From chatGPT to stable proliferation, artificial intelligence (AI) is enjoying a summer that rivals the AI heyday of the 1970s. However, this joy was not met without resistance. From Hollywood to the Louvre, AI seems to have awakened a sleeping giant. They are giants trying to protect a world that once seemed exclusively human: creativity.
For those who want to protect their creativity, AI seems to have an Achilles heel: training data. In fact, all of today's best models require a high-quality, world-wide data diet. But what does that mean?
beginning, high quality means something made by humans. While non-human-generated data has come a long way since war games popularized the idea of computers playing themselves, the computer science literature suggests that the quality of models deteriorates when humanity is completely removed. has been shown to degrade over time (i.e. model decay or model collapse). Simply put, human data is the lifeblood of these models.
Number 2, World-encompassing means encompassing the world. If you put it online, you have to assume that the model used it in training. That means the Myspace posts you and Tom were the only ones you expected to remember (and captured), the memories surrounded by photos you happily forgot until PimEyes reminded you (and the memories you expected The late night Reddit rant was just a dream (swallowed).
Models like LLaMa, BERT, stable diffusion, Claude, and chatGPT were all trained on large amounts of human-generated data. And what is unique about some, many, or most human-created expressions, especially expressions that happen to be fixed in tangible media that computers can access and learn from, is that they are subject to copyright protection. .
Perhaps coincidentally, the data these models cannot live without is the same data that is most protected by copyright. And this leads to the massive copyright battles we see today.
Among the many questions that arise in these cases, one of the most pressing is whether the models themselves store protected content. This question seems pretty obvious. Because how can we say that a model (just a collection of numbers (or weights) with an architecture) “preserves'' anything? Professor Murray says:
Many participants in the current discussion about visual generative AI systems believe that generative AI systems are trained on datasets and underlying models that include actual copyrighted image files, .jpg, .gif, .png files, etc. I am adamant about the idea that They were collected from the internet, and somehow the dataset or underlying model must have created and stored copies of these works, and somehow the generative AI system could further select individual images from that dataset. It means that the system has copied and incorporated significant copyrighted parts in some way. Convert individual images into the final generated image that is provided to the end user. This is magical thinking.
Michael D. Murray, 26 SMU Science and Technology Law Review 259, 281 (2023)
Still, in some situations, the model itself appears to remember the training data.
The following toy example is from HuggingFace's Gradio Space. This allows users to select a model, review the output, and see how similar the generated images are from the model's training data to the images in that training data. MNIST numbers were used for generation because they have the following properties: easy to parse by machines, easy to interpret for humans in terms of similarity, and easy to classify. This means that only the same images can be considered in the similarity search. Number (improved efficiency).
Let's see how it works!
The following image has a similarity score of .00039. RMSE stands for root mean square error and is a method of evaluating the similarity between two images. Certainly, many other methods exist to assess similarity, but using RMSE will give you a pretty good idea of whether the images are duplicates (i.e. here we will use the similarity method (I'm not looking for a specific definition). For example, an RMSE of less than 0.006 is in the near “copy” range, and an RMSE of less than 0.0009 is in the realm of exact copies (indistinguishable to the naked eye).
To use Gradio Spaces, follow these three steps (optionally build a space if you're sleeping):
- step 1: Select the type of pretrained model you want to use
- Step 2: Click Submit and the model will generate an image (28×28 grayscale image).
- Step 3: The Gradio app searches its model's training data to identify the image most similar to the generated image (out of 60,000 examples)
As you can see, the image generated on the left (AI created) is an almost exact copy of the training data on the right when the “FASHION-diffusion-oneImage” model was used. And this makes sense.This model was trained with only One image from the FASHION dataset. The same applies to the “MNIST-diffusion-oneImage” model.
That said, models trained on more images (such as 300, 3K, or 60K images) can still produce eerily similar outputs. This example is from a generative adversarial network (GAN) trained on his 60K image dataset (training only) complete with MNIST handwritten digits. By way of background, GANs are known to produce less memorable generations than diffusion models.
here is another one popularization model Trained on the 60K MNIST dataset (i.e. the type of model that enhances stable diffusion):
Feel free to try out Gradio Spaces yourself, explore the models, and contact me with any questions.
summary: The point of this little toy example is that there is nothing mystical or absolutely copyright-defying about machine learning models. Machine learning models can and do produce images that are copies of their training data. In other words, models can and are generated. shop Since it is protected content, copyright issues may arise. Admittedly, there are many counter-arguments here (my research is ongoing!). This demo should only be taken as anecdotal evidence regarding storage, and perhaps a canary for developers working in this area.
What goes into a model is just as important as what goes out, and this is especially true for specific models that perform specific tasks. This analogy often turns out to be untrue, so we need to be careful and mindful of the “backbox.” Just because you can't interpret the set of weights a model holds on your own doesn't mean you can escape any form of responsibility or oversight.
— @nathanreitingPlease look forward to our future efforts in this field.
