What shapes feature representations? Exploring datasets, architectures, and training

In naturalistic learning problems, a model's input contains a wide range of features, some useful for the task at hand, and others not. Of the useful features, which ones does the model use? Of the task-irrelevant features, which ones does the model represent? Answers to these questions are important for understanding the basis of models' decisions, as well as for building models that learn versatile, adaptable representations useful beyond the original training task. We study these questions using synthetic datasets in which the task-relevance of input features can be controlled directly. We find that when two features redundantly predict the labels, the model preferentially represents one, and its preference reflects what was most linearly decodable from the untrained model. Over training, task-relevant features are enhanced, and task-irrelevant features are partially suppressed. Interestingly, in some cases, an easier, weakly predictive feature can suppress a more strongly predictive, but more difficult one. Additionally, models trained to recognize both easy and hard features learn representations most similar to models that use only the easy feature. Further, easy features lead to more consistent representations across model runs than do hard features. Finally, models have greater representational similarity to an untrained model than to models trained on a different task. Our results highlight the complex processes that determine which features a model represents.

[1]  Ha Hong,et al.  Explicit information for category-orthogonal object properties increases along the ventral stream , 2016, Nature Neuroscience.

[2]  David D. Cox,et al.  On the information bottleneck theory of deep learning , 2018, ICLR.

[3]  Prateek Jain,et al.  The Pitfalls of Simplicity Bias in Neural Networks , 2020, NeurIPS.

[4]  Matthias Bethge,et al.  ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness , 2018, ICLR.

[5]  Daniel L. K. Yamins,et al.  A Task-Optimized Neural Network Replicates Human Auditory Behavior, Predicts Brain Responses, and Reveals a Cortical Processing Hierarchy , 2018, Neuron.

[6]  Yaoda Xu,et al.  Limited correspondence in visual representation between the human brain and convolutional neural networks , 2020, bioRxiv.

[7]  Nikolaus Kriegeskorte,et al.  Diverse deep neural networks all predict human IT well, after training and fitting , 2020, bioRxiv.

[8]  M. Bethge,et al.  Shortcut learning in deep neural networks , 2020, Nature Machine Intelligence.

[9]  Ting Chen,et al.  The Origins and Prevalence of Texture Bias in Convolutional Neural Networks , 2019, NeurIPS.

[10]  Yann LeCun,et al.  What is the best multi-stage architecture for object recognition? , 2009, 2009 IEEE 12th International Conference on Computer Vision.

[11]  Sanjeev Arora,et al.  Implicit Regularization in Deep Matrix Factorization , 2019, NeurIPS.

[12]  Adam Gaier,et al.  Weight Agnostic Neural Networks , 2019, NeurIPS.

[13]  Naftali Tishby,et al.  Opening the Black Box of Deep Neural Networks via Information , 2017, ArXiv.

[14]  Ha Hong,et al.  Performance-optimized hierarchical models predict neural responses in higher visual cortex , 2014, Proceedings of the National Academy of Sciences.

[15]  Surya Ganguli,et al.  A mathematical theory of semantic development in deep neural networks , 2018, Proceedings of the National Academy of Sciences.

[16]  Katherine L. Hermann,et al.  Exploring the Origins and Prevalence of Texture Bias in Convolutional Neural Networks , 2019, ArXiv.

[17]  Toniann Pitassi,et al.  Learning Fair Representations , 2013, ICML.

[18]  Kaiming He,et al.  Rethinking ImageNet Pre-Training , 2018, 2019 IEEE/CVF International Conference on Computer Vision (ICCV).

[19]  知秀 柴田 5分で分かる!? 有名論文ナナメ読み:Jacob Devlin et al. : BERT : Pre-training of Deep Bidirectional Transformers for Language Understanding , 2020 .

[20]  Bruce L. McNaughton,et al.  Integration of New Information in Memory: New Insights from a Complementary Learning Systems Perspective , 2020, bioRxiv.

[21]  Razvan Pascanu,et al.  Overcoming catastrophic forgetting in neural networks , 2016, Proceedings of the National Academy of Sciences.

[22]  Alexei A. Efros,et al.  What makes ImageNet good for transfer learning? , 2016, ArXiv.

[23]  B. Schölkopf,et al.  Learning explanations that are hard to vary , 2020, ICLR.

[24]  Geoffrey E. Hinton,et al.  Similarity of Neural Network Representations Revisited , 2019, ICML.

[25]  Antonio Torralba,et al.  Comparison of deep neural networks to spatio-temporal cortical dynamics of human visual object recognition reveals hierarchical correspondence , 2016, Scientific Reports.

[26]  Dimitrios Pantazis,et al.  Dynamics of scene representations in the human brain revealed by magnetoencephalography and deep neural networks , 2015, NeuroImage.

[27]  Chico Q. Camargo,et al.  Deep learning generalizes because the parameter-function map is biased towards simple functions , 2018, ICLR.

[28]  Ali Farhadi,et al.  What’s Hidden in a Randomly Weighted Neural Network? , 2019, 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).

[29]  David J. Schwab,et al.  Training BatchNorm and Only BatchNorm: On the Expressive Power of Random Features in CNNs , 2020, ICLR.

[30]  Quoc V. Le,et al.  Do Better ImageNet Models Transfer Better? , 2018, 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).

[31]  Surya Ganguli,et al.  Exact solutions to the nonlinear dynamics of learning in deep linear neural networks , 2013, ICLR.

[32]  D. Navon Forest before trees: The precedence of global features in visual perception , 1977, Cognitive Psychology.

[33]  Mohammad Rostami,et al.  Generative Continual Concept Learning , 2019, AAAI.

[34]  Jimmy Ba,et al.  Adam: A Method for Stochastic Optimization , 2014, ICLR.

[35]  Surya Ganguli,et al.  An analytic theory of generalization dynamics and transfer learning in deep linear networks , 2018, ICLR.

[36]  Surya Ganguli,et al.  Continual Learning Through Synaptic Intelligence , 2017, ICML.

[37]  Michael McCloskey,et al.  Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem , 1989 .

[38]  Blake Lemoine,et al.  Mitigating Unwanted Biases with Adversarial Learning , 2018, AIES.

[39]  Nikolaus Kriegeskorte,et al.  Frontiers in Systems Neuroscience Systems Neuroscience , 2022 .

[40]  Nikolaus Kriegeskorte,et al.  Individual differences among deep neural network models , 2020, Nature Communications.

[41]  Sen Jia,et al.  How Much Position Information Do Convolutional Neural Networks Encode? , 2020, ICLR.

[42]  Yonatan Belinkov,et al.  Fine-grained Analysis of Sentence Embeddings Using Auxiliary Prediction Tasks , 2016, ICLR.

[43]  Nikolaus Kriegeskorte,et al.  Deep Supervised, but Not Unsupervised, Models May Explain IT Cortical Representation , 2014, PLoS Comput. Biol..

[44]  Mikhail Belkin,et al.  Reconciling modern machine-learning practice and the classical bias–variance trade-off , 2018, Proceedings of the National Academy of Sciences.

[45]  Andrea Vedaldi,et al.  Deep Image Prior , 2017, International Journal of Computer Vision.