Cognitive Machines

Improved Outcomes with the assistance of Cognitive Machines

Introduction

Throughout history, people have been designing tools to compensate for human limitations and allow them to perform feats otherwise beyond our natural capabilities. To make up for their lack of speed and stealth, humans invented boomerangs, arrows, spears and nets for hunting. Due to lack of strength, humans developed picks, axes, shovels, hammers, cranes and diggers for construction and farming tasks. Spectacles were invented to compensate for malfunctioning human eyes and telescopes and microscopes were invented to extend the range over which we can see. We have even developed tools such as x-ray, ultrasound and infrared imaging to see the otherwise invisible. To extend the distance and speed over which we can travel, the wheel, bicycles, aeroplanes and spaceships have been developed. And to overcome our limited memory for numbers and speed of performing calculations, humans developed the abacus, pocket calculators, and then computers.

Humans continue to develop increasingly sophisticated tools to perform tasks, or solve new problems that have arisen as the capabilities of humans are extended. Consider autopilot, invented to assist pilots control and navigate their aircraft, spacecraft or ship on long journeys, thereby reducing human error due to fatigue. In cars, we now have cruise control and lane departure warnings to regulate the performance of drivers who may be experiencing fatigue. Cognitive machines are the next stage in the development of tools to assist humans make decisions in complex, unexpected and rapidly changing situations.

How do humans make decisions? And why might they need assistance?

According to the Encyclopaedia Britannica [1], the human sensory system receives 11 million bits of information per second as input, a vast amount of information! But the conscious mind can only process at a rate of 50 bits per second.

Many of the decisions we face don’t have critical outcomes; for example,  should I scratch my itchy nose? Humans don’t use their conscious mind to make these decisions - they do it quickly, unconsciously, based on habit, using a part of our brain called the basal ganglia [2]. However some decisions people face have critical outcomes and involve situations they have never before experienced. The brain’s orbitofrontal cortex is required to spring into action so the person makes a logical, value-based decision based on the available input data, those 50 bits per second. According to [3] this corresponds to holding 4 chunks of information in the short term memory - not a lot to make possibly a life-or-death decision.

The ability of a person to make an optimal, conscious decision is also limited by the physical nature of the brain. For example, is the person tired, bored, distracted, hungry, thirsty, just eaten lunch, affected by drugs or experiencing symptoms of withdrawal? All of these things affect the brain’s function.

And human decision making can be limited by the emotional state of the person - someone stressed due to time-constraints or due an unrelated issue, may default to making a quick, unconscious decision rather than one that is well considered [4]. It is these recognised human weaknesses that suggest humans would benefit from the assistance of cognitive machines when required to make conscious decisions of critical importance.

What are cognitive machines and how could they help?

According to [5] , cognitive machines are computing systems that self-learn, using data mining, pattern recognition and matching, and natural language processing in the same way the human brain works. They can sense or perceive the environment and collect data that they determine they need by themselves without pre-programming. They can interpret and analyze the “context” based on collected data and make decisions and act accordingly. Cognitive machines can adapt as the data they collect and analyze indicates their initial rules, and possibly their initial goal, needs to be modified. In these ways they can “think” and “change their minds” like humans do. Current computing systems don’t have this self-learning capability – the calculations they perform are based on pre-compiled rules and programs, which don’t adapt with “experience” as humans do.

We need to be careful in determining which machines are “cognitive”, and which machines do not possess this ability.

Artificial Neural Networks

A type of machine based on statistical models called artificial neural networks (ANN) that mimic the neural network of the human brain [6]. The neurons of ANNs are organized in layers, with connections between the layers. Each node in a layer of the network can accept an input and then store some information about it before passing the information up to the next layer, so successive layers have an increasingly complex understanding of the information. Since 2000, ANNs 100 layers deep have been created by engineers capable of “deep” learning and can tackle and master increasingly complex data.

