Cognitive Science: An Introduction/Creativity, Planning, and Imagination

Cultural Evolution's Creativity
Cultures evolve over time. This evolution is independent of the genetic evolution that is simultaneously occurring in human beings over time. This cultural evolution results in surprisingly creative solutions to problems.

As we have seen in the previous section, individuals are capable of impressive creative acts. But it turns out that if these innovations are not shared with a social group, over time innovation stagnates. More minds catch more errors, recombine things in more ways, and even benefit from lucky accidents. If the learning from one generation to the next is cut off, due to war or some other disaster, the culture often loses innovations it might have enjoyed for centuries, such as how to make a kayak.

But what sense does it make to talk about a culture being creative independently of the individuals in it? Many cultural practices that are important for survival involve complex patterns of behavior (see the section on how to prepare manioc in the section on How Humans Got So Smart for an example), and the people in the culture often do not know what each step does. No one person came up with the behavioral pattern, and no one person might even understand it. So when we have a complex, effective innovation for something like food preparation, to what creative process do we give credit? The culture itself, over time, evolved the innovation. This is why we can think of a culture as being an entity that can be creative.

Genetic Evolution's Creativity
Just as cultural evolution can produce creative results, solutions to the problems species face can be solved by solutions arising in evolution that strike many of us as creative. An example of this is countershading.

Creativity and Artificial Intelligence
Looking back far enough, the history of artificial intelligence becomes synonymous with the history of computers. When Alan Turing invented The Bombe, a machine capable of deciphering the Enigma code used by the German army in World War Two, questions arose as to the potential for computers to demonstrate intelligent behavior. His popular ‘Turing Test’ is used as a benchmark to gauge the intelligence of AI programs even today.

The first AI computer program is widely accepted to be ELIZA, developed between 1964 and 1966 MIT. The program was a tool for natural language processing and capable of simulating a conversation with a human.

However, very few steps were taken before the end of the millennium to develop Machine Learning in AI programs, which is the basis for measuring any sort of creative capacity in contemporary Artificial Intelligence.

What is Creativity?
The Cambridge Dictionary defines “creativity” as ‘the ability to produce original and unusual ideas, or to make something new or imaginative. Boden claims creativity is a definitive quality of human intelligence generally. Avdeeff describes a marker for creativity as being the development of individual style and its manipulation.

The Neuroscience of Creativity
According to Dietrich, creativity is representative of cognitive flexibility, and exemplifies the ability to break patterns while maintaining sensibility. The neural circuits meant for processing information to perform non-creative tasks are the same as those that perform creative tasks. The prefrontal cortex is responsible for combining information in new ways and assessing whether the combination is meaningful or merely novel. Insight is the first step in the creative process, and the brain puts conscious effort to turn an insight into a useful idea. Creative thinking is based on two neural processing modes: deliberate and spontaneous. The spontaneous mode is beyond rationality, and is the popular way to think of creativity. An example of this would be the chemist Kekule discovering the structure of Benzene by dreaming of a snake in the shape of a ring. In this mode, insight is immediately useful and task-appropriate.

Conscious working memory does not select the combinations of information based on logical structures or beliefs and values and overlooks which products ‘slip out’ of the unconscious processing mode. In the deliberate processing mode, creativity is the result of rigorous conscious effort. Attention is provided to the combinations of information and therefore the insight conforms to external and internal structures, beliefs, values, and expectations. Common examples include Bach’s symphonies, Einstein’s theory of relativity, and the discovery of DNA, all of which involved carefully transforming insight into useful products.

The stochastic combinational process, as Dietrich describes, is a telling mental protocol which highlights the difference between human creative thought and artificial creative thought. When thoughts are unguided by attentional restrictions such as norms and logical structures, the combinational ideas and sequence of thoughts manifesting in consciousness are loosely connected by any coherent structure and chaotically combined together. These are generally ignored by the conscious working memory process and “slip through” when attention is less active. This suggests that creativity contains a high degree of unpredictability on the neurological level. It is possible this is the reason that true artificial creativity has been very difficult to achieve, and why the parameters for creativity are so difficult to quantify and transcribe algorithmically.

There are four basic types of creative insight :

1. Deliberate mode-cognitive structures

These are initiated by the prefrontal cortex, and almost always task-relevant. The process is characterized by structured responses between prefrontal brain and rear lobes.

2. Deliberate mode-emotional structures

These are initiated in the frontal attentional network and conform to personal values and beliefs, but are not necessarily domain-specific. Mostly, in the process, affective memory is retrieved from emotional structures rather than cognitive structures.

3. Spontaneous mode-cognitive structures

These types of insights arise from associative unconscious thinking. They are initiated in the rear lobes of the brain rather than the prefrontal cortex. The insight emerges without conscious effort into working memory and gains attention either on its own or as an augmentation of the information already in the attentive working memory.

