Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
194 tokens/sec
GPT-4o
7 tokens/sec
Gemini 2.5 Pro Pro
45 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Expand-and-Cluster: Parameter Recovery of Neural Networks (2304.12794v4)

Published 25 Apr 2023 in cs.NE

Abstract: Can we identify the weights of a neural network by probing its input-output mapping? At first glance, this problem seems to have many solutions because of permutation, overparameterisation and activation function symmetries. Yet, we show that the incoming weight vector of each neuron is identifiable up to sign or scaling, depending on the activation function. Our novel method 'Expand-and-Cluster' can identify layer sizes and weights of a target network for all commonly used activation functions. Expand-and-Cluster consists of two phases: (i) to relax the non-convex optimisation problem, we train multiple overparameterised student networks to best imitate the target function; (ii) to reverse engineer the target network's weights, we employ an ad-hoc clustering procedure that reveals the learnt weight vectors shared between students -- these correspond to the target weight vectors. We demonstrate successful weights and size recovery of trained shallow and deep networks with less than 10\% overhead in the layer size and describe an `ease-of-identifiability' axis by analysing 150 synthetic problems of variable difficulty.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (71)
  1. An initial alignment between neural network and target is needed for gradient descent to learn. In International Conference on Machine Learning, pp.  33–52. PMLR, 2022.
  2. Git re-basin: Merging models modulo permutation symmetries. arXiv preprint arXiv:2209.04836, 2022.
  3. Towards understanding ensemble, knowledge distillation and self-distillation in deep learning. arXiv preprint arXiv:2012.09816, 2020.
  4. Analytic study of families of spurious minima in two-layer relu neural networks: a tale of symmetry ii. Advances in Neural Information Processing Systems, 34:15162–15174, 2021.
  5. Knowledge distillation: A good teacher is patient and consistent. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.  10925–10934, 2022.
  6. Parameter identifiability of a deep feedforward relu neural network. arXiv preprint arXiv:2112.12982, 2021.
  7. Mlpgradientflow: going with the flow of multilayer perceptrons (and finding minima fast and accurately). arXiv preprint arXiv:2301.10638, 2023.
  8. Functional vs. parametric equivalence of relu networks. In 8th International Conference on Learning Representations, 2020.
  9. Cryptanalytic extraction of neural network models. In Annual International Cryptology Conference, pp.  189–218. Springer, 2020.
  10. Only train once: A one-shot neural network training and pruning framework. Advances in Neural Information Processing Systems, 34:19637–19651, 2021.
  11. Otov2: Automatic, generic, user-friendly. In The Eleventh International Conference on Learning Representations, 2022.
  12. On the global convergence of gradient descent for over-parameterized models using optimal transport. Advances in neural information processing systems, 31, 2018.
  13. Cooper, Y. The loss landscape of overparameterized neural networks. arXiv preprint arXiv:1804.10200, 2018.
  14. Geometry of the loss landscape in overparameterized neural networks: Symmetries and invariances. In International Conference on Machine Learning, pp.  9722–9732. PMLR, 2021.
  15. Gradient descent provably optimizes over-parameterized neural networks. In International Conference on Learning Representations, 2019.
  16. Toy models of superposition. Transformer Circuits Thread, 2022. https://transformer-circuits.pub/2022/toy_model/index.html.
  17. The role of permutation invariance in linear mode connectivity of neural networks. In 10th International Conference on Learning Representations: ICLR 2022, 2022.
  18. Stable recovery of entangled weights: Towards robust identification of deep neural networks from minimal samples. Applied and Computational Harmonic Analysis, 62:123–172, 2023.
  19. Robust and resource-efficient identification of two hidden layer neural networks. Constructive Approximation, pp.  1–62, 2019.
  20. Robust and resource efficient identification of shallow neural networks by fewest samples. Information and Inference: A Journal of the IMA, 10(2):625–695, 2021.
  21. Finite sample identification of wide shallow neural networks with biases. arXiv preprint arXiv:2211.04589, 2022.
  22. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations, 2019.
  23. Agnostic learning of a single neuron with gradient descent. Advances in Neural Information Processing Systems, 33:5417–5428, 2020.
  24. Guaranteed recovery of one-hidden-layer neural networks via cross entropy. IEEE transactions on signal processing, 68:3225–3235, 2020.
  25. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pp.  249–256. JMLR Workshop and Conference Proceedings, 2010.
  26. Dynamics of stochastic gradient descent for two-layer neural networks in the teacher-student setup. Advances in neural information processing systems, 32, 2019.
  27. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015.
  28. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016.
  29. Sparsity in deep learning: Pruning and growth for efficient inference and training in neural networks. J. Mach. Learn. Res., 22(241):1–124, 2021.
