3 ms·

This made me think of something, so bear with me for a second. I took Andrew Ng's Machine Learning class on Coursera, and one of the things he talked about was

This made me think of something, so bear with me for a second. I took Andrew Ng's Machine Learning class on Coursera, and one of the things he talked about was how, given a certain data set and a logistic regression approach, the algorithm might decide to just always give it a value of 1, since 98% of the time the value is one. The examples given here were all predicted to have 95%+ similarity, but it seems to me that given this dataset, I could be reasonably convinced that it just assumes all images are reasonably similar. In other words, the author provides no counter-examples that demonstrate that the algorithm can detect dissimilar images, such as a photo of a car and a clipart of pizza. I would expect that to have very low similarity, and given the algorithm's demonstrated ability to discern between similar and dissimilar images, I would be more convinced of perceptual image hashing as a viable technique.

667 12y ago
It works, at least the examples above work fairly well-- I've tested a few for an automated image processing app I'm working on.
andrea_s 12y ago
While your comment applies on a broader sense (we need to test both for cases in which we expect the algorithm to find similarity and cases in which we expect to find dissimilarity - lest we embrace confirmation bias), I believe your reference to machine learning is a bit off. In this case, we are not looking at a model training, we are just evaluating the outcome of a (static) algorithm. On the other hand, in the context of ML you should try to include as much information as possible in your training dataset. This is a concept that comes by in many fields, some of them just tangentially related to machine learning (e.g. excitation signals for system identification, speaking both in terms of frequency domain and differential equations). tl;dr: there is no "training set" to speak in the presented article - hence, concepts such as overfitting and information content (of the training set) do not quite apply here.
- Jemaclus 12y ago
  I wasn't making a direct comparison between machine learning and this particular task. The article just made me think of that machine learning problem, and I would like to see more examples that show a wide range of results. For instance, there are no examples of false positives -- how often would that happen in an image processor? Probably a lot, but I wouldn't know it based on this particular article.

667 12y ago

It works, at least the examples above work fairly well-- I've tested a few for an automated image processing app I'm working on.

andrea_s 12y ago

While your comment applies on a broader sense (we need to test both for cases in which we expect the algorithm to find similarity and cases in which we expect to find dissimilarity - lest we embrace confirmation bias), I believe your reference to machine learning is a bit off. In this case, we are not looking at a model training, we are just evaluating the outcome of a (static) algorithm. On the other hand, in the context of ML you should try to include as much information as possible in your training dataset. This is a concept that comes by in many fields, some of them just tangentially related to machine learning (e.g. excitation signals for system identification, speaking both in terms of frequency domain and differential equations). tl;dr: there is no "training set" to speak in the presented article - hence, concepts such as overfitting and information content (of the training set) do not quite apply here.

Jemaclus 12y ago
I wasn't making a direct comparison between machine learning and this particular task. The article just made me think of that machine learning problem, and I would like to see more examples that show a wide range of results. For instance, there are no examples of false positives -- how often would that happen in an image processor? Probably a lot, but I wouldn't know it based on this particular article.