MIT researchers developed a training method that teaches vision-language generative AI models to localize a specific object, like a person’s pet, in a new scene.

Single-photon sources are key components of quantum communication technologies. However, conventional designs use decoupled single-photon emitters and photon transmission methods, resulting in high transmission loss, limiting practical applicability. Now, researchers from Japan have developed a new method, where a single rare-earth ion is used to generate and guide single photons directly within an optical fiber at room temperature. It is low cost and can become a key component of upcoming quantum communication technologies.

Conventionally, deep neural networks (DNNs), including convolutional neural networks (CNNs), are trained using backpropagation—a standard algorithm in AI learning. However, backpropagation suffers from several limitations, such as high computational cost and overfitting. Researchers have now developed a new training approach called the Visual Forward–Forward Network (VFF-Net), which overcomes these challenges. By eliminating the need for backpropagation, VFF-Net enables more efficient, less resource-intensive training while maintaining high accuracy and robustness.