当前位置：网站首页>Aike AI frontier promotion (2.14)

Aike AI frontier promotion (2.14)

2022-07-04 10:46:00 【Zhiyuan community】

LG - machine learning CV - Computer vision CL - Computing and language AS - Audio and voice RO - robot

Turn from love to a lovely life

1、[LG] EquiBind: Geometric Deep Learning for Drug Binding Structure Prediction

H Stärk, O Ganea, L Pattanaik, R Barzilay, T Jaakkola

[Technical University of Munich & MIT]

EquiBind： Prediction of drug binding structure based on geometric depth learning . Predict how drug like molecules bind to specific protein targets , It is a core problem of drug discovery . A very fast calculation combination method , It will enable key applications such as rapid virtual screening or Pharmaceutical Engineering . Existing methods are computationally expensive , Rely on a large number of candidate samples , Plus the score 、 Sorting and fine-tuning steps . This article uses EquiBind Challenge this model , A kind of SE(3) Depth learning model of equivariant Geometry , Yes i） Receptor binding site ( Blind docking ) and ii） Ligand binding posture direction direct-shot forecast . Compared with traditional and recent baselines ,EquiBind Significant speed improvement and better quality . Besides , When it is combined with the existing fine-tuning Technology , Show additional improvements , But the cost is the increase of running time . A new fast tuning model is proposed , be based on von Mises The angular distance and the global minimum value of the given input atomic point cloud adjust the torsion angle of the ligand rotatable key , The differential evolution strategy avoids the previously expensive energy minimization .

Predicting how a drug-like molecule binds to a specific protein target is a core problem in drug discovery. An extremely fast computational binding method would enable key applications such as fast virtual screening or drug engineering. Existing methods are computationally expensive as they rely on heavy candidate sampling coupled with scoring, ranking, and fine-tuning steps. We challenge this paradigm with EQUIBIND, an SE(3)-equivariant geometric deep learning model performing direct-shot prediction of both i) the receptor binding location (blind docking) and ii) the ligand’s bound pose and orientation. EquiBind achieves significant speed-ups and better quality compared to traditional and recent baselines. Further, we show extra improvements when coupling it with existing fine-tuning techniques at the cost of increased running time. Finally, we propose a novel and fast fine-tuning model that adjusts torsion angles of a ligand’s rotatable bonds based on closed-form global minima of the von Mises angular distance to a given input atomic point cloud, avoiding previous expensive differential evolution strategies for energy minimization.

2、[LG] Riemannian Score-Based Generative Modeling

V D Bortoli, E Mathieu, M Hutchinson, J Thornton, Y W Teh, A Doucet

[University of Oxford]

Generation model based on Riemann fraction . Score based generation model (SGM) Is a new generation model , Demonstrated significant empirical performance . Use diffusion to gradually add Gaussian noise to the data , The generative model is a " Denoise " The process , By approximating this " noise " Time inversion of diffusion to obtain . However , at present SGM The basic assumption is , The data is supported on a Euclidean manifold with plane geometry . This hinders these models in robots 、 Applications in geoscience or protein modeling , These applications depend on distributions defined on Riemannian manifolds . To overcome this problem , This paper proposes a generation model based on Riemann fraction (RSGM), The current SGM Extended to compact Riemannian manifold environment , Its main advantage is its scalability to high dimensions , Due to the diversity of available loss functions , It is suitable for a wide class of manifolds , And its modeling ability for complex data sets . Explain the method with earth and climate science data , And explain RSGM How to speed up by solving the Schrodinger bridge problem on manifolds .

Score-based generative models (SGMs) are a novel class of generative models demonstrating remarkable empirical performance. One uses a diffusion to add gradually Gaussian noise to the data, while the generative model is a “denoising” process obtained by approximating the time-reversal of this “noising” diffusion. However, current SGMs make the underlying assumption that the data is supported on a Euclidean manifold with flat geometry. This prevents the use of these models for applications in robotics, geoscience or protein modeling which rely on distributions defined on Riemannian manifolds. To overcome this issue, we introduce Riemannian Score-based Generative Models (RSGMs) which extend current SGMs to the setting of compact Riemannian manifolds. We illustrate our approach with earth and climate science data and show how RSGMs can be accelerated by solving a Schrödinger bridge problem on manifolds. Keywords— Diffusion processes, Generative modeling, Riemannian manifold, Score-based generative models, Schrödinger bridge

3、[LG] Deconstructing The Inductive Biases Of Hamiltonian Neural Networks

N Gruver, M Finzi, S Stanton, A G Wilson

[New York University]

Hamiltonian neural network inductive bias deconstruction . Neural networks inspired by physics , Such as Hamilton or Lagrange neural network , Take advantage of strong inductive bias , It greatly surpasses other acquisition dynamics models . However , These models are challenging when applied to many real-world systems , For example, those systems that do not conserve energy or contain contact , This is a common setting for robots and reinforcement learning . This paper studies the inductive bias that makes the physics heuristic model successful in practice . High performance HNN The inductive deviation decomposition of the model is its component , namely NeuralODE、 Sympathetic 、 Learn the conservation of energy function and second-order structure . Contrary to the traditional view ,HNN The generalization improvement of comes from the assumption that the system can be expressed as a single second-order differential equation , Avoid artificial complexity from the coordinate system , Instead of conjugate structures or conservation of energy . Peel off HNN Other components of , The rest is a simpler one 、 More computationally efficient models , And less restrictive , It can be directly applied to non Hamilton System . By relaxing the inductive bias of these models , It can match or exceed the performance of energy conservation system , At the same time, it greatly improves the performance of the actual non conservative system . Extend this approach to common Mujoco Environment build transition model , It shows that the model can properly balance inductive bias and the flexibility required by model-based control . The model obtained by application is challenging Mujoco Build a transition model in the sports environment , Gratifying achievements have been made .

