In order to maximize creative behavior, humans and computers need to collaborate in a manner that will leverage the strengths of both. A 2017 mathematical proof shows two limits to how innovative a computer can be. Humans can help counteract these demonstrated limits. Humans possess many mental blind spots to innovating (e.g., functional fixedness, design fixation, analogy blindness, etc.), and particular algorithms can help counteract these shortcomings. Further, since humans produce the corpora used by AI technology, human blind spots to innovation are implicit within the text processed by AI technology. Known algorithms that query humans in particular ways can effectively counter these text-based blind spots. Working together, a human-computer partnership can achieve higher degrees of innovation than either working alone. To become an effective partnership, however, a special interface is needed that is both human- and computer-friendly. This interface called BrainSwarming possesses a linguistic component, which is a formal grammar that is also natural for humans to use and a visual component that is easily represented by standard data structures. Further, the interface breaks down innovative problem solving into its essential components: a goal, sub-goals, resources, features, interactions, and effects. The resulting human-AI synergy has the potential to achieve innovative breakthroughs that either partner working alone may never achieve.
- human-computer interface
- artificial intelligence
- intelligence augmentation
Recent critiques of IBM Watson in the business world (
More generally, any computational approach to innovation and creativity (e.g., Machine Learning, Deep Learning, AI in general) has limits to how creative or innovative it can be. The 2017 mathematical proof details two of these limitations . Humans can help counter these limits. On the other hand, humans have many known mental blind spots to innovation, including
From these findings, it makes sense to create a human-computer interface for innovation that is both human- and computer-friendly so that the computer can help humans be more innovative and humans can return the favor for the computer. The overall result thus far has been a human-computer partnership that has already found novel solutions to such tough problems as how to significantly reduce concussions in American football players and how to adhere a coating to the non-stick surface Teflon [11, 12]. This human-computer synergy has the potential to achieve even greater innovative breakthroughs.
This chapter first articulates new definitions of creativity, innovation, feature, and effect. These definitions permit quantified arguments about the innovation process. Next, the main points of the proof will be presented. All the proof’s details are contained in Ref. . The main conclusion is that no computational approach can fully take over the creative or innovative process.
Then, several of the weaknesses to human innovation will be presented along with their effective algorithmic counter-techniques. A full description of human weaknesses and programmable counter-techniques are contained in Ref. . How these human weaknesses become computer weaknesses is explained with an emphasis on how the programmable counter-techniques can also improve the innovation of any AI technology. Finally, the human-computer interface that permits humans to counter computer limits and the computer to counter human weaknesses will be presented. This interface called
2. Proven computer limits to innovation
Section 2.1 articulates the new definitions for creativity/innovation, feature, and effect, which then permit the quantification of the size of the space of innovation for physical objects. The space of innovation for a given object is shown to consist of all possible effects that the object could produce when interacting with every other possible object, material, force, energy, and condition (e.g., barometric pressure and gravity strength). Section 2.2 quantifies the number of interactions that are possible between an object of interest and all other objects, materials, forces, and energies in the world. Section 2.3 builds on Section 2.2 by exploring all the ways that two given objects could interact within various conditions to produce interesting effects. The number of possible interactions and possible effects is so astronomically large that the fastest supercomputer today could not examine them even it started working from the invention of the first computer. Section 2.4 articulates other reasons why a computer could not predict certain effects of an interaction. Finally, Section 2.5 shows how humans can help counteract the challenges that computers have when innovating.
