Learning predictable and robust neural representations by straightening image sequences
Authors: Julie Xueyan Niu, Cristina Savin, Eero Simoncelli
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the power of this objective in training deep feedforward neural networks on smoothly-rendered synthetic image sequences that mimic commonly-occurring properties of natural videos. |
| Researcher Affiliation | Academia | 1Center for Neural Science, New York University 2Center for Data Science, New York University 3Center for Computational Neuroscience, Flatiron Institute |
| Pseudocode | No | The paper presents mathematical formulations for the objective functions and describes the network architecture, but it does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The implementation can be found at https://github.com/xyniu1/learning-by-straightening. |
| Open Datasets | Yes | For the first set of experiments, we created a sequential MNIST dataset... For a second set of experiments, we generated a sequential CIFAR-10 dataset... |
| Dataset Splits | No | The paper describes various training procedures and dataset augmentations, but it does not explicitly state the specific percentages or sample counts for training, validation, or test splits for any of the datasets used. |
| Hardware Specification | Yes | All pretraining was run for 1000 epochs, which takes roughly 5 hours on 4 A100 Nvidia GPUs. |
| Software Dependencies | No | The paper mentions using 'an adapted version of the solo-learn library' and the 'LARS optimizer', but it does not provide specific version numbers for these or other underlying software dependencies (e.g., Python, PyTorch, TensorFlow). |
| Experiment Setup | Yes | For data augmentation, each frame has an independent probability to be grayscaled (p = 0.1) or solarized (p = 0.1). Each sequence has a probability of p = 0.5 to be horizontally flipped. For the straightening objective we set α = 15/9 and β = 1/9. ... We used the LARS [34] optimizer with learning rate 0.3, weight decay 1e-4, batch size 256 to train our straightening model. |