AI is transforming the world but is facing serious problems

When we refer to AI today we are talking about “recognition.” It is important to note that recognition is only a small part of what you need to have something “intelligent.” I have documented some 12 different aspects of "consciousness" some of which may be necessary to have truly human level of intelligence, some may not.

Some proponents of AI will say recognition is all you need or that it will somehow transition from recognition to human level AI in a small steps that might be happening any day now. Others are convinced the technology today or soon will be so dangerous we need to have ethics and rules about building AI.

I agree we need to be careful about deploying AI but not because it is likely to be like "The Terminator" but because AI will be misused and deployed by humans to hurt humans.

My basis for saying this is that AI actually is far more limited than some people seem to fantasize today and the road to improvement of this AI is not going to be as smooth as some may imagine.

The Road to Today's AI hasn't been smooth

AI circa 1980 seemed very much a problem we could solve quickly. I believe many people believed at the time that "the learning algorithm" our brain used couldn't be that hard to figure out and it would be a short few years to develop AGI (General level intelligence.)

A legendary figure in AI at the time called the end of it. He pointed out that calling something a variable in a program "learning" didn't imply the programmed learned. A crash in AI happened after that article and for 10 years the field was treated with disdain and people looked at you oddly if you said you were working on that stupid idea called AI.

In 1990 AI became rules systems. It was thought that if we taught a system enough rules and allowed it to make deductive reasoning based on these rules that eventually we would have something that acted intelligent. Again a few years of this investment boom in this technology and it became anathema again.

In 2000 machine learning became in the fore. This was simply using mathematics to discover probabilities and statistical coincidences in data that could be used to make predictions. We had the first voice recognition systems and text recognition. It was faulty and no amount of learning seemed to rise about a certain level of incompetence. The field was still considered dead.

In 2010 we had the evolution of several new neural cell model AI systems that using convolutional mathematics produced the first neural like networks that showed much better results than pure machine learning by math did. Language recognition and other things improved dramatically. Image recognition started to really work.

This brings us to the current heyday. Excitement about AI is robust and the fact it works pretty well for some tasks has investors and pundits going crazy about the possibilities.

The truth is that today's AI applied across a lot of applications produces significantly more value than before and is transforming a lot of business and technology.

My point is not that AI is bad or useless. On the contrary I am very enthusiastic but I think to properly improve we also have to know where the limitations are and what we don't know and how it could be dangerous.

We have to be cognizant that in the past we have hit roadblocks and it is very likely we will hit such roadblocks now as well.

Recognition is what the new AIs do. Recognition means the computer can recognize speech, objects, or other correlations in data that looks like intelligence. However good that AI might get at recognition there is a big step between recognition and action.

AI’s don’t plan

In every case where we use AI today we may use the AI to recognize X but the action we take based on that recognition is always chosen by a system of programmed behaviors that follow the more traditional programming model where a human decides by writing non-learning code how to react to such objects or recognition.

In other words an AI might recognize a person’s face but what to do with that recognition is something we program. If the computer or AI punches the person then this is how the person who wrote the code decided they wanted the AI to act.

Programming of this type is not learned. We can’t write programs today that vary their behavior in arbitrary ways based on AI. An AI can be programmed to look at possible range of behaviors and choose statistically which path is more likely to succeed. However the paths and the way we do this is planned by humans not the AI.

For instance, an AI may recognize something and we may even have the program notice that certain actions produce bad results and modify its behavior but the range of possible variation in behaviors is limited by the programming that the AI is given. It can’t think. It may look like it thinks but it can’t. Not yet.

This is called the “planning” part of consciousness. Planning is the aspect of conscious of intelligent beings that they decide what to do based on what it has recognized.

Some animals can react to stimuli they recognize. The more primitive the animal the more direct the action to stimuli. In humans planning can involve many steps that for instance allows a person to do math very differently than a computer. For instance, simple arithmetic, such as adding numbers can't be learned.

In this study recently done using the best AI it was shown that using today’s algorithms for AI they can’t learn basic arithmetic.

If you think about this, it makes sense. We are taught at an early age how to do arithmetic by for instance adding each column and carrying the overflow to the next digit.

To learn arithmetic by AI requires presenting thousands of examples of addition. Unfortunately the computer does not “get” the abstraction that we are carrying over the overflow to the next digit. That is something we do explicitly by planning. Our brains execute the plan. We don’t just look at 47 + 193 and recognize the answer is 240. We compute it using a learned algorithm. We don’t have AI that recognize algorithms or learn algorithms.

