Revisiting Model Stitching to Compare Neural Representations
- URL: http://arxiv.org/abs/2106.07682v1
- Date: Mon, 14 Jun 2021 18:05:10 GMT
- Title: Revisiting Model Stitching to Compare Neural Representations
- Authors: Yamini Bansal, Preetum Nakkiran, Boaz Barak
- Abstract summary: We consider a "stitched model" formed by connecting the bottom-layers of $A$ to the top-layers of $B$, with a simple trainable layer between them.
We show that good networks of the same architecture, but trained in very different ways, can be stitched to each other without drop in performance.
We also give evidence for the intuition that "more is better" by showing that representations learnt with (1) more data, (2) bigger width, or (3) more training time can be "plugged in'' to weaker models to improve performance.
- Score: 8.331711958610347
- License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
- Abstract: We revisit and extend model stitching (Lenc & Vedaldi 2015) as a methodology
to study the internal representations of neural networks. Given two trained and
frozen models $A$ and $B$, we consider a "stitched model'' formed by connecting
the bottom-layers of $A$ to the top-layers of $B$, with a simple trainable
layer between them. We argue that model stitching is a powerful and perhaps
under-appreciated tool, which reveals aspects of representations that measures
such as centered kernel alignment (CKA) cannot. Through extensive experiments,
we use model stitching to obtain quantitative verifications for intuitive
statements such as "good networks learn similar representations'', by
demonstrating that good networks of the same architecture, but trained in very
different ways (e.g.: supervised vs. self-supervised learning), can be stitched
to each other without drop in performance. We also give evidence for the
intuition that "more is better'' by showing that representations learnt with
(1) more data, (2) bigger width, or (3) more training time can be "plugged in''
to weaker models to improve performance. Finally, our experiments reveal a new
structural property of SGD which we call "stitching connectivity'', akin to
mode-connectivity: typical minima reached by SGD can all be stitched to each
other with minimal change in accuracy.
Related papers
- Neural Metamorphosis [72.88137795439407]
This paper introduces a new learning paradigm termed Neural Metamorphosis (NeuMeta), which aims to build self-morphable neural networks.
NeuMeta directly learns the continuous weight manifold of neural networks.
It sustains full-size performance even at a 75% compression rate.
arXiv Detail & Related papers (2024-10-10T14:49:58Z) - Improving Network Interpretability via Explanation Consistency Evaluation [56.14036428778861]
We propose a framework that acquires more explainable activation heatmaps and simultaneously increase the model performance.
Specifically, our framework introduces a new metric, i.e., explanation consistency, to reweight the training samples adaptively in model learning.
Our framework then promotes the model learning by paying closer attention to those training samples with a high difference in explanations.
arXiv Detail & Related papers (2024-08-08T17:20:08Z) - Learning Goal-Conditioned Representations for Language Reward Models [10.94845204766088]
We propose training reward models (RMs) in a contrastive, $textitgoal-conditioned$ fashion.
We show this way of training RM representations enables improved $textitsteerability$ because it allows us to evaluate the likelihood of an action achieving a particular goal-state.
We additionally find that these representations can perform fine-grained control by conditioning on desired future goal-states.
arXiv Detail & Related papers (2024-07-18T20:23:11Z) - On Layer-wise Representation Similarity: Application for Multi-Exit Models with a Single Classifier [20.17288970927518]
We study the similarity of representations between the hidden layers of individual transformers.
We propose an aligned training approach to enhance the similarity between internal representations.
arXiv Detail & Related papers (2024-06-20T16:41:09Z) - Model Stitching: Looking For Functional Similarity Between
Representations [5.657258033928475]
We expand on a previous work which used model stitching to compare representations of the same shapes learned by differently seeded and/or trained neural networks of the same architecture.
We reveal unexpected behavior of model stitching. Namely, we find that stitching, based on convolutions, for small ResNets, can reach high accuracy if those layers come later in the first (sender) network than in the second (receiver)
arXiv Detail & Related papers (2023-03-20T17:12:42Z) - Improved Convergence Guarantees for Shallow Neural Networks [91.3755431537592]
We prove convergence of depth 2 neural networks, trained via gradient descent, to a global minimum.
Our model has the following features: regression with quadratic loss function, fully connected feedforward architecture, RelU activations, Gaussian data instances, adversarial labels.
They strongly suggest that, at least in our model, the convergence phenomenon extends well beyond the NTK regime''
arXiv Detail & Related papers (2022-12-05T14:47:52Z) - Part-Based Models Improve Adversarial Robustness [57.699029966800644]
We show that combining human prior knowledge with end-to-end learning can improve the robustness of deep neural networks.
Our model combines a part segmentation model with a tiny classifier and is trained end-to-end to simultaneously segment objects into parts.
Our experiments indicate that these models also reduce texture bias and yield better robustness against common corruptions and spurious correlations.
arXiv Detail & Related papers (2022-09-15T15:41:47Z) - How Well Do Sparse Imagenet Models Transfer? [75.98123173154605]
Transfer learning is a classic paradigm by which models pretrained on large "upstream" datasets are adapted to yield good results on "downstream" datasets.
In this work, we perform an in-depth investigation of this phenomenon in the context of convolutional neural networks (CNNs) trained on the ImageNet dataset.
We show that sparse models can match or even outperform the transfer performance of dense models, even at high sparsities.
arXiv Detail & Related papers (2021-11-26T11:58:51Z) - Distill on the Go: Online knowledge distillation in self-supervised
learning [1.1470070927586016]
Recent works have shown that wider and deeper models benefit more from self-supervised learning than smaller models.
We propose Distill-on-the-Go (DoGo), a self-supervised learning paradigm using single-stage online knowledge distillation.
Our results show significant performance gain in the presence of noisy and limited labels.
arXiv Detail & Related papers (2021-04-20T09:59:23Z) - Improving Robustness and Generality of NLP Models Using Disentangled
Representations [62.08794500431367]
Supervised neural networks first map an input $x$ to a single representation $z$, and then map $z$ to the output label $y$.
We present methods to improve robustness and generality of NLP models from the standpoint of disentangled representation learning.
We show that models trained with the proposed criteria provide better robustness and domain adaptation ability in a wide range of supervised learning tasks.
arXiv Detail & Related papers (2020-09-21T02:48:46Z) - From Boltzmann Machines to Neural Networks and Back Again [31.613544605376624]
We give new results for learning Restricted Boltzmann Machines, probably the most well-studied class of latent variable models.
Our results are based on new connections to learning two-layer neural networks under $ell_infty$ bounded input.
We then give an algorithm for learning a natural class of supervised RBMs with better runtime than what is possible for its related class of networks without distributional assumptions.
arXiv Detail & Related papers (2020-07-25T00:42:50Z)
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.