The Persian Rug: solving toy models of superposition using large-scale symmetries

• URL: http://arxiv.org/abs/2410.12101v2
• Date: Tue, 22 Oct 2024 17:48:56 GMT
• Title: The Persian Rug: solving toy models of superposition using large-scale symmetries
• Authors: Aditya Cowsik, Kfir Dolev, Alex Infanger,
• Abstract summary: We present a complete mechanistic description of the algorithm learned by a minimal non-linear sparse data autoencoder in the limit of large input dimension. Our work contributes to neural network interpretability by introducing techniques for understanding the structure of autoencoders.
• Score: 0.0
• License:
• Abstract: We present a complete mechanistic description of the algorithm learned by a minimal non-linear sparse data autoencoder in the limit of large input dimension. The model, originally presented in arXiv:2209.10652, compresses sparse data vectors through a linear layer and decompresses using another linear layer followed by a ReLU activation. We notice that when the data is permutation symmetric (no input feature is privileged) large models reliably learn an algorithm that is sensitive to individual weights only through their large-scale statistics. For these models, the loss function becomes analytically tractable. Using this understanding, we give the explicit scalings of the loss at high sparsity, and show that the model is near-optimal among recently proposed architectures. In particular, changing or adding to the activation function any elementwise or filtering operation can at best improve the model's performance by a constant factor. Finally, we forward-engineer a model with the requisite symmetries and show that its loss precisely matches that of the trained models. Unlike the trained model weights, the low randomness in the artificial weights results in miraculous fractal structures resembling a Persian rug, to which the algorithm is oblivious. Our work contributes to neural network interpretability by introducing techniques for understanding the structure of autoencoders. Code to reproduce our results can be found at https://github.com/KfirD/PersianRug .

Related papers

• Scaling and renormalization in high-dimensional regression [72.59731158970894]
  This paper presents a succinct derivation of the training and generalization performance of a variety of high-dimensional ridge regression models. We provide an introduction and review of recent results on these topics, aimed at readers with backgrounds in physics and deep learning.
  arXiv  Detail & Related papers  (2024-05-01T15:59:00Z)
• Efficient and Generalizable Certified Unlearning: A Hessian-free Recollection Approach [8.875278412741695]
  Machine unlearning strives to uphold the data owners' right to be forgotten by enabling models to selectively forget specific data. We develop an algorithm that achieves near-instantaneous unlearning as it only requires a vector addition operation.
  arXiv  Detail & Related papers  (2024-04-02T07:54:18Z)
• Universal Neural Functionals [67.80283995795985]
  A challenging problem in many modern machine learning tasks is to process weight-space features. Recent works have developed promising weight-space models that are equivariant to the permutation symmetries of simple feedforward networks. This work proposes an algorithm that automatically constructs permutation equivariant models for any weight space.
  arXiv  Detail & Related papers  (2024-02-07T20:12:27Z)
• Compression of Structured Data with Autoencoders: Provable Benefit of Nonlinearities and Depth [83.15263499262824]
  We prove that gradient descent converges to a solution that completely disregards the sparse structure of the input. We show how to improve upon Gaussian performance for the compression of sparse data by adding a denoising function to a shallow architecture. We validate our findings on image datasets, such as CIFAR-10 and MNIST.
  arXiv  Detail & Related papers  (2024-02-07T16:32:29Z)
• A Dynamical Model of Neural Scaling Laws [79.59705237659547]
  We analyze a random feature model trained with gradient descent as a solvable model of network training and generalization. Our theory shows how the gap between training and test loss can gradually build up over time due to repeated reuse of data.
  arXiv  Detail & Related papers  (2024-02-02T01:41:38Z)
• Fundamental Limits of Two-layer Autoencoders, and Achieving Them with Gradient Methods [91.54785981649228]
  This paper focuses on non-linear two-layer autoencoders trained in the challenging proportional regime. Our results characterize the minimizers of the population risk, and show that such minimizers are achieved by gradient methods. For the special case of a sign activation function, our analysis establishes the fundamental limits for the lossy compression of Gaussian sources via (shallow) autoencoders.
  arXiv  Detail & Related papers  (2022-12-27T12:37:34Z)
• An Information-Theoretic Analysis of Compute-Optimal Neural Scaling Laws [24.356906682593532]
  We study the compute-optimal trade-off between model and training data set sizes for large neural networks. Our result suggests a linear relation similar to that supported by the empirical analysis of chinchilla.
  arXiv  Detail & Related papers  (2022-12-02T18:46:41Z)
• Git Re-Basin: Merging Models modulo Permutation Symmetries [3.5450828190071655]
  We show how simple algorithms can be used to fit large networks in practice. We demonstrate the first (to our knowledge) demonstration of zero mode connectivity between independently trained models. We also discuss shortcomings in the linear mode connectivity hypothesis.
  arXiv  Detail & Related papers  (2022-09-11T10:44:27Z)
• uGLAD: Sparse graph recovery by optimizing deep unrolled networks [11.48281545083889]
  We present a novel technique to perform sparse graph recovery by optimizing deep unrolled networks. Our model, uGLAD, builds upon and extends the state-of-the-art model GLAD to the unsupervised setting. We evaluate model results on synthetic Gaussian data, non-Gaussian data generated from Gene Regulatory Networks, and present a case study in anaerobic digestion.
  arXiv  Detail & Related papers  (2022-05-23T20:20:27Z)
• Provable Benefits of Overparameterization in Model Compression: From Double Descent to Pruning Neural Networks [38.153825455980645]
  Recent empirical evidence indicates that the practice of overization not only benefits training large models, but also assists - perhaps counterintuitively - building lightweight models. This paper sheds light on these empirical findings by theoretically characterizing the high-dimensional toolsets of model pruning. We analytically identify regimes in which, even if the location of the most informative features is known, we are better off fitting a large model and then pruning.
  arXiv  Detail & Related papers  (2020-12-16T05:13:30Z)
• The data-driven physical-based equations discovery using evolutionary approach [77.34726150561087]
  We describe the algorithm for the mathematical equations discovery from the given observations data. The algorithm combines genetic programming with the sparse regression. It could be used for governing analytical equation discovery as well as for partial differential equations (PDE) discovery.
  arXiv  Detail & Related papers  (2020-04-03T17:21:57Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.