Reproduce “Collaborative Feature Learning from Social Media” on OpenResearch

About this paper

Chen Fang, Hailin Jin, Jianchao Yang, Zhe Lin

Researchers from Dartmouth College, Adobe Research, and Snapchat introduce "collaborative feature learning," a novel method for obtaining image feature representations without explicit category labels. This approach leverages implicit user behavior data from social media to train a Deep Convolutional Neural Network (DCNN), yielding features that outperform ImageNet features in image similarity and style classification, and achieve competitive results in object classification on benchmark datasets.

Key insights

• Implicit user behavior data, such as viewing patterns on social media, can effectively serve as a label-free source of supervision for learning robust and meaningful image features.
• Latent factors derived from collaborative filtering capture rich high-level visual and semantic information about images, discerning both coarse categorical and fine-grained stylistic nuances.
• Converting continuous latent factors into discrete pseudo classes via clustering offers a more stable and efficient training paradigm for DCNNs compared to directly regressing to continuous latent values.

Summary by alphaXiv.