4. Spontaneous mode-emotional structures

During this process, biological structures largely in control of emotional responses produce a “loud” signal that demands the attention of the conscious working memory. Without this signal, there is no way to gain access to the insight. The insights produced in this manner include mainly artistic expression, and create a profound emotional effect.

Making Artificial Intelligence Creative
According to Boden, AI can use three techniques to create new ideas: producing novel combinations of familiar ideas; exploring the potential of conceptual spaces; and transforming data in a way that enables the generation of previously impossible ideas. The first technique is widely associated with poetic imagery and forms of analogy, which computers are becoming increasingly more nimble with. The second technique is most commonly researched in computer models for intelligence. This technique involves defining parameters of the conceptual space first, which can be difficult even for professionals of a given field. Adequate knowledge of a conceptual space has to then be transcribed into a computable knowledge-base or semantic network before it can be explored. However, more emphasis is being put on the third technique in modern research. Most people would define creativity in this way across professions and disciplines. Boden also claims when creativity is acknowledged in jazz musicians or scientists for example, it is generally an acknowledgement of the individual’s capability to transform data to pave the way for new and previously impossible ideas.

The dominant discourse on AI believes in a version of the philosopher Laplace’s deterministic argument for intelligence. Laplace claimed that a truly intelligent being will be able to derive a formula for the universe which predicts its future and recounts its past based on the trajectories of every particle within it. The being will perfectly know everything about each aspect of the universe, and be able to recount a “story” of the universe from the beginning forward into infinity.

Approaches
An integral part of computational creativity research in artificial intelligence is recognizing that there are systematic patterns within human creative fields, which historically have been thought as unpredictable and non-orderly. These systematic patterns become the foundation on which artificial programs can produce novelties that are non-identical to items found in a computer’s database.

In addition, when considering combinational creativity, the AI may also produce new connections between items in its semantic network. This is a goal that characterizes creative AI research starting from the late 2000s. Before this, innovations in AI creativity were the production of novel items in a database using the same, pre-programmed connections, such as with ELIZA.

In most cases where creativity research in AI has succeeded, the AI program and the human programmer assist each other rather than either one developing a project or artifact alone. There are two major ways in which this complementary workstyle functions: the human programmer edits and adjusts the work produced by the AI software, referred to as content creation; and the AI software enhances and upscales a finished item produced by a human, referred to as content enhancement.

For Artificial Intelligence in general, Ong & Gupta outline five pillars of research: rationalizability, resilience; reproducibility; realism; and responsibility. Overall, these pillars are meant to support the growth of AI research across its different fields and applications. Many of the pillars are aimed at addressing the opaqueness of machine learning methods and to encourage the constructions of clear goals and strategies before the implementation of methods.

Narratives and scripts
‘Sunspring’ is a movie whose script is written almost entirely by an AI program called Benjamin. Launched in 2016, the movie is live-action and its post-production workflow was executed using AI-based decisions and monitoring. In 2016, IBM’s Watson platform created the first AI-produced movie trailer. The trailer was designed for 20th Century Fox’s horror movie, Morgan, and the project was led by the Manager of Multimedia and Vision at IBM Research, John Smith.

Games
In video game narratives, creative AI has made significantly faster progress than in most industries. AI Dungeon is a web-based game which generates storylines in real time and interacts with player input. The algorithm is trained by more than 10,000 label contributions to ensure smooth user interaction. Procedural generation randomizes the content to add variability and uniqueness to each game.

Visual art
Some attempts have been made in the content creation area of creative AI research in visual art.

The Next Rembrandt, a 3D printed painting made in 2016, was produced by AI software solely based on taking data from the famous Dutch painter Rembrandt’s portfolio and training using deep learning algorithms and facial recognition techniques. The first painting created solely by AI was auctioned in 2018.

Music
Musical creativity in modern AI research is based on a model called Long Short-Term Memory, which takes a transcribed musical idea as input and uses transformational creativity to manipulate it in meaningful ways. Several important innovations have occurred in the past decade which demonstrate the plausibility and quality of genuine AI-generated music.

Amper is the world’s first AI music composer, which released its beta software in 2014. It features a more human mood and style including optimizations that impose minute variations rather than strict and “robotic” adherence to time. This technique brings it closer to having a ‘ human spirit’. Software such as Amper is used mainly to speed up the production of royalty-free backdrop music, and is useful for content creation such as podcasts and YouTube videos. For this purpose, the music that AI creates does not need to be emotionally engaging or even musically interesting, it just creates an atmosphere that augments the primary focus of a certain piece of content. The music being produced by AI and used by human content creators can be considered a type of content creation on the part of the AI machine, but its primary function is content enhancement. Since this function affects the parameters and expectations for creative AI, it is appropriate to label it as content enhancement rather than creation.

