This occurs when the model incorrectly classifies healthy individuals as having the disease. While this leads to patient anxiety, unnecessary follow-up tests, and increased costs, it is generally considered less harmful than failing to detect an actual disease case (a false negative).
For critical applications like early cancer detection, low sensitivity is unacceptable, as missing real cases can lead to delayed treatment and potentially fatal outcomes.
On the other hand, low precision is characterized by a high number of false positives (FP).
'Of all the people with the disease, what fraction did the model correctly detect?' High sensitivity corresponds to a low number of false negatives (FN), meaning the model is less likely to miss a diagnosis.
🧵 Sensitivity vs Precision in Medical AI
-----------------------------------------
In the field of medical AI, sensitivity is a crucial performance measure. It indicates how effectively the model identifies individuals who actually have a specific disease. Essentially, it answers the question:
At first, all entries of this matrix are set to zero. Then, each prediction is compared with its true label, and their corresponding entry is incremented by one. Confusion matrix entries are used to calculate counting metrics like TP, FP, FN, and TN.
github.com/GoktugGuverc...
It shows how many times the model predicted each class compared to ground truth labels. While the rows represent the predictions, the columns refer to ground-truth labels. Each row and column are defined to denote a specific class index.
🧵Confusion Matrix:
--------------------
Classification models classify the samples into predefined categories. To evaluate how well these classifiers perform, we use confusion matrix. It is actually tabular way of visualizing and assessing model performance.
SmolVLM paper is out and it's packed with great findings on training a good smol vision LM!
Andi summarized them below, give it a read if you want to see more insights 🤠
Yıllarca github.com/acikkaynak ‘ta projeleri, sonra github.com/eser/topluluklar ‘la insanları bir araya getirme alışkanlığımla şimdi de BlueSky’da hesapları bir araya getirme inisiyatifi başlattım 😅 ben aklıma gelen isimleri ekledim. sizler de “beni de ekle” yazarsanız eklerim.
go.bsky.app/TLtndiT
Beni de ekleyebilirsiniz :)
Ben de benzer bir karar aldım. Orada daha fazla community olmasına, ve oranın daha çok kullanılmasına rağmen dediğin üzere uyguladığı son politikalarla iyice X’ten bir soğuma geldi ve hani ders de almıyor o kadar şikayet edilmesine rağmen.
Recurrent Neural Networks vs Transformers:
------------------------------------------------
- Core M...
Read more: https://longer.blue/posts/4FuiiTEfIK
Heyy, hoşgeldin
Cancer Research
============
I am proceeding cancer research repository to illuminate cancer drugs, structure of proteins, and protein-protein interaction networks. The following sections are ready for usage:
- 36 Different Cancer Drugs
- Protein Database with UniProt
github.com/GoktugGuverc...
Image is from https://blog.roboflow.com/yolov7-breakdown/
🧵YOLO (You Only Look Once) was introduced by Joseph Redmon et al. in 2015 and has since undergone several iterations. The core idea behind YOLO is to perform object detection by treating it as a regression problem, where bounding box coordinates and class probabilities.
🧵 Understanding Vision Transformers
Before we jump into ViTs, let's quickly recap the basics. In traditional computer vision, Convolutional Neural Networks (CNNs) have been ruling the roost for image-related tasks.
🧵The Jacobian Matrix and Backpropagation
The Jacobian matrix represents the partial derivatives of a vector-valued function with respect to its input variables. In deep learning, it helps us compute the gradients of the loss function with respect to the network's parameters.
In that case, embedding vectors generated for different images
become constant. This is called dimensional collapse.
In the general concept of this process, random transformation techniques are applied to the images to come up with positive matching pairs of similar images. At that point, the representations and features extracted by the network may collapse into one single point.
Self-supervised learning techniques construct and learn an embedding space by minimizing distances between embedding vectors of input images or just clustering them in a multidimensional graph.
Compared with CNNs, transformers are
computationally more expensive and require much more data for successful training. Besides, transformer features do not tend to exhibit distinct properties.
Possible alternative to convolutional neural networks for visual recognition task is transformers. The strategy inspired by natural language processing to train visual transformers (ViT) is to pretrain on large amount of data, and finetune on target dataset.
Hafiften buraya geçmeye başladım, twitterdaki şu kısıtlamalar kabak tadı vermeye başladı
Asıl sıkıntılı olan durum bence şu. Takip edilen birçok bilimsel içerik ve account burada yok. Misal, google-health’in publish ettiği makalelere bakıyorum twitter’da ama bunu burada yapmam mümkün değil. Hala invitation olayı, bi tık sıkıntı