2.1. New definitions
Any creative/innovative solution to a problem is built upon at least one commonly overlooked or new feature of the problem. A feature that is commonly overlooked or new is called obscure. The above description is called the
If the solution was based upon a commonly noticed feature, then it would get a low rating on originality and a high rating on obviousness . For example, if a scented jar candle company came out with a new scent called
Given that features are a crucial aspect of creativity, a definition adapted from the philosopher Nietzsche permits the number of features of an object to be quantified . Nietzsche states: “The features of a ‘thing’ are its effects on other ‘things’: if one removes other ‘things,’ then a thing has no features” . From this perspective, every feature emerges from interactions and is not intrinsic to the object itself. Certainly, color is not intrinsic to an object, but results from light interacting with the object and our retinas, which results in processing in the human visual cortex. Change the circumstances of the interaction and the color changes. Change the lighting. Put on sunglasses. Experience trauma in the visual cortex. These and other changes can result in a change in color.
As a further example, the mass of an object now appears to be the result of the object interacting with Higgs bosons . Mass and length of an object change as its speed increases as it nears the speed of light . Even the size of an object depends on the gravitational field that it is experiencing. A table of a certain size might be stable within one gravitational field but collapse in another gravitational field because its legs cannot hold up the weight of its tabletop. Any feature of an object, in fact, can be described as the effect of interactions.
Given these definitions of creativity and feature, we are able to quantify the number of features, interactions, and effects by defining a feature as an effect that results from interacting the object of interest with other objects, materials, forces (e.g., centripetal and centrifugal), and energies (acoustic, magnetic, chemical, biological, human, thermal, electrical, hydraulic, pneumatic, mechanical, electromagnetic, and radioactive: ). Given that some amount of a material (e.g., a patch of velvet or a chunk of steel) can be considered an object, we can leave out material from the definition of feature above. Also, the lists of forces and energies may increase someday, especially as we better understand dark matter and dark energy, but these lists are currently stable but potentially dynamic in the future.
For our calculations, let us estimate that there are 10 million objects in the world. In April 2015, the US Patent Office issued its nine millionth patent , and this number does not include the patents unique to patent offices of other countries or the trade secrets contained in no patent databases. Further, this estimate leaves out natural objects (e.g., stone) and common objects (e.g., ball) that are also excluded from all patent databases. Further, the number of patented objects grows everyday as new patent applications are submitted on a daily basis. However, 10 million is a reasonable estimate for the present time, and it is an easy number with which to do calculations.
Given an object of interest, how many interactions are possible with 10 million objects? Strictly speaking, there are 210,000,000 possible subsets of 10 million things, which is approximately 1080, so our object of interest could interact with every possible subset of objects. More realistically, however, an engineer might interact their object of interest with between one and five other objects, which would result in on the order of 1027 subsets. Computers have existed for on the order of 109 seconds, so to examine all subsets of five or fewer objects would require examining 1027/109 = 1018 subsets per second since the 1950s. The fastest supercomputer as of June 2015, the Tianhe-2, computes on the order of 1016 floating-point operation per second . So, if the Tianhe-2 existed since the first computer existed, it could still not examine all the possible interactions of our object of interest with a reasonable number of subsets of possible objects. This calculation only allows one floating-point operation to process each subset. Further, it does not take into account all the possible conditions that these interactions could take place in: differing barometric pressures, humidity, temperature, lighting intensity, radiation, magnetic fields, strength of gravitational field, and so on.
In sum, even with our conservative estimates, the current fastest supercomputer could not fully explore the space of possible interactions for our object of interest in a reasonable amount of time.
2.3. Many ways to interact
The assumption made in the previous section is that, given two or more objects, it is obvious how they should interact. A spoon is used to stir the contents of a coffee cup, for example. That is what
Or, again set the coffee cup upright on the counter and place the spoon horizontally so it rests across the opening of the cup. Turn the spoon over so the curved part is facing upward and play a game of trying to balance various objects on the curved surface so they do not fall into the cup. Or, shake a spoon around in an empty cup to make a rattling sound. Or, turn a coffee cup over so that the open end is facing down. Place a spoon into the open end of the empty coffee cup and set the contraption on the counter. The spoon will force one side of the coffee cup to elevate a bit, forming a trap. When the spoon is disturbed by a mouse, for example, the coffee cup will fall and flatten, possibly trapping the mouse.
