It's possible... For some tasks (denoising, inpaining, high resolution,...) we can use the only sample for training a deep model. One of the best paper on this technology:

Article Deep Image Prior

And in addition there are some explonation materials on:

https://dmitryulyanov.github.io/deep_image_prior

But the key feature is that it's possible only for the narrow area of problems. But we are looking towards to a new research... maybe we can transfer deep neural prior methods in a wide range of problems in future.

Nice suggestions at all

You may have a look on one shot learning proposal by Fei Fei.

https://authors.library.caltech.edu/5407/1/LIFieeetpam06.pdf

The answer is NO.

Usually, we need a large number of data to train the DNN to make all the weights well trained. If the training sample is single, then the DNN will be underfit.

I agree with Xiliang Zhang sir and Christian Bauckhage sir . You can not do that

It's one of the biggest challenges in Deep Learning right now.

It always reminds me of Arch Learning by Winston. That was such a lovely idea, I hope we can implement it in the future.

The standard solution to this problem is called a matching network. A recent, well covered paper on this came out of Google Deepmind, Article Matching Networks for One Shot Learning

The primary difference between a matching network and a normal network is that, for each test sample, you compare it with a known example that you may or may not have shown the machine before. The DNN is then trained to answer the question, "Are these two the same or different?"

In simple terms, if I want to know what sort of animal is in my photo, I show it my photo alongside a photo of a dog, a cat, a horse and a lemur and I ask it in each case, "Is this a match?"

This can make the test process very long, as you have to iterate through at least one example of every possible answer. For this reason, a blocking stage is often employed, i.e. some means of limiting the number of possible answers is used. In the photo example, this might mean using a standard DNN to determine whether the photo is of a mammal, a bird, a reptile, an insect, etc.

The heart of most of these networks is a Siamese Neural Net. This architecture puts items, in my case, the photo that I want to identify and the possible match photo, through an identical neural net. It then, in some way, measures the distance between the outputs.

A pretty good introduction to all of this is here:

https://hackernoon.com/one-shot-learning-with-siamese-networks-in-pytorch-8ddaab10340e

A paper that I've found quite helpful in covering the range of options for matching networks is this one: Conference Paper Deep Learning for Entity Matching: A Design Space Exploration

A modern application of this sort of idea that is out in the field already is automatic facial recognition against passport data. So, in many airports, the check that you match your passport photo is automatic. That is not because a classifier has been trained on every face in the country. It is because the machine can whether your face is a match with the one in the photo.

As you can imagine (and the two papers cited here give a pretty strong sense of this) there are lots of variations on this basic idea but the heart of all of them is the same - it is taking a test sample, comparing it with a known sample, and asking, "Are these two a match?"

I hope that that helps.

Hello Wasseem,

As others have rightly pointed out, it's very much unlikely to succeed by training a deep neural network from scratch using one sample per class for the training set. That said, there are two options I foresee that may be useful to you as follows:

(1) You may want to try a pre-trained model. For example, there are several available models (VGG16, VGG19, ResNet, ResNeXt, DenseNet, etc.) that were pre-trained on the extremely large ImageNet dataset with about 1.2M samples and 1000 different classes. The hope is that the prior knowledge in these pre-trained models serve as a generic feature extractor and as such a strong regularizer for other tasks so that it's less probable to overfit. You may want to try these pre-trained models with or without fine-tuning. In the case that you decide to fine-tune the models using one sample per class which is an extreme scenario, I'll suggest that you consider/impose large weight decay penalty or dropout probability for regularization; you can also combine both!

(2) One-shot learning based on a Siamese network is another approach to consider. Here the nature of the overall cost function can be suitable for your scenario. Note that typcially few samples per class would be required for training. Again, building the Siamese network on pre-trained models may yield more interesting results as opposed to training the model from scratch.

I hope this helps :)

Best,

Oyebade

No chance, using actual methods.

There is an approach called as One Shot Learning. You can search for it, but one or the other way you need to have big dataset to get higher accuracy.

Machine learning is about designing models from data. If you don't have data, you don't have data. It's as simple as that. Reinventing pattern matching methods under new names won't help.

A problem with training data and pre-selected features is that there is no guarantee that they will be reflected in unseen data. A way around this might be to measure commonality between candidate images and single references. Some results are reported here: https://www.ee.ucl.ac.uk/~fstentif/EI2019%20paper.pdf

Actually in such cases when training data is scare, one should opt for transfer learning. In the approach, take a pre-trained model on some image dataset like imagenet and use that model as feature extractor- extract features at last layer of the model and then depending upon problem you have options like-

1. if you want to match test image in 1:1 mode against a sample registered in the database- compute distance between features extracted from registered image and test sample and then using a threshold verify if images are of same class or not.

2. if you want to identify class of test image through some classifier- then train a SVM like classifier over the features extracted (i.e. using deep model) from training data of target problem and use this classifier model to identify the class of features extracted for test sample.

It leads to underfitting of the network. But we can still create different variations of the same image like applying rotations, scaling... and feed them to the model in case of CNN but still will be a problem. I am not sure about the other networks.