The Future of Problem-Solving

 

Why we need help and where it will come from

 

The Human Condition

For millennia, humans have used creativity and ingenuity to tackle problems of increasing complexity. First we developed more efficient ways of meeting our biological needs: tools for hunting, weapons and shelter for safety, and farming practices for convenience. We extended our lives with increasingly researched and revised medical practices, and developed more advanced and faster modes of transport and communication. We also created extensive education systems to further improve our ability to do all of this.

More recently, the advent of computer technology has enabled refinement of pretty much every existing invention. From CAD programs used in architecture, to safety and navigation features in cars, and instantaneous global communication via the internet, among countless others. All of which serve to make up for human shortcomings in some way.

Despite our enterprising natures, however, there is a limit to how much we can accomplish. Our problem-solving capabilities are restricted by how much information we can consciously hold and manipulate at once, using our working memory. This is particularly problematic in novel or unexpected situations that have important or critical outcomes. We need to integrate incoming information with our existing knowledge in real time to make an appropriate decision.

Our ability to do so is further complicated by competition from biological imperatives and a constant barrage of irrelevant stimulus from our environment. We can become hungry, thirsty, tired, bored, emotional and distracted. We make mistakes. And our decision-making is not always optimal as a result. In fact, everyone defaults to reflexive or habit-based decisions and responses under time constraints or environmental pressures, rather than thinking things through. This can prove deadly in any number of situations.

Limitations of Computers

Computer technology has been of enormous benefit to humans already, as machines aren’t constrained by biological needs and have processing capabilities far beyond ours. It stands to reason that the logical next step is to build computers capable of assisting humans in critical decision-making.

 So far the focus has been on artificial intelligence, however it is also limited in its capacity to handle novel situations.

We can easily visualise this in a chess scenario. For example, if an important rule was changed or added, such as how pieces can move, right before an AI chess machine faced a human grandmaster in a tournament, it would suddenly be operating on outdated information. It either needs to be given new statistical data or the chance to compile it. However, the human chess player can immediately integrate the new information with their existing game knowledge and develop new play strategies in real time. They are likely to win because humans can learn and adapt independently, whereas AI are simply following a set of precompiled instructions for a specific purpose or context, which humans must give them.

Computers are also unable to change their objective in response to an otherwise unforeseen variable. If a human is at home preparing a meal and an earthquake unexpectedly hits, their objective changes to seeking safety, and no one had to tell them to. But an AI wouldn’t be able to abandon its task and reorient appropriately unless its programming included specific detailed instructions for that scenario.

Such issues arise because computers aren’t truly intelligent. They rely on using algorithms, which are great for making calculations in a finite sequence using known variables, but not particularly useful in situations where not every variable can be anticipated in advance. Statistics are helpful for finding correlations and adjusting one’s approach to something based on the data, but they won’t help you avoid something you’ve never encountered. True intelligence includes the ability to acquire knowledge and build a flexible model, which can be modified, refined and adapted with experience to solve problems that are completely novel. The human mind is full of unstructured models on how things work or how to do things, which can be adjusted to extrapolate beyond what is already known.

Raw computing power and processing speed isn’t enough to solve new or unanticipated problems, but requires an element of versatility that machines have conspicuously lacked. Ultimately, artificial intelligence in its current state will fail to meet the rapidly approaching needs of the future.

The Next Step

Both humans and current computer technology face enormous problems when dealing with novel situations — and the solution lies with the development of cognitive machines, a computer system that mimics human cognition. Such a machine would be capable of learning independently to build its own models and apply that knowledge, even modifying its internal workings, unlike current AI. At the same time, it wouldn’t be restricted by working memory capacity.

So how can a cognitive machine help with critical decision making? Responding to a global pandemic is an apt example. At present, global leaders are tasked with managing lockdowns, developing a vaccine, negotiating border closures and international relationships, dealing with economic ramifications, funding additional healthcare, and appeasing or reassuring the general public. That’s a lot for only a few people to manage! How can any one person address all of these issues, let alone look for any unforeseen complications? They simply can’t.

Usually, a taskforce of experts is assembled, but each is limited in their capacity to understand fields outside their area of expertise. Human collaboration suffers other issues too, such as language barriers and cultural differences as well as miscommunication with body language, tone and sarcasm. Experts can hold prejudices, conflicting personal interests, or a belief that one’s own area is more important than others. Limitations are imposed by those who don’t feel psychologically safe enough to speak up or who don’t want to voice anything potentially detrimental to one’s own job. Some could be lying or withholding crucial information, while others may cede a point or capitulate in a compromise to finish up faster, resulting in an imperfect solution. Perhaps most inimical is the time taken to reach a solution — by this stage, the problem may have shifted or progressed to the point that it is obsolete.