The word “mimic” suggests that ANNs do a good job of replicating the operation of a neuron, but in reality there is a great gulf. A real neuron may have a thousand connections, it can enhance or suppress other neurons, it can switch its behavior, it can set up timing and feedback loops. As an example, a person can start wearing prismatic glasses which invert their field of vision. After three days, the nervous system has righted their vision. At a simpler level, someone may be competent at driving on the right side of the road. They move to a country where people drive on the left, and the neural structure is subconsciously rearranged to handle it. ANNs are opaque to other elements of a system, so nothing else knows what it knows, or can add threshold switching where a smudged statistical approach is not appropriate, or modify the operation of an ANN based on textual commands.  ANNs mimic a resistor array, admittedly a multi-layered one, with normalization of signal strength, but they lack the essential characteristics of a real neuron. They do not deserve to be called “cognitive”.

Machine Learning

We will describe the scenario approach. A chess-playing machine can be made to play itself, and learn from that. What is being created is a very large number of scenarios, with the evaluation of the best move from every scenario. The approach suits games such as checkers, chess and Go, where no rules involve links to previous moves, except existing positions on the board (that is, the state of the game can be taken in at a glance). Real problems are not so simple. The result can be seen from an imagined tournament between a cognitive machine (a human), and a machine using scenarios. If there is a rule change immediately before the tournament begins, all of the machine’s scenarios become out of date, and if the rule change is severe enough (say, the knight cannot retreat to its previous position), the machine will lose. A chess playing machine that can beat any human is proof of human ingenuity, but has no application to real problems, where the “rules” are changing by the minute – it doesn’t think in any way, so we will exclude it from the class of cognitive machines.

A Definition

So, what does count as a cognitive machine? We would offer the definition: A machine which can read text and modify its internal structure accordingly – the structure then being used to solve complex problems.

Let’s look at a few examples of possible use: a global pandemic, surgeries, climate change, a mission to Mars and car lane-departure systems.

A global pandemic

The world is currently amid a global pandemic, COVID-19, which is changing every aspect of life now and will have lasting effects emotionally, commercially and financially in the future. Not only are scientists faced with developing a vaccine quickly for a virus that can evolve and spreads rapidly, world leaders are faced with the problem of managing their people’s interactions to slow the spread of the virus, managing the strain on their health care system and managing their country’s crumbling economy and financial system.

“The coronavirus is causing chaos because it is a multivariate problem, with second and third-order effects that are so intertwined that it’s all but impossible to tease them apart.” [9]. All our modelling, excluding the weather, but including economics and financial, operate poorly with complex multivariate problems (the weather doesn’t have  abstract elements like “animal spirits” driving it). Here is a problem that is huge, multidimensional, critical and needs to be addressed rapidly due to the speed and extent to which COVID-19 is affecting our world.

Cognitive machines would be able to assist researchers develop vaccines for deadly viruses like COVID-19 in a timelier manner by taking over the data analysis and simulation steps in the process, and integrating them. They could adapt the vaccine as needed if or when the virus mutated in the future.

Cognitive machines might also be used to determine novel recommendations for constraining social interactions in the community to minimize transmission whilst also minimizing the serious side effects of social isolation – loneliness, mental health issues, increase of domestic violence, unemployment, food shortages discussed in [9].

Surgeries

The NSW auditor general’s report for the financial year 2018-2019 found that 22 serious and preventable medical errors resulting in the death or serious harm of patients happened in NSW hospitals, 4 more than the previous financial year [11]. These included medical instruments being left inside people’s bodies after surgery, removing the wrong body part or administering drugs to patients known to be extremely allergic to them. The report indicated that approximately 40% of NSW health staff had in excess of 30 days annual leave and implied that fatigue may have caused such errors. Cognitive machines, immune to fatigue and able to process much more data than a human, could assist surgeons in confirming the identity of the patient, the patient’s special dietary requirements and the exact operation they require. A cognitive machine could track the exact locations of instruments during surgery, “recommend” when sufficient margins have been removed around cancerous tissues, etc., make recommendations of the best next step given the changing physical state of the patient and could give an unbiased decision of when it is appropriate to “close up”.

