These titles and abstracts have been generated by a fine tuned GPT-2 model. The model has 345M parameters and has been trained on ~30k papers from arxiv
The results are generated at a temperature=0.7
The model has learnt a lot of contextual meaning from the abstracts and the generated titles and abstracts are really good for a training time of just 40 minutes.
This text has been completely thought up by a trained model and not written by a human!
Unsupervised Temporal Compressive Imaging: A New Method for Classification
Temporal Compressive Imaging (TcI) is a popular method for deep convolutional neural networks (CNN) for image classification. TcI uses a convolutional
transform to estimate the temporal sequence of a CNN. This context-sensitive recall mechanism is coupled with a "mind-body-mind" model, which is a
powerful framework for visual-spatial-spatial navigation. We show that TcI can be used for image classification in an unsupervised fashion. First,
we show that, with the help of a convolutional-transform, the temporal sequence is obtained by the convolutional-transform. Then, we introduce a TcI
classifier to explain the temporal sequence obtained by the convolutional transform with respect to the convolutional-transform. These results are
interpretable to further classify the images. They show that TcI can be used for both low-level and high-level learning tasks. We also use TcI for
image classification in the context of the epileptic person. We validate our methods on a large dataset of images and show that they are able to solve
the tasks of the epileptic person.
- Train on ~30k papers
- Deploy on website as POC
- Train on more papers
- Upload fine-tuned weights
- A more efficient way to store and display generated samples
This is just for fun. I have no intention of causing harm or infringing on anyone's copyright.