These are just a few of the ways to interact the spoon and the coffee cup to achieve an interesting effect. To consider all the ways that these two objects could interact, we would have to take into account every possible spatial relation between the two objects; every possible speed, acceleration, and deceleration of the two objects with respect to each other; every possible type of movement (linear, nonlinear, spinning at various angles and speeds); every possible surface that they may rest upon; every possible lighting condition, wind condition, heat condition, radiation level, magnetic field strength, electrical current flow, barometric pressure, humidity, earthquake or turbulence condition, and gravity strength; every possible extra object involved in the interaction (e.g., ice cube, marble, liquid coffee, and a human); as well as other conditions that we are probably overlooking.
If any of these conditions is actually measured by a continuous variable, then the number of different interactions between the spoon and coffee cup is truly computably nonenumerable. Even if all these conditions are measured by discrete variables that extend to a finite number of decimal places, then the number of possible interactions is outlandishly large. All these digits of precision on a variable are probably unnecessary in most cases, but when one is approaching a phase transition (e.g., liquid coffee approaching gas or the ceramic coffee cup possibly becoming superconductive), then many decimal places might be necessary to understand the onset of the transition. If one is approaching a previously unknown phase transition, then the slightest change in one condition, as measured by a change in the 100th decimal place for that variable, for example, could produce a radically different effect.
In sum, although we calculated that there may be about 1018 possible interactions between one object and up to five other objects out of 10 million possible objects, taking into account the incredible number of ways that any two objects can interact with each other plus all the possible conditions that those interactions could take place in, raises our number of interactions at least several orders of magnitude and quite possibly many orders of magnitude . The overall result is a number of possible interactions that becomes increasingly beyond the ability of current and projected supercomputers to explore even if they were running since the invention of the first computer.
When quantum computers come fully into being, then all the above calculations will need to be redone. There has been work showing how quantum computers could handle certain computably enumerable sets . However, if any of the conditions (e.g., heating, humidity, radiation, etc.) actually requires a continuous variable for its measurement, then the number of possible interactions is truly continuous and thus not computably enumerable. If all the conditions can be measured with discrete variables, then it is possible yet unclear whether the set of interactions is the type of set that is computably enumerable by a quantum computer, according to the specifications in Ref. . Even if the set of possible interactions were computably enumerable, however, any gaps in the theories involving those interactions—as described in the next section—would make the set of derived effects from the set of interactions uncomputable.
2.4. Predicting effects computationally
Can a computer compute the effects of a set of objects or entities that are interacting? It depends on whether a theory exists that derives the particular effects under consideration. Sometimes, theory is ahead of empirical measurement and sometimes empirical measurement is ahead of theory. For the former, Einstein’s General Relativity, developed between 1907 and 1915, predicted that light would bend around massive objects such as our Sun . It took until 1919, however, until Arthur Eddington verified this prediction by measuring starlight that moved around a total solar eclipse . For the latter, empirical measurement determined that galactic clusters did not have sufficient mass to account for their rotational speeds, so the existence of dark matter was posited as a way to increase gravitational effects present in galactic clusters .
If no theory exists to predict a particular effect of an interaction, then no algorithm exists to compute that effect. Given our previous example of a coffee cup interacting with a spoon, if there are gaps in the theories for how the interaction would proceed in a possible condition (e.g., lighting, wind, heat, radiation level, magnetic field strength, electrical current flow, barometric pressure, humidity, earthquake or turbulence condition, and gravity strength), then no computer could predict the effects within that particular configuration. That particular combination of conditions would have to be empirically measured. Thus, a computer’s ability to list out a particular combination of conditions does not mean that the computer could successfully predict the effects of the interaction taking place within that amalgam of conditions.
2.5. Humans countering computer limits
Humans are needed to carry out the empirical measurements that neither a computer can carry out nor a robot has not been set up to execute. Further, with our vast experience of interacting with the physical world, humans already know many effects of interactions but have yet to encode them for a computer. If humans have not yet experienced the interaction, often we can comfortably predict the main effects of that interaction after running a mental simulation in the sensorimotor cortices of our brain [23, 24].