This simple example illustrates the very real problem that AI is limited to today. A deep learning machine might recognize how to add 112135 + 135321 since it could memorize the sum of the individual digits and learn that the sum is simply the sum of the corresponding digits. I can imagine a convolutional algorithm doing this.

However, when the 2 digits exceed 10 the network would have to recognize that the sum requires it to bump the next digit. That is not something I see any network being able to figure out. It requires a higher level of abstraction and planning.

We have no way of computers learning planning like this or recognizing planning like this.

If a computer launches a nuclear bomb as part of current AI it is not because the computer decided malevolently to do this but because someone programmed the action response to recognize a certain threat scenario is to launch our nuclear bombs. We cannot ascribe malice or intelligence to the AI that does this because the computer didn’t decide what to do and if we let it do something based on a recognized scenario it is because we specifically programmed that response in, not because of the fault of the AI or computer but of the human that programmed it.

We should be clear about that and be careful when we program actions to be sure that the consequences of any programmed action is non-lethal or reversible or safe.

This planning limitation of learning is severe limitation and fundamentally makes a whole range of things unlearnable by AI's today. It is not clear if anyone has any idea how to approach learning planning.

Deep Learning and Machine Learning recognition are two different types of recognition

Deep Learning is different than Machine Learning.

Machine learning uses mathematical techniques and probabilities to calculate what are most likely characteristics. The simplest way to think of this is average. I may average a set of numbers and if it goes above a certain number take an action or say I recognize this or that. Machine Learning is simply applying complex mathematical formulas to find correlations.

You may not know but there are many mathematical analysis techniques we have perfected to find statistical correlations and abnormalities in data.

Certainly we do things like this as humans. If I see a certain horse win a lot I might think it is a good bet. You might call that intelligence but it is simply applying math. A computer can apply math faster and it might be able to compute many different correlations on data that we might call intelligent. In fact, it could very easily see correlations we don’t see and seem almost miraculously intelligent.

In my mind this is the same as using a big machine to do labor. It seems miraculously strong and better than a human but it can only do what we tell it to do.

I could compute all the math to do this but the brain uses more abstract things combined with things like average to produce a guess. Our brains find characteristics of the problem that math alone can’t see so easily.

This is why Deep Learning has worked. Deep Learning works by simulating what we think of as a network of neurons. However, that is a little grandiose way of thinking about it. Really what is happening is that we are performing a convolution on the data to normalize it along different paths and then applying these in different ways. These mathematical functions are then compared with actual results and we find which correlations of which convolutions results in better recognition.

We don’t know if this is what neurons do but it has proven to be a better way to recognize some things especially sensory data like vision and sound. Deep Learning techniques have gotten more and more sophisticated and they work far better than ML(Machine Learning) approaches to this kind of data.

However, not always. Sometimes machine learning produces better results. Sometimes different variations of neural networks work better. We don't know apriori ahead of time which approach will work better for which kind of learning problem. Therefore, the craft today consists of different libraries that you try and tweak to find which learning technique works best for the problem.

This is another problem with the AI approach we have today. There is no “standard” neuron model or ML approach that works uniformly across data sets or types of input. That means AI scientists have to try different things to find what technology or cellular network or ML technique produces the best results.

We have only one brain. The brain works the same for vision, smell, the ears or other sensory inputs. The brain is remarkably uniform in structure and it has been postulated that you could hook up virtually any input source to a brain and have it learn it.

For instance, if we had artificial eyes that could see a much wider spectrum of light our brains would undoubtedly learn how to process that input and create a world view and make conclusions and do things using those eyes as easily as the eyes we have today.

Our brains are extremely plastic to the type of data it gets and plastic about the type of learning. Our brains have a powerful general purpose recognition algorithm that works with virtually any type of data and develops many levels of abstractions quickly and reliably.

The proof is in the various types of senses other creatures have and how their brains adapt to it using the same basic building parts. Also, we have seen plasticity of the brain when we substitute different electronic senses for failed vision or other senses.

What this points out is that our algorithms are still not general learning ( or recognition ) algorithms but instead are still immature.

AI is very slow and reaches a peak and stops or even degrades as we try to learn more.

The recognition we have today with AI and ML is immature because we have to use millions of data to train the simplest models and if we put in more data the models can get worse and start to diverge. We generally have to label the training data carefully or the AI won’t learn.

This need for copious data is a serious problem and AI scientists spend the vast majority of their time working on the data and picking the features and labels to assign.

