Unsupervised Feature Learning via Non-Parametric Instance Discrimination

Inspiration

---

• The top-5 classification error is significantly lower than top-1 error
• second highest responding class in the softmax output to an image is more likely to be visually correlated

---

Apparent Similarity is learned not from semantic annotations, but from visual data themselves

---

Instance-level discrimination

---

• An image is distinctive in its own right, and each could differ significantly from other images in the same semantic category
• Major challenge - The number of classes is the size of the entire training set

---

Method

1. Non-Parametric Softmax Classifier

• {Wj}: weight vector for class j, {Vj}: embedded feature for all the images j
• use {Vj} instaded of {Wj}
• {Vj} for all the images are needed --> maintain a feature memory bank for storing them

2. Noise-Contrastive Estimation (NCE)

• Computing the non-parametric softmax classifier for every instance is costly
• cast multi-class classification into a set of binary classification problem (positive/negative)
  • discriminate between data samples and noise samples
  • data samples (positive samples):  past embedded feature for data i in memory bank
  • noise samples (negative samples): randomly sampled embedded data for other data in memory bank

3. [Test] Weighted k-Nearest Neighbor Classifier

• compute the embedded feature of test image
• compare it against the embeddings of all the images in the memory bank using cosine similarity
• top k nearest neightbors would then ben used to make the prediction via weighted voting (k=200)

---

Experiment

1. Setting

• backbone network (AlexNet, VGG16, Resnet18, Resnet50) and linear classification (SVM)
• Dataset: ImageNet ILSVRC

2. Evaluate Performance

• Perform linear SVM on the intermediate features from conv1 to conv5
• Perform KNN on the output features

 

 

 

 

Reference

https://velog.io/@dongdori/InfoNCE-Metric-Learning

https://www.youtube.com/watch?v=wyBzB9iRveI