Knowledge Vault 2/79 - ICLR 2014-2023
Moderators: Adriana Romero-Soriano Yale Song ICLR 2021 - Outstanding Paper Session 1
<Resume Image >

Concept Graph & Resume using Claude 3 Opus | Chat GPT4 | Gemini Adv | Llama 3:

graph LR classDef alex fill:#f9d4d4, font-weight:bold, font-size:14px; classDef eigengame fill:#d4f9d4, font-weight:bold, font-size:14px; classDef sde fill:#d4d4f9, font-weight:bold, font-size:14px; classDef meshgraphnets fill:#f9f9d4, font-weight:bold, font-size:14px; A[ICLR 2021
Outstanding Paper Session 1] --> B[Alex: neurosynthesis of binaural
speech for VR. 1] B --> C[Captured data around mannequin
to train model. 2] C --> D[Model warps mono signal
to synthesize binaural. 3] D --> E[Key components: warping, WaveNet,
modified L2 loss. 4] B --> F[Outperforms baselines, enables realistic
VR audio. 5] A --> G[Ian: Eigengame reformulates PCA
as multi-player game. 6] G --> H[Top-k PCA is Nash
equilibrium of game. 7] G --> I[Distributed updates via matrix-vector
products, scalable. 8] I --> J[Computes top components quickly
on huge dataset. 9] G --> K[Novel applications, outperforms baselines. 10] A --> L[Yang: generative model reverses
SDE perturbing data. 11] L --> M[Neural net estimates score
function at noise levels. 12] L --> N[Samples generated by solving
reverse-time SDE. 13] L --> O[Exact likelihoods computed by
transforming SDE to ODE. 14] L --> P[State-of-the-art sample quality, enables
conditional generation. 15] A --> Q[Tobias: MeshGraphNets for mesh-based
simulations using GNNs. 16] Q --> R[Encodes physical system as
graph of mesh elements. 17] R --> S[GNN predicts mesh evolution,
generalizes to complex scenarios. 18] Q --> T[Models cloth, fluids, elasticity
by changing mesh quantities. 19] Q --> U[Adaptive remeshing allocates resolution,
generalizes to larger meshes. 20] class A,B,C,D,E,F alex; class G,H,I,J,K eigengame; class L,M,N,O,P sde; class Q,R,S,T,U meshgraphnets;

Resume:

1.-Alex from Facebook Reality Labs presents research on neurosynthesis of binaural speech from mono audio for realistic sound in virtual reality.

2.-They captured data of speech around a mannequin to train an end-to-end model with a neural time warping component.

3.-The model warps the mono signal over time based on transmitter/receiver positions to synthesize binaural audio matching the ground truth.

4.-Key components are neural time warping, a WaveNet-style network, and a modified L2 loss balancing optimization of amplitude and phase.

5.-The approach outperforms digital signal processing baselines and enables realistic, immersive audio rendering for VR applications.

6.-Ian Gemp presents Eigengame, reformulating PCA as the solution to a multi-player game to make it more naturally distributed and decentralized.

7.-They prove the top-k PCA solution is a Nash equilibrium of the game they construct with an interpretable utility function for each player.

8.-Eigengame's distributed updates rely only on matrix-vector products, allowing it to scale to massive datasets on TPU hardware.

9.-On a huge dataset from a pre-trained ResNet-200, Eigengame computes the top 32 principal components in 9 hours using 32 TPUs.

10.-Eigengame enables novel applications like interpretable analysis of neural networks and outperforms baselines like frequent directions.

11.-Yang Song presents a generative model based on learning to reverse a stochastic differential equation (SDE) that perturbs data into noise.

12.-A neural network is trained to estimate the score function (gradient of the log probability) of the perturbed data distribution at each noise level.

13.-Samples are generated by drawing random noise and numerically solving the learned reverse-time SDE using the score network.

14.-Exact likelihoods can be computed for the model by transforming the reverse-time SDE into an ODE and using the instantaneous change of variables formula.

15.-The model achieves state-of-the-art sample quality and likelihood scores on image datasets like CIFAR-10 and enables various conditional generation tasks.

16.-Tobias Pfaff presents MeshGraphNets, a method for learning mesh-based simulations using graph neural networks.

17.-Mesh-based finite element methods are crucial for efficient and adaptive physical simulations but rarely used in machine learning.

18.-Their model encodes a physical system as a graph with nodes/edges representing mesh elements and their connectivity in physical and world space.

19.-A learned graph neural network is applied to the input graph to predict the time evolution of the mesh in a way that generalizes to complex unseen scenarios.

20.-The same architecture can model cloth, fluids, and elasticity by simply changing the input and output mesh quantities.

21.-Learned adaptive remeshing allows the model to allocate more resolution to important regions and generalize to much larger meshes than seen during training.

Knowledge Vault built byDavid Vivancos 2024