Episode SummaryWhen we think about AI, we often think about optimizing some particular task. In most circumstances through computation there is an optimal chess move, or an optimal way to determine pattern in data, or solve a math problem, or route info through servers. Most of us are aware of these uses, but what about creative tasks? Can these also be optimized? If we want to give a computer information and tell it to create powerpoint slides, is there an optimal way to create such slides? Dr. Philippe Pasquier’s computational research is focused on artificial creativity. In this episode, we talk about how to define a very new field, train machines in this area, and also discuss trends and developments that might permit such technology to thrive in the next 10 years.

View Companies Related to Your Industry

Explore companies, products, and service providers on TechEmergence:

GuestDr. Phillipe Pasquier

Expertise: Artificial Intelligence and Cognitive Sciences, Computational Creativity

Recognition in BriefAfter studying computer science and cognitive sciences in Europe (France and Belgium), Philippe Pasquier completed his PhD. in the field of artificial intelligence at Laval University (Québec, Canada). He has been working as a postdoctoral research fellow in the Department of Information Systems at the University of Melbourne (Australia). Since January 2008, he has been an assistant professor in the School of Interactive Arts and Technology (SIAT) of Simon Fraser University’s Faculty of Applied Sciences (Vancouver, Canada), where he is conducting both a scientific and an artistic research agenda.

Current AffiliationsSimon Fraser University, Founder of the ACM Movement and Computation, Chair of the Musical Metacreation Workshops, CEO of Metacreative Technologies, and Advisor of Generate Inc.

Can Machines be Creative?

Most of us can understand, at least in simple form, the meaning of computation and its role in artificial intelligence. But what about creative computation aka artificial creativity, how would you define that concept? There’s still debate about how to define this embryonic subfield of more traditional artificial intelligence, and it’s one preoccupation of Dr. Philippe Pasquier. Well, more than a preoccupation really; mucho f Pasquier’s research explores and makes use of the ideas growing out of creative artificial intelligence.

“It’s looking at exploring automation of creative processes, it’s called sometimes computational creativity, sometimes it’s called artificial creativity, but the idea always is try to endow machines with creative behaviors. As a field it investigates both creativity as it is, can we understand creativity, can we model it…then it also inquires creativity as it could be.” In other words, Pasquier and his team think about whether they can come up with processes that human beings are incapable of achieving at present, but that machines could potentially be capable of producing through their own creative powers.

In contrast to the accepted role of AI as a goal-oriented process, the notion of creativity is not associated with a set outcome. Tasks, such as choreography as one example, often end up in a spot unassociated with the beginning stages of the planning process. As Philippe noted, “There is not (even) an optimal way to serve a beer.”

Many creative fields, including musical composition and interpretation, video game creation, drawing, painting, writing new narratives and poetry, and design-related tasks, seem relegated, in most people’s minds, to the forever domain of human beings. But many of these “softer” domains are beginning to be opened up to manipulation by AI.

Pasquier’s lab experiments with using AI for visual art and movement generation, but their specialty is in using AI for music composition and interpretation. “We develop agents that do those tasks, and then we compare the rhythm between each other’s but we also compare the performance of the agents with human composers, and then we…try to find out if people can tell – it’s a bit like a musical Turing test if you will – the difference between machine-generated music and human-written music,” says Philippe.

AI started by looking at rational problems, likely because they’re more easily defined, said Philippe. With the work done in his lab and elsewhere, researchers are starting to discover new uses for AI in creative tasks, particularly in the entertainment industry, and deciphering approaches and algorithms that can be used in conjunction with these tasks.

But creativity is a complex process that humans don’t understand fully, so it’s only natural to wonder how we can expect AI to exhibit creative responses if we ourselves don’t understand its ins and outs. Then again, the same thing can be said for how our minds – and sophisticated algorithms – process information (echoed by Google’s Demis Hassabis in the U.K.’s The Financial Times). Although not fully understood, in principle it may be possible to employ creativity in some form as a machine-learning process.

