Analyzing the Sample Complexity of Self-Supervised Image Reconstruction Methods
Authors: Tobit Klug, Dogukan Atik, Reinhard Heckel
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We then study self-supervised denoising and accelerated MRI empirically and characterize the cost of self-supervised training in terms of the number of additional samples required, and find that the performance gap between self-supervised and supervised training vanishes as a function of the training examples, at a problem-dependent rate, as predicted by our theory. |
| Researcher Affiliation | Academia | Tobit Klug, Dogukan Atik, Reinhard Heckel School of Computation, Information and Technology Technical University of Munich tobit.klug, dogukan.atik, reinhard.heckel {@tum.de} |
| Pseudocode | No | The paper does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | The repository at https://github.com/MLI-lab/sample_complexity_ss_ recon contains the code to reproduce all results in the main body of this paper. |
| Open Datasets | Yes | We use cropped patches of size 128 128 from Image Net [37] to create training sets with 100 to 300k images. We conduct our experiments on a subset of the fast MRI brain dataset [46]. |
| Dataset Splits | Yes | We use 50k of those images to design training sets SN of varying size N with Si Sj for i < j, 300 for validation and the remaining 4700 for testing. For the real-world camera noise, 'validate on the first 10 scenes, and report the test performance on the remaining 30 scenes from the validation set.' |
| Hardware Specification | Yes | All experiments were conducted on NVIDIA A40, NVIDIA RTX A6000 and NVIDIA Quadro RTX 6000 GPUs. |
| Software Dependencies | No | The paper mentions software like 'Adam optimizer' and 'RMSprop optimizer' and 'ESPIRi T', but it does not provide specific version numbers for any of these software dependencies. |
| Experiment Setup | Yes | We use cropped patches of size 128 128 from Image Net [37] to create training sets with 100 to 300k images. We pick the best run out of different runs with independently sampled network initialization. We use cropped images of size 100 100 from Image Net [37] as ground truth images x and obtain undersampled complex-valued measurements in the frequency domain as y = MFx and y = M Fx. For small training set sizes the curves in Figure 8 use the best performance out of up to 5 training runs with different random network initialization, indicated by the additional points underneath the different curves. We tested two different batch sizes 1 and 10 for training set sizes up to 60k images... All models are trained with the Adam optimizer [21] with β1 = 0.9 and β2 = 0.999. Starting from a small learning rate 1.25 10 6 we double the learning rate after every epoch as long as the validation error improves. We halve the learning rate once the validation PSNR did not improve for eight consecutive epochs. |