  30. Network trimming: A data-driven neuron pruning approach towards efficient deep architectures. arXiv preprint arXiv:1607.03250, 2016.
  31. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018.
  32. High accuracy and high fidelity extraction of neural networks. In 29th USENIX security symposium (USENIX Security 20), pp.  1345–1362, 2020.
  33. Beating the perils of non-convexity: Guaranteed training of neural networks using tensor methods. arXiv preprint arXiv:1506.08473, 2015.
  34. Repair: Renormalizing permuted activations for interpolation repair. arXiv preprint arXiv:2211.08403, 2022.
  35. Adam: A method for stochastic optimization. In ICLR, 2015.
  36. Self-normalizing neural networks. Advances in neural information processing systems, 30, 2017.
  37. Feature selection with the boruta package. Journal of statistical software, 36:1–13, 2010.
  38. Lasserre, J. B. Global optimization with polynomials and the problem of moments. SIAM Journal on optimization, 11(3):796–817, 2001.
  39. LeCun, Y. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998.
  40. Optimal brain damage. Advances in neural information processing systems, 2, 1989.
  41. Cones of matrices and set-functions and 0–1 optimization. SIAM journal on optimization, 1(2):166–190, 1991.
  42. The landscape of empirical risk for nonconvex losses. The Annals of Statistics, 46(6A):2747–2774, 2018.
  43. Misra, D. Mish: A self regularized non-monotonic activation function. arXiv preprint arXiv:1908.08681, 2019.
  44. On the connection between learning two-layer neural networks and tensor decomposition. In The 22nd International Conference on Artificial Intelligence and Statistics, pp.  1051–1060. PMLR, 2019.
  45. Algorithms for hierarchical clustering: an overview. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 2(1):86–97, 2012.
  46. Data-independent structured pruning of neural networks via coresets. IEEE Transactions on Neural Networks and Learning Systems, 2021.
  47. The building blocks of interpretability. Distill, 3(3):e10, 2018.
  48. I know what you trained last summer: A survey on stealing machine learning models and defences. arXiv preprint arXiv:2206.08451, 2022.
  49. Notes on the symmetries of 2-layer relu-networks. In Proceedings of the Northern Lights Deep Learning Workshop, volume 1, pp.  6–6, 2020.
  50. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017.
  51. Fundamental bounds on learning performance in neural circuits. Proceedings of the National Academy of Sciences, 116(21):10537–10546, 2019.
  52. Reverse-engineering deep relu networks. In International Conference on Machine Learning, pp.  8178–8187. PMLR, 2020.
  53. Trainability and accuracy of artificial neural networks: An interacting particle system approach. Communications on Pure and Applied Mathematics, 75(9):1889–1935, 2022.
  54. Learning internal representations by error propagation. In Rumelhart, D. E., McClelland, J. L., and Group, P. R. (eds.), Parallel Distributed Processing, volume 1, chapter 8, pp.  318–362. MIT press Cambridge, MA, 1986.
  55. On-line learning in soft committee machines. Physical Review E, 52(4):4225, 1995.
  56. Spurious local minima are common in two-layer relu neural networks. In International conference on machine learning, pp.  4433–4441. PMLR, 2018.
  57. A hierarchy of relaxations between the continuous and convex hull representations for zero-one programming problems. SIAM Journal on Discrete Mathematics, 3(3):411–430, 1990.
  58. Model fusion via optimal transport. Advances in Neural Information Processing Systems, 33:22045–22055, 2020.
  59. Data-free parameter pruning for deep neural networks. In Proceedings of the British Machine Vision Conference (BMVC), pp.  31.1–31.12, September 2015.
  60. An embedding of relu networks and an analysis of their identifiability. Constructive Approximation, pp.  1–47, 2022.
  61. Sun, J. Provable nonconvex methods/algorithms, 2021. URL https://sunju.org/research/nonconvex/.
  62. Sussmann, H. J. Uniqueness of the weights for minimal feedforward nets with a given input-output map. Neural networks, 5(4):589–593, 1992.
  63. Tian, Y. Student specialization in deep rectified networks with finite width and input dimension. In International Conference on Machine Learning, pp.  9470–9480. PMLR, 2020.
  64. Ai feynman: A physics-inspired method for symbolic regression. Science Advances, 6(16):eaay2631, 2020.
  65. Affine symmetries and neural network identifiability. Advances in Mathematics, 376:107485, 2021.
  66. Neural network identifiability for a family of sigmoidal nonlinearities. Constructive Approximation, 55(1):173–224, 2022.
  67. Federated learning with matched averaging. In International Conference on Learning Representations, 2020.
  68. Learning a single neuron with gradient methods. In Conference on Learning Theory, pp.  3756–3786. PMLR, 2020.
  69. From symmetry to geometry: Tractable nonconvex problems. arXiv preprint arXiv:2007.06753, 2020.
  70. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021.
  71. Recovery guarantees for one-hidden-layer neural networks. In International conference on machine learning, pp.  4140–4149. PMLR, 2017.
Citations (2)
View on Semantic Scholar

Summary

We haven't generated a summary for this paper yet.

Ai Sparkles Streamline Icon: https://streamlinehq.com Summarize Now
Youtube Logo Streamline Icon: https://streamlinehq.com

YouTube