Data Science and Machine Learning
Until recently, everyone was delighted with convolutional networks in the tasks of "understanding" images, the work of transformers, such as BERT and GPT, in the tasks of understanding and generating text, GANs in the tasks of generating deepfake photos and videos.
Imagine a world in which you no longer need to train neural networks. If you think "fit predict" sounds tragic, then "zero shot" will just shoot them all.
AI researchers at OpenAI have created two neural networks. The first, called DALL · E, which generates images from natural language textual descriptions. Second, CLIP is an image classifier that doesn't need to be trained!
- Zero shot: Is it possible to train your high-precision classifier for your task without training on any image at all?
- Generating an image from a text description: How to get a cool logo and not pay? How well does Artificial Intelligence understand the text, how well does it “draw”?
- Neural network trends of the future. What to expect?