Physics-inspired neural networks (NNs), such as Hamiltonian or Lagrangian NNs, dramatically outperform other learned dynamics models by leveraging strong inductive biases. These models, however, are challenging to apply to many real world systems, such as those that don’t conserve energy or contain contacts, a common setting for robotics and reinforcement learning. In this paper, we examine the inductive biases that make physics-inspired models successful in practice. We show that, contrary to conventional wisdom, the improved generalization of HNNs is the result of modeling acceleration directly and avoiding artificial complexity from the coordinate system, rather than symplectic structure or energy conservation. We show that by relaxing the inductive biases of these models, we can match or exceed performance on energy-conserving systems while dramatically improving performance on practical, non-conservative systems. We extend this approach to constructing transition models for common Mujoco environments, showing that our model can appropriately balance inductive biases with the flexibility required for model-based control.

4、[CV] FEAT: Face Editing with Attention

X Hou, L Shen, O Patashnik, D Cohen-Or, H Huang

[Shenzhen University & Tel Aviv University]

FEAT： Attention based face editing . The use of pre training generators for latent space has recently been shown to be based on GAN An effective means of face manipulation . Its success largely depends on the natural unwrapping of the hidden space axis of the generator . However , Face manipulation is often intended to affect local areas , While ordinary generators often do not have the necessary spatial decomposition . In this paper StyleGAN Based on the generator , Put forward a way , Explicitly encourage face manipulation to focus on a predetermined area by incorporating learned attention maps . In the process of editing image generation , Attention map as a mask , Guide the fusion between original features and modified features . The guidance of hidden space editor is through CLIP Realized ,CLIP Recently, it has been proved to be effective for text driven editing . Extensive experiments have been carried out in this paper , It shows that this method can decompose based on text description and controllable face manipulation by focusing only on relevant regions . Qualitative and quantitative experimental results show that the proposed method is superior to other methods in face region editing .

Employing the latent space of pretrained generators has recently been shown to be an effective means for GANbased face manipulation. The success of this approach heavily relies on the innate disentanglement of the latent space axes of the generator. However, face manipulation often intends to affect local regions only, while common generators do not tend to have the necessary spatial disentanglement. In this paper, we build on the StyleGAN generator, and present a method that explicitly encourages face manipulation to focus on the intended regions by incorporating learned attention maps. During the generation of the edited image, the attention map serves as a mask that guides a blending between the original features and the modified ones. The guidance for the latent space edits is achieved by employing CLIP, which has recently been shown to be effective for text-driven edits. We perform extensive experiments ∗Corresponding author and show that our method can perform disentangled and controllable face manipulations based on text descriptions by attending to the relevant regions only. Both qualitative and quantitative experimental results demonstrate the superiority of our method for facial region editing over alternative methods.

5、[CV] DiffusionNet: Discretization Agnostic Learning on Surfaces

N Sharp, S Attaiki, K Crane, M Ovsjanikov

[University of Toronto & École Polytechnique & CMU]

DiffusionNet: Surface discretization independent learning . This paper presents a new method in 3D A general method of deep learning on the surface , Very effective insight into space communication based on simple diffusion layer , Use spatial gradient features to inject directional information . The resulting network is automatically robust to surface resolution and sampling changes —— This is a fundamental attribute that is essential for practical applications . The proposed network can be discretized in various geometric representations , Such as triangular mesh or point cloud , You can even train on a representation , Then applied to another representation . The diffusion space support is optimized as a continuous network parameter , The range ranges from pure local to complete global , Eliminates the burden of manually selecting neighborhood size . The only other component of this method is the multilayer perceptron applied independently at each point , And spatial gradient feature supporting directional filter . The resulting network is simple 、 Robust and efficient . ad locum , Focus on triangular mesh surfaces , And show the most advanced results of various tasks , Including surface classification 、 Segmentation and non rigid correspondence .

We introduce a new general-purpose approach to deep learning on 3D surfaces, based on the insight that a simple diffusion layer is highly effective for spatial communication. The resulting networks are automatically robust to changes in resolution and sampling of a surface—a basic property which is crucial for practical applications. Our networks can be discretized on various geometric representations such as triangle meshes or point clouds, and can even be trained on one representation then applied to another. We optimize the spatial support of diffusion as a continuous network parameter ranging from purely local to totally global, removing the burden ofmanually choosing neighborhood sizes. The only other ingredients in the method are a multilayer perceptron applied independently at each point, and spatial gradient features to support directional filters. The resulting networks are simple, robust, and efficient. Here, we focus primarily on triangle mesh surfaces, and demonstrate state-of-the-art results for a variety of tasks including surface classification, segmentation, and non-rigid correspondence.