There are some Deep Learning cellular models which don’t need to be trained with labeled data. Unfortunately, this doesn’t produce the best results and leads to some grotesque errors in cases.

In the following article we see how an object recognition system puts a hand with the object it is recognizing. Literally it considers the hand holding the object and the object itself as being one and the same.

It is important to note that recognition is a useful tool by itself even if it isn't perfect

The fact is a lot of human jobs are tedious and don’t require creativity. They are very likely to be able to be supplemented or even replaced by robots or computers.

If I can speak to an AI about my flight and it understands basic commands really well maybe I don’t need to talk to a human often. That’s a good thing as far as efficiency. Humans may still play a role but they can delegate a lot of activity to a computer just as we do today with robots making cars.

Thus, the recognition being perfect or being human level AI is not that critical a need to make this a practical technology to deploy and use.

Whether this reduces the jobs for people enough to cause a labor crisis isn’t clear in my mind yet but it could. We have thought this frequently as we develop more and more machine help for us.

At every stage of new replacements for humans we have always found we move to more personal services and to higher level jobs to create, design, service, customize these robots or AI’s and to use the technology to build more and better vehicles. We then need more managers, more accountants, more artists, more HR people, more programmers, more therapists, health professionals.

It is not clear at all if such changes will continue to expand the labor demand even as we replace more and more millions of other jobs as we have in the past. Later I will explain an area of AI that may generate a lot of additional jobs for people.

Deep Learning is not plastic or has cross skill capabilities

I mentioned this before but what I am specifically referring to here is the idea that a concept learned in one area is transferable to another area.

I can recognize a voice, tie it to a picture of a person and to other characteristics of a person. The same brain can tie these things together and use the recognition of one thing to reinforce the other.

Computer AI does each specific domain without regard to another. Thus a vision algorithm might recognize objects in general but be terrible at facial recognition. A facial recognition algorithm would do poorly to recognize sounds and also wouldn’t have similar abstractions of voice to face.

The first problem is that each of the abstractions created by each AI is specific to the domain that it learns. Don’t confuse it with other input or it will not understand. A vision algorithm may see an eyebrow but it wouldn’t be able to tie the written or audible word eyebrow to the visual recognition.

In fact like in the object case what we call a toy might have an abstraction in a vision system that contains part of a persons hand. There is no abstraction in words for this. In other words the abstractions that each learning system creates is not necessarily the same as another system in another domain would recognize and thus the two aren’t tied and can’t reinforce each other.

In other words the domains are not able to cross learn. This is a huge deficit for AI learning and limits its usefulness.

It further distinguishes from the human method of learning which leverages this cross learning to facilitate learning large number of things in single steps.

Like in the case of our planned learning of arithmetic which allows a human to learn math infinitely better than a computer, learning cross domain and correlating different domains allows a human to make leaps of learning at millions of time the speed of a machine learning algorithm.

The ironic thing is how computers are able to do complex math calculations in a trillionth of the time of a human brain but learning to do simple addition using current techniques is beyond it's ability.

I can for instance learn a concept from reading which I then apply against thousands of memories to suddenly make sense of and add to my understanding and recognition of past things as well as future things much faster.

This could be compensated for by a computer system because we could apply millions of data points in a second with enough computers. But without the cross domain correlation certain types of learning and correlation simply can’t happen at all.

I learn that a certain person has a disease maybe orally from a friend. When I see the person with the disease I operate differently because I know he has that disease. The context of that knowledge informs my behavior. A computer will not have the ability to contextualize and to use abstractions learned from other avenues to buttress my recognition or behavior.

Right now cellular neural models are limited to 2 levels of abstraction. This means that learning higher order abstraction which enables faster learning but also allowing a human to contextualize information to create higher levels of information is not possible yet.

Today a computer can recognize a face for instance. It has no way of knowing a face is part of a body or a body is part of a crowd or type of crowd. It can't work its way up from the pieces of a car's engine to recognize the engine.

A human learns maybe hundreds of layers of abstractions. We understand a car, different kinds of cars, different specifications of the cars mean different performance characteristics, means how it will ride, how fast it will get us from a to b or safely, These aren't useful or necessary for our AI today but it points out the limitations.

It is not surprising

A lot of what humans learn is dependent on massive amount of cultural information. I may see a caricature of that face in another context I recognize it as the same person. Even if a person is wearing a disguise his voice may clue me in. This kind of information would be impossible for a computer to recognize without absorbing years of cultural training.

