Access training data and a fully trained deep learning model
|
Data set = 5 million images
|
Infrastructure = Tensor flow
|
Model type = Inception V3
|
Neon Open's model architecture
ArchitectureModified Inception v3
Uses PReLUs instead of ReLUs for stability One auxiliary head for regularization1024-feature final layer, followed by softmax to get prediction |
TrainingTrain on NxNxD dense submatrices of full ranked comparisons matrix
Apply Laplacian smoothing across demographics Batch normalization Gradient clipping Modified RankNet loss |
Behavioral task used to collect training data
Data was collected on Amazon's Mechanical Turk from US participants over the age of 18. Age and gender information for each participant was collected.
|
Task
Flash triplets of images at subject
Locations of images randomized on screen Force subject to make a quick decision Images are repeated to ensure consistency Provides a weak sample that A > B, A > C Images sampled by scraping videos off the internet Trial images randomly chosen, weighted so that images with fewer labels are more likely to be surfaced in task |
Controls to minimize cheaters
Required training process for task
Internal inconsistency in the task is flagged Interacting too quickly is flagged Demographics for a subject are validated by asking for the demographics again a few days later |
Get model and code on GitHub
Selected papers
Lebrecht, S., Bar, M., Feldman Barrett, L., & Tarr, M.J. Micro-valences: Perceiving Affective Valence in Everyday Objects. Frontiers in Psychology 3. 2012: doi:10.3389/fpsyg.2012.00107
Lebrecht, S., & Tarr, M. J. Can Neural Signals for Visual Preference Predict Real-World Choices? BioScience. November 2012 / Vol. 62 No. 11
See more papers
Helpful resources
We find the following resources useful for people working in computer vision, video, and deep learning and so wanted to share!
Training data
ImageNet
Image Aesthetics: AVA
8M YouTube videos
Deep learning infrastructure
TensorFlow
Caffe
Computer vision toolkits
Open CV
ImageNet
Image Aesthetics: AVA
8M YouTube videos
Deep learning infrastructure
TensorFlow
Caffe
Computer vision toolkits
Open CV