Trading robust representations for sample complexity through self-supervised visual experience
Authors: Andrea Tacchetti, Stephen Voinea, Georgios Evangelopoulos
NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experimental evaluation of learned image representations |
| Researcher Affiliation | Collaboration | The Center for Brains, Minds and Machines, MIT Mc Govern Institute for Brain Research at MIT Cambridge, MA, USA {atacchet, voinea, gevang}@mit.edu; Currently with Deep Mind. Currently with X, Alphabet. |
| Pseudocode | No | The paper describes methods mathematically but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described in this paper. |
| Open Datasets | Yes | Table 1: Datasets, orbits and canonical definitions used in our experimental evaluations. Overall, the sets include 3D viewpoint, light and unconstrained transformations; MNIST is the only one with analytic, known transformation orbits. |
| Dataset Splits | Yes | For each dataset, we employ an Embedding set to learn representations and a separate Validation set for determining the optimal number of SGD iterations (early stopping hyperparameter) by maximizing the mean, across multiple re-splits of the set, of task-specific performance metrics. The representations are used for encoding Train and Test sets from a Supervised set for the downstream recognition task. Performance is reported as the mean/std of multiple train/test re-splits of the Supervised set. For all experiments, there is no overlap between embedding, validation, and supervised sets, or between train/test splits, within the validation and the supervised sets. |
| Hardware Specification | Yes | The DGX-1 used for our experiments was donated by NVIDIA. |
| Software Dependencies | No | The paper mentions software components like 'Adam optimizer' and 'VGG networks' but does not specify their version numbers or other software dependencies with versions. |
| Experiment Setup | Yes | Loss minimization was carried out with mini-batch Stochastic Gradient Descent using the Adam optimizer. For MNIST, we used mini-batches of size 256, for Multi-PIE, 72, for NORB, 256, for Multi-PIE, 72 and for the Late Night dataset, 128 images. The selection of triplets for ST, OT and OJ followed the soft negative selection process from [25]. The values for λ1 and λ2 were set equal (λ1 = λ2 = 1). |