In this way, humans can help flesh out and teach the computer many effects that the computer does not currently know and is presently unable to derive. Further, humans are good at crafting theories that make predictions of effects that then can be empirically tested. So, humans can encode their theories that a computer can then use to derive effects. Although a computer will continue to learn new effects taught to it by humans and derive effects based on new theories, given the computable nonenumerability of effects, humans will continue to maintain their rightful place in innovation—even with the onset of quantum computers (see previous Section 2.3 M
3. Human weaknesses to innovation and counter-techniques
In this section, we present five human blind spots to being creative and innovative (i.e.,
3.1. Functional fixedness
Most people first try to light the candle and drip wax around the rings. However, the rings are too heavy to be fastened securely with a wax bond. The key is to notice that the candle’s wick is a string. Remove the string by scraping the wax away on the steel cube and tie the rings together.
People who used the GPT solved 67% more problems than a control group . The idea is to break an object into its parts while you ask two questions. First, can the object be broken down further into smaller parts? For example, in Figure 1,
Software that exists can find the solution to the
3.2. Design fixation
Although the number of features of a candle (or any object) is intractable and not computably enumerable, classifying the types of features that any object could possess into an extensive category system has been a highly effective method for overcoming
We initially listed 32 categories of feature types for objects, but now use a 50-category system . We asked people to list as many features as they could for many common objects (e.g., candle, umbrella, etc.). We then categorized their answers based on our 32-category system. On average, people overlooked 20.7 of the 32 categories (67.4%) for each of the objects . For each object, they overlooked different types of features. For example, for a rocking chair, they would notice motion—that the chair was designed to move in a certain way. For a candle, however, no one we tested ever noticed that a candle is motionless when it burns. Its flame flickers, but the candle itself does not move.
To be innovative, you need to build upon a feature that has been commonly overlooked and, based upon our findings, the majority of feature types of common objects are overlooked. Therefore, there is plenty of room to create novel variations for even the most common of objects.
For example, a candle that moves from its own dynamics is an under-explored type of candle. Examining the other overlooked features for a candle, we found that no one noticed that a candle loses weight when it burns. Thus, we leveraged weight loss to produce a candle in motion. By placing a candle on one side of a scale-like object and a counterweight on the other side, the candle moved upward slowly as it burned down. For fun, we placed a snuffer above the candle so that it eventually moved into the snuffer and extinguished itself. The
Computationally, in ConceptNet 5.5 , a candle has no connection with being motionless or losing weight while burning, while a rocking chair has many connections related to motion. ConceptNet 5.5, as an example of many textual and data sources, would not be a good source for noticing overlooked features that could become the basis of a novel design. The overlooked features need to be uncovered through another method such as using the extensive category system of feature types discussed above.
3.3. Goal fixedness
Focusing on the verb, people are able to list between 5 and 11 synonyms of a verb . Humans drastically underperform when compared to the synonyms that are present in a good thesaurus. In
For example, suppose a person was working on fastening the rings together in the
3.4. Assumption blindness
Any phrasing of the goal belies many assumptions . For example, a company was stuck on trying to adhere a coating to the nonstick surface Teflon. Everything they tried failed. However, some analysis of the verb
Noticing three of the assumptions was crucial to a solution: (1) using a chemical process between (2) two surfaces where (3) contact is crucial to the solution. Exploring alternatives to these assumptions led to a novel solution: (1) using a magnetic process among (2) three surfaces where (3) contact is not crucial to the solution. Specifically, a magnetic surface is placed behind the Teflon surface, while the coating with some ferrous content is placed in front of the Teflon surface. The coating sticks through the Teflon to the magnetic surface and forms a kind of
In general, there is a master list of 50 types of features that any physical solution might possess : including size, shape, material, quantity, type of energy used (e.g., chemical, magnetic, etc.), spatial relations among the parts, symmetry, and motion. To uncover some important assumptions, simply proceed through the list and ask if the verb under consideration assumes anything about each of these feature types.
