Can Deep Learning Learn the Principle of Closed Contour Detection?

Zhang, Xinhua; Watkins, Yijing; Kenyon, Garrett T.

doi:10.1007/978-3-030-03801-4_40

Cited by 2 publications

(4 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Despite using notably different tasks such as abstract or spatial reasoning, they arrive at similar conclusions: They also observe drops in performance in the generalization regime and that interpolation is, in general, easier than extrapolation, and also hint at the modularity of models using distractor symbols [7]. Lastly, posing the concept of using correct generalization as a necessary condition to check whether an underlying mechanism has been learned has also been proposed in [22,104,108].…”

Section: Other Related Benchmark Studiesmentioning

confidence: 95%

“…Abstract reasoning Model performances on OOD generalizations are also intensively studied from the perspective of abstract reasoning, visual and relational reasoning tasks [7,22,83,98,104,106,107,108]. Most related, [7,104] also study similar interpolation and extrapolation regimes.…”

Section: Other Related Benchmark Studiesmentioning

confidence: 99%

“…In essence, machines seemingly learn to solve the tasks they have been trained on using auxiliary and spurious statistical relationships in the data, rather than true mechanistic relationships. Pragmatically, models relying on statistical relationships tend to fail if tested outside their training distribution, while models relying on (approximately) the true underlying mechanisms tend to generalize well to novel scenarios [7,19,22,61,62,72,86,104,108]. To learn effective statistical relationships, the training data needs to cover most combinations of factors of variation (like shape, size, color, viewpoint, etc.).…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Visual Representation Learning Does Not Generalize Strongly Within the Same Domain

Schott¹,

Kügelgen²,

Träuble³

et al. 2021

Preprint

View full text Add to dashboard Cite

An important component for generalization in machine learning is to uncover underlying latent factors of variation as well as the mechanism through which each factor acts in the world. In this paper, we test whether 17 unsupervised, weakly supervised, and fully supervised representation learning approaches correctly infer the generative factors of variation in simple datasets (dSprites, Shapes3D, MPI3D). In contrast to prior robustness work that introduces novel factors of variation during test time, such as blur or other (un)structured noise, we here recompose, interpolate, or extrapolate only existing factors of variation from the training data set (e.g., small and medium-sized objects during training and large objects during testing). Models that learn the correct mechanism should be able to generalize to this benchmark. In total, we train and test 2000+ models and observe that all of them struggle to learn the underlying mechanism regardless of supervision signal and architectural bias. Moreover, the generalization capabilities of all tested models drop significantly as we move from artificial datasets towards more realistic real-world datasets. Despite their inability to identify the correct mechanism, the models are quite modular as their ability to infer other in-distribution factors remains fairly stable, providing only a single factor is out-of-distribution. These results point to an important yet understudied problem of learning mechanistic models of observations that can facilitate generalization.Preprint. Under review.

show abstract

Section: Other Related Benchmark Studiesmentioning

confidence: 95%

Section: Other Related Benchmark Studiesmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Visual Representation Learning Does Not Generalize Strongly Within the Same Domain

Schott¹,

Kügelgen²,

Träuble³

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“… 35 applied predictive coding 36 to neural network architectures and proposed that predictive coding could generate similar spatial and temporal neuronal dynamics as human vision systems. Apart from those directly analyzed illusory contours in deep learning models, some studies investigate the gestalt closure law, 37 , 38 , 39 , 40 , 41 which is considered involved in the perception of illusory contours. 42 These papers generally address the illusory contour perception of neural networks from two perspectives.…”

Section: Introductionmentioning

confidence: 99%

Challenging deep learning models with image distortion based on the abutting grating illusion

Fan¹,

Zeng²

2023

Patterns

View full text Add to dashboard Cite

Can Deep Learning Learn the Principle of Closed Contour Detection?

Cited by 2 publications

References 9 publications

Visual Representation Learning Does Not Generalize Strongly Within the Same Domain

Visual Representation Learning Does Not Generalize Strongly Within the Same Domain

Challenging deep learning models with image distortion based on the abutting grating illusion

Contact Info

Product

Resources

About