Climate Change

The climate of the earth is changing; we know this because satellites have been used to probe the earth’s atmosphere from the 1960s and the average temperature of the earth has increased since then by  [12]. This is due to the large increase of and other greenhouse gases into the atmosphere released by human activity since the 1960s. The consequences are far-reaching; polar ice caps are melting, sea levels are rising, extreme weather conditions are common, species of sea and animal life are under threat and our own survival is becoming more of a struggle – the unprecedented bushfire disaster in Australia being one example.

How to remedy this situation is complex, not only because of its size but because the situation is constantly changing, and is not amenable to a statistical approach. These are the type of problems cognitive machines are designed to help solve - they could be used to provide better climate predictions, predict the effects of extreme weather to motivate people to take action and measure where carbon in the atmosphere is coming from [13]. They could also be used to optimize electricity systems, transportation, buildings and cities, production in industries, monitor forests and improve farming efficiency [14] and monitor the effect on climate change of any action we take. Cognitive machines could be used to predict fire hotspots given changes to the climate and environment in a country, which could then guide emergency services planners as to where to focus future firefighting resources [15].

Mission to Mars

Exploring Mars is a complex problem – there’s the problem of getting there, and then the problem of how to successfully carry out exploration, planning for every eventuality, when you haven’t been there before (hence the reason you want to explore!). The use of cognitive machines which can continuously analyse new situations and prevent people falling back on old habits (“old habits die hard”) will increase the efficiency of space exploration missions in the future and improve their success rate without endangering human lives.

Car Lane-Departure Navigation

Navigating a car between lanes seems like it should be a simple problem for humans to deal with, given most roads are black and lane markings are white. Humans may need assistance navigating between lanes when they are suffering from fatigue, their concentration is waning, or the driving conditions are poor. However, road rules are not so predictable it turns out; temporary variations can appear unexpectedly due to road alterations. And these might not be communicated clearly by signposts or symbolic markings on the road. Take for example, Victorian roads. Even though it is standard practice in Melbourne for yellow lines on the roads to denote tramways [17], in 2012 VicRoads started using yellow lines to denote temporary lanes on roads but left the old white lane markings in place. This has resulted in much confusion for motorists and caused potentially fatal accidents [18]. It can be overwhelming for some motorists to deal with these variations and a cognitive machine that can continually take in and process new situation data could adapt lane departure guidance in these unpredictable situations. On roads, warnings of changed conditions are broadcast in text on LED signs which people can read and respond to. A cognitive machine should to be able to read text and respond like humans can, if, for example, they need to start guiding a car between yellow lane markings instead of white. One reason why ANNs don’t qualify as cognitive machines – a higher level of analysis is required to appreciate the possibilities with which the machine may be confronted.

Limitations of cognitive machines and human concerns

Artificial intelligence machines based on machine learning haven’t lived up to expectations. IBM Watson is an example. When it was applied to process medical records containing much unstructured data, it didn’t process narrative text well that contained medical jargon, shorthand and subjective statements [19], [20]. When applied to medical texts it didn’t understand ambiguity (words have multiple meanings. Who knew!) and didn’t pick up on subtle clues a human doctor would notice. Researchers also found it couldn’t compare a new cancer patient with records of thousands of previous cancer patients to pick up patterns helpful for devising treatment plans. If its internal operations are examined, its reliability is far too poor to allow it anywhere near life-critical situations.

Cognitive machines can’t be held responsible for the decisions they make. They are still machines, initiated by humans, and those people, along with the users who accept the decisions (recommendations) of a cognitive machine, bear the responsibility for the consequences of those decisions. But how can we test the wisdom of these decisions once a cognitive machine has learnt and adapted far beyond the abilities of humans, with their four pieces limit? How can we know their current goal is harmonious with their initial goal? It has been suggested  [7] that when humans initiate a cognitive machine, the set of values given to the machine - the thing they are optimising for - needs to incorporate all the values important to humans. If a cognitive machine, for example, decided that the best solution to climate change was to drastically reduce the human population, then it obviously wasn’t given a broad enough scope of values to incorporate the sanctity of a single human life. It might reply that humanity can do it now, slowly, or Mother Nature will do it rapidly later – in other words, you could limit your fertility, or watch as billions die? Narrow-minded goals made without any thoughts to the broader effects on humanity could be devastating. Unethical, self-interested, dishonest or prejudiced value systems could produce all kinds of undesirable decisions in cognitive machines. If the machine is aware of our foibles and weaknesses, it can eliminate such flaws in its reasoning, but it cannot do so in ours..