These types of assumptions are contained in neither ConceptNet 5.5 nor, most likely, any other current text or data source. These assumptions need to be unearthed carefully through a method such as the one described above.
3.5. Analogy blindness
Gick and Holyoak [9, 10] were the first to show experimentally how difficult it is for humans to notice by themselves how an idea from one area could be adapted as a solution in another area. For example, they had participants read a brief military story that held the crucial idea for solving a surgery problem [9, 10]. Thirty percent solved the surgery problem after mere exposure to the military problem, but 80% solved it after being told to use the military problem to help solve the surgery problem.
Building upon the work of Julie Linsey and colleagues [32, 33, 34, 35, 36], who focused upon looking at synonyms of the main verb expressing the goal of the problem, McCaffrey and Krishnamurty  went one step further to explore the synonyms of both the verb and noun phrase of the goal. For example, consider the goal
Next, we entered each of the phrases into Google in the form “concussions diminish trauma” . This step helped us determine which phrases were under-explored in the context of concussions. We found that
Obtaining synonyms from
4. Human-computer interface to achieve synergy
In order for humans to counter computer limits and computers to counter human weaknesses, an interface is needed, which is comprised of data structures that both humans and computers can easily populate. In order to make the human-computer interaction efficient, the interface needs to be both human- and computer-friendly. Building upon our new definitions, we can define a problem as a set of desired effects and define a solution as a sequence of interactions that ultimately produces the desired effects named in the problem.
We define the problem solving grammar for innovation in Extended Backus-Naur Form (EBNF: ), which is a compact notation mostly used for defining the syntax of computer programming languages. For our grammar, we only need a few of EBNF’s symbols: “::=” means “is defined as,” a superscripted “+” means there can be one or more of the preceding item, and a superscripted “*” means there can be zero or more of the preceding item.
A goal is a set of desired effects. Any effect can be described as an action verb that describes a change (or a nonchange), a noun phrase to name that which needs changing (or should be kept from changing), and a list of prepositional phrases that describe important constraints and relations. A feature is synonymous with an effect, but sometimes a shorthand can be used: an adjective (e.g., heavy) or a noun phrase (e.g., a heavy, metal rectangle). A resource is either an object (e.g., hammer), a material (e.g., velvet), an energy (e.g., magnetic), or a force (e.g., centrifugal).
Each of these grammatical forms is both human- and computer-friendly. Each phrase has a natural English form, and each phrase is regular so that it is easy for a computer to parse.
5. Current implementation and applications
The current implementation of
In this way,
Every innovative solution is built upon an obscure (i.e., commonly overlooked or new) feature of the problem [13, 27]. Both computers and humans tend to overlook different sets of obscure features based on their differing search biases. These differences are somewhat complementary so that computers and humans can help each other uncover obscure features that the other partner would miss [6, 11]. The result is significantly more unearthed features of the problem, resulting in a higher chance of unearthing the key obscure features required for a novel solution to the problem. Further, computers cannot completely take over the creative and innovative process due to the fact that the set of features of any object is not computably enumerable, so it cannot be fully explored by a computational device . Working together through a computer- and human-friendly interface called
Specifically, imagine IBM Watson plugged into the
This article is based on work supported by National Science Foundation Grants 1534740, 1331283, and 1129139. Any opinions, findings, and conclusions or recommendations expressed in this publication are those of the author and do not necessarily reflect the views of the National Science Foundation.
Conflict of interest
This research is associated with my company, Innovation Accelerator, Inc. (www.innovationaccelerator.com), and may lead to the development of software products, in which I have a business and/or financial interest. I have in place an approved plan for managing any potential conflicts arising from this arrangement.