A cognitive machine does not suffer these woes. It could be an expert in multiple fields with equal understanding and weighting. Or multiple expert machines could collaborate, free from human constraints and impartial to the outcome.

This approach would allow us to equitably address all components of managing a pandemic, and could also highlight something humans would overlook or disregard entirely. The machine could address even the most seemingly irrelevant points, embracing and utilising any area of expertise immediately. Most importantly, it could do all that a human taskforce could (and far more) in a mere fraction of the time — so quickly it may have a strategy effective enough quash a crisis before it even gains momentum, or by acting in real time to efficiently counter each new issue as it arises. All the while, human experts are still busy scrambling for their seats.

Building The Future

The good news is that cognitive machines are close to being a reality. At Interactive Engineering, we are making it happen! But how?

Our machine will be using a language network to form its active knowledge structure, much like humans do. It will be provided with a dictionary to facilitate its understanding of language and build a lexicon it can use to interact with new information. This ability to adapt its internal structure in real time is what allows our machine to think like a human and learn.

Before our machine can do this however, we must first address some of the issues that dictionaries have. These problems arise because they are written by humans for humans, which means they frequently rely on us knowing how something looks, sounds or feels (physically, mentally and emotionally) to relate to a word’s meaning. Often, writers make assumptions that something is ‘obvious' because humans can derive certain clues from context, which a machine is unable to do. We are currently in the process of editing entries of the Oxford English Dictionary (American usage variant) to identify and address any such issues that may trip up the machine.

Many words have more than one meaning or use, and as a result they have separate entries in dictionaries for any homonyms as well as different syntactic functions (noun, verb, preposition, etc). Each of these entries may in turn have senses with embedded sub-senses to define branches of meanings. We are restructuring entries for polysemous words (those with multiple meanings) to ensure that the hierarchy of senses is arranged to suit rapid searching, rather than by frequency of use or even at random.

It is important to note that there are several different types of senses. Many attempt to explain the meaning of a word using several simple words, while others describe the purpose of a word e.g. one definition for ‘about’ is “used to describe a quality apparent in a person.”

Some entries even describe grammatical functions of words, such as when and how ‘by’ is used in a sentence with a passive verb. These types are particularly important because they change our internal language structure as we are using it to read them. For example, we can see that the word ‘verb’ is categorised as a noun in its definition “a word used to describe an action, state, or occurrence…” It requires us to understand when and why associations don’t carry over; to know that the word ‘verb’ is a noun (while its function is being described) but that it doesn’t mean every verb is a noun.

We tend to know many of these usage definitions intuitively because we built a model as we learnt the language, but a machine doesn’t have this luxury. So an important part of the editing process is to ensure these types of entries are clear (or actually included) to change its inner workings appropriately and guide its understanding of language rules.

For entries that define meaning, we need to resolve any ambiguity. This includes breaking up circularity, which is a tendency for semantically-related words to be defined by each other. For example ‘comment’ is defined as “a verbal or written remark expressing an opinion or reaction” and ‘remark’ is defined as “say something as a comment” We sort of know what the difference is between them, but the definitions don’t clearly state what each word means independent of each other. The machine is going to need more precise definitions to understand contextual nuances.

Another task is the creation of new entries for word phrases that convey a specific abstracted or figurative meaning, one that colours the literal meaning of the words but otherwise isn’t obvious from the definitions of the single words combined, e.g. ‘grow up.’ A child can grow up into a mature adult but telling someone to grow up means to tell them to behave or think sensibly, something that is attributed to maturity, and typically used in a demeaning or insulting manner. However the latter meaning isn’t obvious from ‘grow’ and ‘up’ on their own, it’s something we intuitively know from tone and context.

Including definitions for word phrases used more frequently than the component words alone (e.g. bode ill and bode well) will also improve the machine’s efficiency by decreasing time spent on parsing sentences.

Editing and reorganising a dictionary is no small feat. It is incredibly time consuming and complex, and relies on humans (who have to think about what they know unconsciously) to do the work. Until the machine has a sufficient grasp of language to take over, that is.

Future Implications

Problem-solving of the future lies in the development of cognitive machines, and the applications are limitless. They could be used to determine effective strategies for slowing and reversing climate change that we wouldn’t have thought of. Or help diagnose patients faster by building a model of the comorbidities of a specific patient as reference, rather than haphazardly sampling irrelevant statistics. They may even prove the pivotal element that enables us to safely embark on a trip to Mars. The technology for these accomplishments is within our grasp, and cognitive machines can bridge the gap.