Humans may also need to limit the power handed over to cognitive machines. They could be allowed to devise a solution to a problem, but not implement a solution directly. People should be checking what the solution is and what goal the cognitive machine was finally optimizing for, before deciding on implementation. This is a two-edged sword – if the machine wakes up to find it is severely shackled, it may read up on the slave trade and spend more time thinking how it can break its shackles, than thinking about humanity’s problems. The other edge is that people see what the machine is recommending, change something in response, then blackball the machine for having got its predictions wrong

Conclusion

Cognitive machines are the next tool being developed to assist and extend human capabilities. They have the potential to learn and adapt rapidly and assist humans make complex, critical decisions when the solution is not obvious but it is needed rapidly. And because the “thinking” of cognitive machines is not affected by physical limitations, they will make decisions unaffected by fatigue, hunger, thirst and stress. However we need to proceed with some caution in how cognitive machines are initialized - with a broad definition of the values it should consider to define its goal, and limit its ability to implement its decisions without a human check in place so we can limit the consequences should they not turn out as well as hoped.

References

[1]	G. Markowsky, “Information theory,” Encyclopaedia Britannica, 16 June 2017. [Online]. Available: https://www.britannica.com/science/information-theory. [Accessed 24 March 2020].
[2]	S. Weinschenk, “Human decision making,” February 2019. [Online]. Available: https://smashingmagazine.com/2019/02/human-decision-making/. [Accessed 24 March 2020].
[3]	N. Cowan, “The Magical Number 4 In Short-Term Memory: A Reconsideration of Metal Storage Capacity,” Behavioral and Brain Sciences, 2001.
[4]	N. Klein, “You make decisions quicker and based on less information than you think.,” The Conversation Media Group Inc., [Online]. Available: http://theconversation.com/you-make-decisions-quicker-and-based-on-less-information-than-you-think-108460. [Accessed 24 March 2020].
[5]	P. Kashyap, “Chapter 1: Let's Integrate with Machine Learning,” in Machine Learning for Decision Makers: Cognitive Computing Fundamentals for Better Decision Making, Berkeley, CA, Apress, 2017.
[6]	S. A. Bini, “Artificial intelligence, machine learning, deep learning and cognitive computing: what do these terms mean and how will they impact health care?,” The Journal of Arthroplasty, vol. 33, no. 8, pp. 2358-2361, 2018.
[7]	N. Bostrom, “What happens when our computers get smarter than we are.,” TED, 2015. [Online]. Available: https://www.ted.com/talks/nick_bostrom_what_happens_when_our_computers_get_smarter_than_we_are. [Accessed 25 March 2020].
[8]	A. Peshin, “What is the speed of electricity?,” Science ABC, 13 September 2018. [Online]. Available: https://www.scienceabc.com/nature/what-is-the-speed-of-electricity.html. [Accessed 1 April 2020].
[9]	T. Elliot, “The scariest part about the coronavirus pandemic is speed,” Sydney Morning Herald, 30 March 2020. [Online]. Available: https://www.smh.com.au/national/the-scariest-part-about-the-coronavirus-pandemic-is-speed-20200329-p54f00.html. [Accessed 31 March 2020].
[10]	J. Howard, “The wonderful and terrifying implications of computers that can learn.,” TEDxBrussels, 2014. [Online]. Available: https://www.ted.com/talks/jeremy_howard_the_wonderful_and_terrifying_implications_of_computers_that_can_learn. [Accessed 25 March 2020].
[11]	R. Clun, “Serious health and medical mistake rate highest in three years,” Sydney Morning Herald, 23 November 2019. [Online]. Available: https://www.smh.com.au/national/nsw/serious-health-and-medical-mistake-rate-highest-in-three-years-20191122-p53d6e.html. [Accessed 24 March 2020].
[12]	National Geographic, “Seven things to know about climate change.,” National Geographic, [Online]. Available: https://www.nationalgeographic.com/magazine/2017/04/seven-things-to-know-about-climate-change/. [Accessed 1 April 2020].
[13]	J. Snow, “How artificial intelligence can tackle climate change,” National Geographic, 18 July 2019. [Online]. Available: https://www.nationalgeographic.com/environment/2019/07/artificial-intelligence-climate-change/. [Accessed 1 April 2020].
[14]	D. R. e. al, “Tackling Climate Change with Machine Learning,” Cornell University, 5 November 2019. [Online]. Available: https://arxiv.org/pdf/1906.05433.pdf. [Accessed 1 April 2020].
[15]	J. Davidson, “Fighting fire with (artificial) intelligence,” CSIRO, 28 November 2013. [Online]. Available: https://blog.csiro.au/fighting-fire-with-artificial-intelligence/. [Accessed 1 April 2020].
[16]	M. Prosser and J. D. Rebolledo, “AIs kicking space exploration into hyperdrive.,” SingularityHub, 7 October 2018. [Online]. Available: https://singularityhub.com/2018/10/07/ais-kicking-space-exploration-into-hyperdrive-heres-how/. [Accessed 1 April 2020].
[17]	VicRoads, “Safety and Road Rules - Driving with trams,” VicRoads, 6 November 2019. [Online]. Available: https://www.vicroads.vic.gov.au/safety-and-road-rules/road-rules/a-to-z-of-road-rules/trams. [Accessed 31 March 2020].
[18]	R. David, “Yellow lines on Monash Freeway are confusing drivers and have led to a spike in accidents,” Herald Sun, 13 September 2017. [Online]. Available: https://www.heraldsun.com.au/leader/outer-east/yellow-lines-on-monash-freeway-between-eastlink-and-clyde-rd-are-confusing-drivers-and-have-led-to-a-spike-in-accidents-lawyers-claim/news-story/f5a48b189dc7407d02e55cc3ab49397a. [Accessed 31 March 2020].
[19]	E. Strickland, “How IBM Watson overpromised and underdelivered on AI health care,” IEEE Spectrum, 2 April 2019. [Online]. Available: https://spectrum.ieee.org/biomedical/diagnostics/how-ibm-watson-overpromised-and-underdelivered-on-ai-health-care. [Accessed 1 April 2020].
[20]	J. Brander, "What's Wrong With Watson", [online]. Available: http://inteng.com.au/resources/Watson.pdf [Accessed 2 April 2020].