We don't train computers for years and even if we did it wouldn't make the cross domain abstraction correlations needed nor would it be able to develop the levels of abstraction from such large data sets. However, it is instructive to note that humans don't become super smart instantly. It takes years of training a babies mind to recognize basic sounds, visual things for the brain to then absorb the higher level abstractions. It's a little unfair to an AI to give it a few million data and expect it to "learn" at the level of a human.

The fact it does as well as it does is remarkable. Given no context it does seem to learn but it is important to distinguish it is different and precedents to the learning are vastly different and therefore the method of learning is different and the results will be different.

For certain recognition a computer might be a lot better eventually. For instance, recognizing fingerprints was mastered a while ago but facial recognition is getting better where computers can recognize a face from a huge database of faces and find criminals in a crowd faster than any human could be expected to do.

The point is that computer learning and human learning are vastly different and will probably end up supplementing each other more often than replacing one or the other.

The trick will be to learn the limitations and benefits of the recognition we get from computers and applying it to help people or companies better serve customers without becoming evil.

Bias in Models is extremely thorny problem and has to be solved

Now I am starting to discuss the problems not of the AI itself but of how it is applied that can be problematic and become unethical or evil.

One of the things we notice is that computer algorithms for recognition can have similar problems to human learning and stereotype building.

There was a story about an AI that learned to talk racist easily. We have seen some examples of loan scoring AI’s which might be racist using inappropriate criteria for making a decision.

Facial recognition systems show similar bias when trained on faces that are proportional to the population in the US it may appear to be racist in not recognizing certain races very well leading to unequal outcomes.

These things would generate massive legal problems for firms employing AI.

The problem is there is no way to say teach a computer NOT to be a racist.

Racism is a higher level abstraction that computers don't understand and therefore can't be taught to recognize. If we could teach an AI this then maybe it could recognize when it was being racist.

Testing an AI to insure it doesn't produce racist results is difficult. You have to think hard about the data you use to train the AI to make sure it is getting balanced input but even though you may do this it may still become racist or biased because of issues that are difficult to correct for without explicit algorithmic override.

There is also a concern about privacy.

AI algorithms are trained on massive amounts of data. If that data is supposed to be confidential personal information it can lead to problems that the AI is learning things from personal information it shouldn’t.

It may be necessary for AI technicians to remove data from training data for customers that ask to have their information withheld due to privacy issues. AI’s may have to be retrained without the private data. This may produce worse results but also for firms auditing the results how do they know if the AI learned about specific people or specific characteristics that related to private information that it shouldn’t learn?

The minefields for AI may be huge and difficult to fix regarding things like bias.

Suing an AI

One issue that has come up is what if I get into an accident because an AI was defective?

What if I don’t get a loan? Can I sue if I believe the AI made a decision based on racist or other illegal criteria?

Somehow organizations will not only have to train AI’s to make good decisions but those decisions cannot have inappropriate bias just as we train humans to be unbiased can computers be “trained” to be unbiased? How would we know?

We could ask corporations to produce reports of decisions based on racial or other criteria. We could set up trials where computers are asked questions that appear to produce biased results. How would a company defend its AI?

There are 2 strategies. One is to track the data carefully that was used to train the AI to demonstrate no biased data was put in that could have leaned the AI to a biased learning. More relevant would be to have datasets of all the different classes of people that we want to protect and make sure to run AI’s through scenarios with all those candidates to make sure they aren’t reflecting illegal bias.

It is hard to ask an AI questions about its decisions

Another approach is if we have a way of knowing why an AI made a certain recognition or decision. If we can ask the AI how it made its decision we might be able to figure out if it is biased.

However, AI’s don’t operate this way. They don’t answer such questions easily unlike a human who can articulate why he or she made a decision a computer may simply say, my algorithm recognized this person as risky to buy a car. That’s not going to be good enough answer.

This means that training AI’s and the data used to train them as well as being able to determine what the AI learned and why it made a certain decision is critical. This implies a significant amount of regulation is ahead for this industry.

If a car gets into an accident and kills someone I can imagine a lot of second guessing.

Did the AI not see something it should have seen? Did it make a decision to save one person’s life to kill another? Was there a way to avoid the accident that the computer didn’t see or wasn’t programmed.

All these scenarios will require the NTSA in the US to do detailed studies of accidents and why a computer made a certain decision. Car companies may not be able to rely on purely learned algorithms to make a decision.

For instance, if the computer sees two obstacles to make sure the computer always chooses the non-human obstacle it may need significant if-then-else type programming that is deterministic instead of more fuzzy learning algorithms that might be better at generally driving and be easier to program than a purely declarative programming of all possible scenarios with the actions determined beforehand by programming.

