ) = − E y | x ∼ p w ∂ 2 log p ( y | x , w ) ∂ w 2 w = H ′′ [ y | x , w ] (5) where H ′′ [ y | x , w ] is the Hessian matrix of Eq. (4)
FisherRF: Active View Selection and Uncertainty Quantification for Radiance Fields using Fisher Information | PDFhe Fisher Information of the model log p ( y | x ; w ) is the Hessian of the log-likelihood function with respect to the model parameters
FisherRF: Active View Selection and Uncertainty Quantification for Radiance Fields using Fisher Information | PDF( y | x ; w ) . In the problem of novel view synthesis, ( x , y ) are the camera pose x and image observa- tion y at pose x , respectively, whereas w are the volumetric parameters of the radiance field
FisherRF: Active View Selection and Uncertainty Quantification for Radiance Fields using Fisher Information | PDF