AI as Creative Collaborators

Like the unprescribed flow of creative acts, learning AI’s creative potentialities starts with experimentation. There are a variety of software programs that can be used to produce visual art and music in the style of an existing artist, an approach called style imitation  and akin to Google’s DeepDream software.“Like humans, (machines) can be creative but they need to be exposed to a lot of (artists) before they get any good at it; like a human composer couldn’t compose if it never listened to any music, for machines it works in very similar ways,” says Pasquier. Systems that emulate style movements in animation, particularly video game animation, are also ramping up and increasing in quality, paving the way for further use in the industry.

In a very real sense, these new machine-learning tools save humans time and allow them to use their creative energies in other facets of a project. For example, motion capture animation is very time-consuming and expensive. Researchers are “teaching” machines the movement styles of human beings, and then getting those same machines to emulate moves. For example, Pasquier mentioned a famous, “anonymous” actor who comes into their studio every so often. During this time, the actor’s body movements are captured on camera. The machine receives this input, then learns to emulate those moves. At some point, the machine can go beyond it’s initial training to generate new movements that were not directly captured, but are of a similar style to the actor’s, all done without the use of human animators.

Of course, creativity isn’t limited to the visual- or design-based arts. If teaching machines to be creative is possible, it seems there could potentially be a wide variety of applications for a diverse set of tasks. How versatile is “creative AI”, and will it ever be used outside of the visual arts, music, and gaming industries?

“We focus on the arts because it’s a niche domain, we generate systems and we can show that there’s no ultimate solution and we can really test the algorithms in those domains, but creativity is everywhere, so the boundary of the field, it’s out to be redefined,” says Phillipe. He also acknowledges that there are several contexts in which rational thinking and creative thinking are used at the same time, such as in creative mathematics.

In other instances, some people do well in negotiation not because they know game theory better, but because they have mastered a nuance of style – a form of creativity – that allows them to more effectively communicate and express their ideas. “We have a lot to learn at that level, it usually doesn’t fit in the type of algorithm and  approach that AI has, and the more we spend time tracking these computers, the more…their interactivity improves…the more important it is that they get to recognize, represents, and eventually manipulate all those subtleties that are very characteristic of humans,” he says.

This weaves into the type of creativity that computers are capable of exhibiting. “When it comes to low-level creativity – everyone is creative, it’s not always considered as being a super-genuis style of creativity, it’s a lower type of creativity, but it’s really where we are right now with computational creativity, and there’s other types of creativity, like transformational creativity and combinatorial creativity for computers to exhibit,” says Phillipe. To date, Pasquier is not aware of a machine that has, for example, created an entirely new type of music of which humans were not aware – disrupting the current paradigm, if you will. Humans are still needed to maintain and transform human culture.

Over the next decade or so, Phillipe believes there’s a bright future for increased collaboration between creative AI and human beings in many “softer” applications.

For example, when you use composition or drawing software or video editing software today, you often have in mind – at least to some extent – a vision of the finished product. However, parts of the process might take days or weeks, even though you can tell the computer, through various demands and manipulation techniques, what you want it to do.

In the lab, Phillip had his team tried to validate this process through purely computer-generated outcomes, using a process similar to pedagogical theories. The machine is given copies of a particular creative product (model), then the team compares the computer’s final product with that of human beings (assess), and finally the systems is told whether the final product is “bad or good” (feedback) before it tries the process all over again (repetition).

“It’s important for humans to keep creativity and be able to drive the system, but right now there’s no real computer-assisted creativity tools, you have to do all the pieces from scratch; because creative computing is new, with the vast majority used for entertainment, the market is still shifting,” says Phillipe. His lab is currently working with companies to find software that can function somewhere in-between the two extremes of purely reactive with no autonomy versus completely autonomous without human input.

Image credit: Simon Fraser University