Transformation Properties of Learned Visual Representations
Authors: Taco Cohen and Max Welling
ICLR 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We trained the model on the NORB dataset (Le Cun & Bottou, 2004). This dataset consists of objects in 5 generic categories: four-legged animals, human figures, airplanes, trucks, and cars. Each category contains 10 instances, of which we used the last 5 for training. Each instance is imaged at 9 camera elevations (30 to 70 degrees from horizontal, in 5 degree increments) and 18 azimuths (0 to 340 degrees in 20 degree increments). Finally, there are 6 lighting conditions for each instance, yielding a total of 5 5 6 9 18 = 24300 images. |
| Researcher Affiliation | Academia | Taco S. Cohen & Max Welling Machine Learning Group Department of Computer Science University of Amsterdam {t.s.cohen, m.welling}@uva.nl |
| Pseudocode | No | The paper describes the |
| Open Source Code | No | The paper does not provide any statement or link indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | We trained the model on the NORB dataset (Le Cun & Bottou, 2004). |
| Dataset Splits | No | The paper describes how the NORB dataset was used for training and testing interpolation/extrapolation. However, it does not provide specific, explicit dataset splits (e.g., exact percentages, sample counts, or formal predefined splits for training, validation, and testing) that would allow for direct reproduction of data partitioning in a standard machine learning sense. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run the experiments (e.g., specific GPU/CPU models, memory, or cloud instance types). |
| Software Dependencies | No | The paper mentions "Theano (Bergstra et al., 2010)" as a tool used, and "adagrad for all optimization (Duchi et al., 2011)", but does not provide specific version numbers for these or other software components necessary for replication. |
| Experiment Setup | Yes | We used a neural network fθ with one hidden layer containing 550 hidden units. The group representation ˆT is determined by a choice of li; i = 1, . . . , L which we chose to be: [0] 20 + [1] 15 + [2] 10 + [3] 10 + [4] 10 + [5] 9 + [6] 8 + [7] 7 + [8] 6 + [7] 7 + [8] 6 + [9] 5, where the number in brackets represents li and the multiplier denotes its multiplicity. The regularization parameters were set to β = 0.1, α = 0.1. We use adagrad for all optimization (Duchi et al., 2011). |