The End Of Knowledge - Vault 6/41 - CVPR - 2019 - Machine learning for robots to think fast

graph LR classDef main fill:#f9d4f9, font-weight:bold, font-size:14px classDef robotics fill:#f9d4d4, font-weight:bold, font-size:14px classDef ml fill:#d4f9d4, font-weight:bold, font-size:14px classDef matrix fill:#d4d4f9, font-weight:bold, font-size:14px classDef sparse fill:#f9f9d4, font-weight:bold, font-size:14px classDef impact fill:#d4f9f9, font-weight:bold, font-size:14px Main[Machine learning for
robots to think
fast] --> A[Robotics and
Machine Learning] Main --> B[Machine Learning
Applications] Main --> C[Matrix Factorization] Main --> D[Sparse Coding] Main --> E[Impact and
Future Directions] A --> A1[Billard: EPFL robotics, machine learning
professor 1] A --> A2[Fast-moving object catching challenges robots 2] A --> A3[ML models uncertainty, enables fast
computation 3] A --> A4[Dynamical systems generate stable control
policies 4] A --> A5[Multi-attractor systems adapt grasp configurations 5] A --> A6[SVMs learn attractor switching boundaries 6] B --> B1[GMMs represent robot configurations for
planning 7] B --> B2[Optimal control generates training data
offline 8] B --> B3[Tactile, vision enable on-the-fly grasp
adaptation 9] B --> B4[ML crucial for deformable object
manipulation 10] B --> B5[Control theory, ML combination is
important 11] B --> B6[Online dictionary learning papers significant
impact 12] C --> C1[Matrix factorization decomposes data, unsupervised
learning 13] C --> C2[Matrix factorization variants yield different
techniques 14] C --> C3[Matrix factorization successful in various
domains 18] C --> C4[Alternating minimization unsuitable for large
datasets 19] C --> C5[SGD scalable but requires careful
tuning 20] C --> C6[Authors proposed fast, tuning-free online
algorithm 21] D --> D1[Sparse coding discovers useful image
features 15] D --> D2[Sparsity principle has long history 16] D --> D3[Sparse coding led to image
restoration breakthroughs 17] D --> D4[Key ideas: fast updates, online
approximation 22] D --> D5[Convergence guaranteed despite optimization challenges 23] D --> D6[Dictionary learning connected to neural
networks 26] E --> E1[Impact from timeliness, software, flexibility 24] E --> E2[SPAMS toolbox found diverse applications 25] E --> E3[Convolutional, multi-layer sparse coding possible 27] E --> E4[Simplicity principle remains relevant for
interpretability 28] E --> E5[Simplicity alone insufficient, consider other
properties 29] E --> E6[Easy-to-use software increases research impact 30] class Main main class A,A1,A2,A3,A4,A5,A6 robotics class B,B1,B2,B3,B4,B5,B6 ml class C,C1,C2,C3,C4,C5,C6 matrix class D,D1,D2,D3,D4,D5,D6 sparse class E,E1,E2,E3,E4,E5,E6 impact

Resume:

1.- Aude Billard is a robotics and machine learning professor at EPFL who has made significant contributions in robot control and learning.

2.- Controlling robots to catch fast-moving objects is challenging due to uncertainty in the object's motion and limited time for computation.

3.- Machine learning can help model uncertainty and enable fast computation of control paths compared to pure control theory approaches.

4.- Dynamical systems can be learned from demonstration to generate stable, closed-form control policies that enable real-time tracking of moving targets.

5.- Multi-attractor dynamical systems allow the robot to adapt its grasp configuration in real-time to catch objects like flying bottles.

6.- Support vector machines can be used to learn switching boundaries between different attractor systems for fast runtime decision making.

7.- Gaussian mixture models can compactly represent probability distributions over robot configurations for collision-free path planning between multiple collaborating arms.

8.- Optimal control can be used to generate training data offline when human demonstrations are unavailable, and reinforcement learning can further optimize policies.

9.- Tactile sensing and vision enable a robot to estimate an object's pose and adapt its grasp on the fly.

10.- Machine learning is crucial for manipulation of deformable objects like food items, where explicit physical models are very difficult to obtain.

11.- The combination of control theory and machine learning is important - control provides feasible solutions and guarantees, while ML efficiently approximates policies.

12.- The online dictionary learning for sparse coding paper from 2009 has had significant impact in the past decade.

13.- Matrix factorization decomposes a data matrix into a dictionary and sparse code, useful for unsupervised learning and efficient representation.

14.- Many variants of matrix factorization exist with different constraints on the factors, yielding techniques like NMF, sparse PCA, structured sparse coding.

15.- Sparse coding, introduced in 1996, automatically discovers useful features like Gabor filters from natural image patches via a sparsity prior.

16.- The sparsity principle of selecting the simplest explanation for data has a long history, from Rinsch and Jeffries in the 1920s onward.

17.- In the 2000s, sparse coding led to state-of-the-art results in image denoising, inpainting, and other restoration tasks.

18.- Matrix factorization also found success in computer vision, winning Pascal VOC and ImageNet challenges, as well as collaborative filtering and bioinformatics.

19.- Alternating minimization is a standard algorithm for matrix factorization but requires loading all data into memory, not suitable for large datasets.

20.- Stochastic gradient descent is more scalable but requires carefully tuning learning rates and can be slow to converge.

21.- The authors proposed an online algorithm that is as fast as well-tuned SGD but does not require learning rate tuning.

22.- Key ideas are: 1) dictionary update is fast if optimal sparse codes are known, 2) optimal codes can be approximated online.

23.- Convergence to stationary points is guaranteed by exploiting problem structure, despite the optimization being non-convex, constrained, and stochastic.

24.- The work's impact comes from timeliness (increasing dataset sizes), robust/practical software, and flexibility to adapt to various applications.

25.- The SPAMS toolbox implementing the algorithm has found diverse applications, from modeling tree leaves to analyzing spatial gene expression patterns.

26.- Dictionary learning has connections to neural networks - sparse coding is related to ReLU activations, and end-to-end learning via backprop is possible.

27.- Convolutional and multi-layer sparse coding is possible, related to approaches like the LISTA neural network architecture.

28.- The simplicity principle behind sparsity remains relevant today for interpretability and model selection, though perhaps in a different form than the L1 norm.

29.- Simplicity alone is likely not enough, and other properties like robustness and stability are also important to consider.

30.- Releasing easy-to-use, robust software together with research publications can significantly increase impact, including on fields beyond the original domain.

Knowledge Vault built byDavid Vivancos 2024