A similar software was developed by Sony CSL Research Laboratory called Flow Machines. These are referred to as “augmented creativity” and cannot generate meaningful tracks on their own, but can help human creators enhance their own music and gain inspiration. Flow Machines differ from Amper in that they play an active role in the musical production and innovation process. In 2018, Flow Machines released their first album, ‘Hello World’, with compositional help by Benoit Carré, known as “SKYGGE”.

The audio uncanny valley
The closer AI-generated music gets to human-produced music, it enters into what software engineers have termed the “audio uncanny valley.” Music in this zone is characterized as unsettling to human ears and doesn’t seem “right,” in that it does not fit a standard expected musical piece composed by a human. The brain seems to reject the technically similar AI-produced track as authentic. The cause for this, Avdeeff claims, lies in Freud’s notion that people become disturbed if they observe a trace of a larger or unfamiliar creative force behind something.

Humans can immediately tell when music appears to lack a human authorship. At one of the first concerts of SKYGGE’s debut album, Hello World, in 2018, the appeal was largely contained in the design and mechanical traits of the production rather than any aesthetic value it may have had. Comments were mixed about the quality of the music itself, with several people saying how it “just doesn’t add up” or “doesn’t seem right”.

Images and animations
Most examples of creativity in AI to enhance or change images and video can be considered content enhancement, as the system is not making creative decisions but merely executing protocols based on vast input and specific goal-oriented operations.

Samsung AI takes an existing portrait and makes it look like the face is talking. Adobe has created the Character Animator software which enables lip synchronization, eye tracking and gesture control in real-time, by picking up input from the webcam and microphone. The software has been adopted by Hollywood and several online creators. Facebook’s Oculus Insight uses “SLAM” (meaning “simultaneous localization and mapping”) to generate maps and position tracking of the user in real-time.

The 2018 Peter Jackson film “‘They Shall Not Grow Old’ used an AI software to colorize 90 minutes worth of archived footage from World War One. The software took an input study database of WW1 equipment and uniforms as reference to convert the footage into a realistic colorized film. Other architectures, such as NesNet and DenseNet, have been effective at conversion from grayscale to natural-looking images with color.

Fictional Representations
Artificial Intelligence has been a part of fictional narratives for centuries. Starting with Mary Shelley’s Frankenstein, the artificially produced sentient or “intelligent” automaton has been a part of popular culture in science fiction primarily, and recently has made its way into other genres. A foundational work of science fiction which inspired some of the world’s leading AI researchers, is Isaac Asimov’s short story Runaround, in which two engineers develop an intelligent robot based on the Three Laws of Robotics. A recent example is the 2013 movie Her, directed by Spike Jonze. The movie is a romantic comedy-drama featuring a highly intelligent AI software with creative abilities which the protagonist falls in love with.

Most representations of AI in popular fiction depict creative AI. Characters such as Morgan and Frankenstein’s monster demonstrate traits that would fit one or more of Boden’s types of creativity. They are not merely observant in a sophisticated way, but can evaluate their observations for the purpose of producing non-trivial actions based on a flexibility of cognitive boundaries. The reason for this may be that creative AI carries more entertainment value than AI that functions purely as a virtual assistant without any novel and productive capacity. Tony Stark’s JARVIS from the Iron Man and Avengers movies from Marvel Studios demonstrates a creative capacity of producing novel combinations of data, such as finding an optimum landing location for Stark when his suit malfunctions mid-flight, and providing relevant creative information which is adaptive to changing parameters, such as dissecting a model of the Möbius strip to help Stark invent time travel.

Challenges
At present, AI is most valuable at the content enhancement level, where human efficiency is low and post-production tasks require increasingly more labor. Creative AI can help in enhancing and restoring data from acquisition as well as be trained to upscale data from relatively low input.

State-of-the-art Machine Learning techniques have been designed to make use of recursive encoding-decoding leading to reiterative content enhancement, but the same techniques rarely translate into the content creation domains for creative AI research.

Ethics
Currently, AI performs with intelligence but not an awareness of the wider context. Artificial intelligence has not yet been programmed to balance the ethical factors surrounding the execution of tasks. Deepfakes, which are becoming increasingly higher quality, can be used for malicious purposes, and AI used in visual art and music can endanger artists and designers to exploitation and fraud. According to Anantrasirichai & Bull, future considerations regarding the development of creative AI techniques should include ethical consideration of wider contextual implications as a parameter.

The three categories for relating ethics to AI, classified by Dignum, are: Ethics by Design, the practice of including ethics into the programming methods; ethics in Design, methods that assess the ethical implications of AI systems within the architecture; and ethics for Design, consisting of protocols which consider and ensure the protection of artist and developer rights.