The pioneers of computer science would be shocked by today`s AI
Ada Lovelace wrote that a machine “has no pretensions to originate anything. It can only do whatever we know how to order it to perform.” To an extent, this is as true now as it was 200 years ago; Humans can create and perceive patterns from (apparent or real) randomness, because humans, for better or worse, see patterns everywhere. But computers can only work in existing patterns; it’s actually impossible for a computer to internally create truly random — and therefore truly original — data.
However, the news is continually awash with stories of AI creating original art, poetry, and even paintings. So what’s going on here? Has AI advanced to the stage where it can be called truly “creative”? And how might we put it to practical use?
To discover the part of AI creative frontiers, I would like to utilize free machine learning accessible software “RunwayML” to generate an example GAN (Generative Adversarial Networks) model. The process of generating material is used to discuss if AI can enhance our creativity and what will it bring to the future of human creativity.
What is creativity?
Before we dive into the GAN model generation process, I would like to make it clear what is our measure of creativity. Due to the apparent “accidental” nature of idea generation, creativity is often treated as unpredictable. However, science has been researching the processes leading to creativity and offered a neuroscientific explanation for one of them. In fact, there are three, distinguished by the sorts of psychological, and cognitive processes involved in idea generation for each of them. They are combinational, exploratory, and transformational creativity.
- Combinational: new combinations of familiar ideas
- Exploratory: generation of new ideas by exploration of a space of concepts
- Transformational: involves a transformation of the search space so new kinds of ideas can be generated.
If AI were considered “creative” out of these types of creativity presented, it would most likely be categorized as combinational creativity as AI requires inputs to generate “creation.”
Gan model theme
Over the last decade, we have seen the online community grow rapidly in importance. In fact, more than 2 billion people actively use Youtube— that’s 1/3 of the world’s population.
Inevitably these people look up to influencers on Youtubers to guide them with their decision-making. Of course, the same is true for…
- Video game streamers
- Social media influencers
- Youtubers
Influencers in the online community are people who have built a reputation for their knowledge and expertise on a specific topic. They make regular posts about that topic on their preferred channels and generate large followings of enthusiastic, engaged people who pay close attention to their views. Brands love influencers because they can create trends and encourage their followers to buy the products they promote.
However, as the number of online community participants increases, content consumption cycles get faster due to competition and accustomedness. Therefore, influencers need an environment for rapid content creation such as…
- Quik posting appealing images
- Efficient video editing
In this paper, I would like to focus on creating a system that generates various youtube thumbnails in just one click.
Generating Gan model
By using Runway, you will be able to quickly discover new models without using any code. In this project, I would like to use a 500+ image dataset from two different methods, Youtube -8M, and Youtube Data API v3.
- Youtube -8M: The YouTube-8M Segments dataset is an extension of the YouTube-8M dataset with human-verified segment annotations
- Youtube Data API v3: Use the API to search for videos matching specific search terms, topics, locations, publication dates, and much more. The APIs search. list method also supports searches for playlists and channels
The main reason why I decided to use these methods is that they both complement each other. Youtube -8M offers a crude video image dataset by using a simple search engine. On the other hand, Youtube Data API v3 provides detailed thumbnail data with various conditions. This range of choices satisfies both beginners and experienced programmers.
GAN training structure and result
During each Training Step, the model learns a little bit more about the visual patterns or features, it finds in your dataset. Determining an effective number of training steps is a bit of a moving target. Too few and your model might not be able to synthesize images with coherent visual forms that are reminiscent of those in the original dataset. Too many and your model will stop improving, not learning anything new, and possibly produce worse images or images that look the same. The maximum number of Training Steps is 25,000, but a good starting point is 3,000. During and after training, you can review the model’s learning progress at various steps. For this case, 3,500 steps were taken, which took approximately four hours to generate GAN models. However, the resulting image turned out to be horrifying and generated an image that has little vibes of people and a speech bubble.
Current limitations and future challenges
- Lack of dataset (sorting)
- Difficulty balancing generator loss and discriminator loss
- Time/computational cost of handling image data on the device
- Processing time
Is AI creative?
It can be implied that the AI process is truly ingenious by understanding the GAN model training process and the philosophy behind Boden’s definition of ingeniousness. Combinatorial inventiveness is the act of using well-known concepts (pictures entered) to create something new (the models). In addition to combinational ingenuity, if AI is allowed to continue creating models, there is a much greater chance that it will create content with transformational ingenuity. Is AI intelligent if Boden’s definition of cleverness was not one of them? Many would contend that it is not inventive because AI follows a predetermined pattern and does not make new discoveries outside of those confines. They might counter that AI has the ability to affect inventiveness.
Last thought
While serving as inspiration is one role AI can play in the creative process, it can also help with more mundane tasks, especially in the digital domain where much of the behind-the-scenes work is far from glamorous.
I personally think that it is about the augmentation of creativity. In the end, the human really is the one being creative, and it’s more about how can you get better efficiencies. “With filmmaking, 99% of the work is actually very mundane. It’s going through hundreds of hours of video in some cases to arrive at the core pieces to use. So there’s still a very good reason to use technology as an assistant here, rather than replace the human in the loop.”
AI techniques, from machine learning to pattern recognition, have already proved ideal in providing assistance in virtually every industry. Healthcare, finance, and retail are just a few that are reaping the benefits of advanced cognition capabilities. There’s no doubt the boundaries of AI’s role in creative endeavors will be pushed. And while it will never replace the human soul of creativity, AI can certainly offer many benefits by serving as a smart, efficient and inspirational assistant.
References:
{1} https://influencermarketinghub.com/what-is-an-influencer/
{2} https://www.kaggle.com/datasets/praneshmukhopadhyay/youtube-thumbnail-dataset
{4} https://www.youtube.com/watch?v=gvNdCRe3T-g
{5} https://github.com/lucidrains/denoising-diffusion-pytorch