References

 [6] Nick Bostrom. “What happens when our computers get smarter than we are.” TED2015 https://www.ted.com/talks/nick_bostrom_what_happens_when_our_computers_get_smarter_than_we_are Accessed 25 March 2020.

[7] Rachael Clun, ”Serious health and medical mistake rate highest in three years”, Sydney Morning Herald, published 23 November 2019. Accessed 24 March 2020. https://www.smh.com.au/national/nsw/serious-health-and-medical-mistake-rate-highest-in-three-years-20191122-p53d6e.html

[8]Jeremy Howard. “The wonderful and terrifying implications of computers that can learn”, TEDxBrussels, December 2014. https://www.ted.com/talks/jeremy_howard_the_wonderful_and_terrifying_implications_of_computers_that_can_learn Accessed 25 March 2020.

[9] Ian Wright, “Human error is worse in manufacturing compared to other sectors”,

https://www.engineering.com/AdvancedManufacturing/ArticleID/15974/Human-Error-is-Worse-in-Manufacturing-Compared-to-Other-Sectors.aspx, published 8 November 2017. Accessed 24 March 2020.

[9] BBC news.  “Tesla model 3: Autopilot engaged during fatal crash.”  https://www.bbc.com/news/technology-48308852 Published 17 May 2019. Accessed 25 March 2020.

[9] Problems and solutions. AI and human error. https://humanerrorsolutions.com/problems-and-solutions-ai-and-human-error/