You need a saturated black and white image

Hi there, I do not think this is a few-shot setting as you have just a single task. You need a distribution over many tasks. What you can do is to use fine-tuning, meaning pretraining your neural network on a larger (possibly related) set of data and then fine tune on your target data. See methods like Big Transfer for an idea of how to do it in practice.

Yes, MAML should be able to adapt to these new datasets as they are not so different from the (meta)training distribution. However, recent work showed that pretraining a ResNet on ImageNet it is a good starting point for adaptation to many different datasets (via dfine-tuning). Give a look at the BiT paper for more details ( arxiv.org/abs/1912.11370 ). If adaptation time is a crucial requirement in your work, then BiT may not be a good choice, and a hybrid approach may be better. Check out my recent paper about CaSE layers for additional details ( arxiv.org/abs/2206.09843 ).

It's not going to work very well because the two set of data are substantially different. You need to train on a larger variety of classes in order to generalize to medical images. Training on something like Meta-Dataset may be better.

There may be different solutions here. 1) You could use image augmentation to produce the query set. 2) Alternatively, you can train a backbone on a large dataset via supervised or self-supervised learning, then directly use that backbone at evaluation time as a prototypical network (remove the last linear layer and use the embeddings in the penultimate layer to build the prototypes).

Grazie Alessandro. Se vuoi utilizzare vettori come input invece che immagini, allora devi usare una rete feedforward (MLP) invece di una CNN. Il resto rimane uguale.

@@MaxPatacchiola Hai per caso qualche riferimento da cui posso prendere spunto?

Hi there, thank you for reaching out. There is no book or tutorial with subtitles. I was doing this in my spare time and now I am too busy to publish new material. At the moment I am offering private courses in 1-to-1 online sessions. I found this to be optimal for the student since the material can be adapted on the student's prior knowledge. If you are interested in something like this send me a message, you can find my contacts on my personal website: mpatacchiola.github.io

@@MaxPatacchiola I am from Iran. I am a master's student in artificial intelligence. My English is weak. I can't pay because my country is embargoed. Please introduce a video with English subtitles or a book like "Hands-On Deep Learning Algorithms with Python", there are few shot learning in this field. Thank you for your kindness. Unfortunately, there were limited resources.🙏🙏

It depends from the implementation. Typically there is a distance metric defined as a function used as part of a loss. The sign inside the distance function should agree with the minimisation objective in the loss. In other words, when the loss is minimised the distance between vectors of the same class should diminish. As a sanity check you can define two vectors and see if the distance diminish as they get closer to each other.

@@MaxPatacchiola Got it!!! Thanks a lot!

Glad you liked the video. I am not familiar with the particular application you are mentioning. I have worked mainly on image classification with RN. I suggest you to search on Google Scholar for papers like that.

@@MaxPatacchiola Thank you🙏

Grazie Nazareno, sono contento che i video ti siano stati di aiuto!

What do you mean with "one class"? A binary classification problem or a regression problem?

@@MaxPatacchiola binary classification problem. I want to use few-shot learning with stroke brain dataset (ISLES 2017). it includes only two classes (ischemic or not) with few samples. In few-shot learning, we train the model to recognize a novel class. In my case, I have only two classes in training and testing dataset. How can I use few-shot learning in this case. Thank you

two-ways many-shots

@@No_noifuw in this case there is no problem, you can apply few-shot learning methods. However, if you want to perform some sort of meta-learning training you need many small datasets from different distributions. The number of data points in a single dataset is small but you should have many of those datasets in order to learn something meaningful. Another solution in you case could be to train with standard supervised learning in a very large dataset that is close to your target data. Then in a second phase just fine-tune your model on the small dataset.

@@MaxPatacchiola Thank you so much I really appreciate your help

Exactly

@@MaxPatacchiola thank you so much! Your series of lecture on few shot learning is awesome!

Hi there! At test time yes, you can do the forward pass on a single image. You need a mini-batch when you are training the network, or part of it.

I mean, when I have pretrained model, how can i adjust it to detect certain objects, using maml, and only limited amount of examples?

Hi there, object detection is not the usual setting where MAML has been applied. You should check in the literature if there has been any adaptation in this direction. One solution could be to use the pretrained net as a backbone and then use MAML to adapt the last layer. Take this with a pinch of salt, I'm not familiar enough with object detection to give a more specific advice.

@@MaxPatacchiola do You mean the fully connected one?

@@MaxPatacchiola and what is the difference? In the video You say that we use specific model, which is my pretrained YOLO model and I cannot see why i can’t use it the way You explain. There is very little about it in literature so i can’t find any specific working implementations 😅

@@krzysztofmajchrzak1881 Yes you can apply MAML in the fully connected layer. It will be less flexible than applying it on the whole model but it should still work. The problem you may have on using it with YOLO is that you still need to find a way to define tasks (support + query) to train the last layer. You will need to adapt an object detection dataset to do that. As I said, I am not too familiar with the object detection literature and I do not know if there is anything you can use.