The latter approach has the advantage that a car company could say the car did X because we programmed it to. Having the autonomous vehicle surprise the car company by making a bad choice because of flaws in the learning model would be very risky just as bias can be a problem.

Companies will need to show all training data and what cars in similar situations in the training data did. They will have to prove they trained the car in such situations and prove it learned appropriate behavior.

If you think about the complexity of all this my concern that AI is going to be removing jobs from the US seems misplaced. We may have millions of people employed auditing and checking and performing lawsuits and discovery for every AI we build that makes the jobs lost by the AI industry look minuscule!

I have shown a large number of limitations of AI that doesn’t necessarily reduce the usefulness of the technology or how awe inspiring it can be but shows that in many cases today’s AI technology has some severe limitations people need to really be careful about.

Planning is not learned. AI’s are really bad at some things especially things that have multi-step reactions that really need to be learned as “planning” not as recognition of underlying sensory data.

Very slow learning. Computer AI’s take millions of examples and frequently need to have finely tuned labeled sample data to produce good results. This makes it computationally difficult.

Over-learning. AI’s can get too much data and get worse if you keep trying to train it. The AI does not recognize when it has learned something.

No Cross Domain learning. AI’s aren’t able to do cross domain learning and leverage information gained in one domain to supplement learning in another domain. No AI can leverage multiple domains simultaneously and produce a common set of abstractions across multiple domains. This dramatically limits recognition ability and the speed of recognition.

No memory. AI’s today don’t have memory of their learning or have an aha type experience. When I learn something I know the moment I “get it,” usually. Even more important I associate contextual information with the learning experience. I discover a loved one is sick. I might remember who told me, the tenor of their voice, the feelings that came over me. Music that was playing at the same time.

No motivation. AI’s aren’t motivated to learn one thing over another. They have no way of prioritizing learning without human intervention. Thus they will take random features that may be purely coincidental and determine they are important.

For instance, the learning of objects including the hand holding the object demonstrates how the learning lacks motivation, emotion or purpose. I call these other qualia part of consciousness. Without purpose a computer cannot evaluate what criteria might be important or not.

These things seem insignificant but they relate to the motivation and prioritizing of the learning that allows me to recall and use that information to learn other things. This is how I use memory. Memory for our consciousness is recursive. I remember remembering. I remember a concept and the emotions, the state of the world and everything else. How this helps humans learn better is clear but it is also a distraction.

We are motivated learners, computers aren't. They don't have a need to survive or get food or get a mate. They don't have to worry about predators. They don't have to worry if the result of their learning means they will die.

I might notice that people aren’t sad at the persons sickness and figure out the person isn’t well liked. The point is that memory, emotions, context help a human decide on the importance of the information contextualize the importance of the information in my learning.

Computers make no such importance and thus their learning is not motivated learning and they have no priorities of things to learn. They learn based purely on the coincidence of data points. Thus they pick features to abstract that are not motivated and could be surprising or stupid or brilliant. It’s different.

This last point is probably one of the crucial aspects of human learning vs computer learning we haven’t figured out. A human is motivated to learn things which help them feed themselves, procreate, keep themselves healthy and safe. An underlying motivation drives our learning and planning activities to insure we survive.

A computer has no survival instinct. It’s learning is purely as a mathematician looking at data in a completely objectively dispassionate way that doesn’t prioritize features based on importance to a goal. That makes the learning a lot slower I believe.

Other aspects of human learning that computers don’t mimic include qualia. Humans for some reason associate innate beauty with some combinations of sensory inputs. Computers see no difference or experience no sense that anything is better than another.

Why colors seem beautiful or sounds are more pleasant or odors produce ecstasy is unknown and its importance for driving human behavior and learning is not understood but it’s not clear a computer would ever understand beauty unless we were able to algorithmically describe beauty to it.

There are things like humor which have been inscrutable to computer learning. The DOD, NSA has been trying for a long time to figure out how to recognize the reference to a bomb in a letter as a malicious thing versus as a humorous device. Humor is a very creative act and computers seem incapable of recognizing humor or replicating humor in particular.

Summary

Computers are getting better at recognizing but it is important to understand their limitations and liabilities as it is to understand what they can do.

A lot of people seem over-enamored of the possibilities of AI without understanding how severe the deficiencies of our current technology are.

It is likely some of the problems I describe above might not be solved or mitigated soon and we will run into another "dead-end." If so, I am sure in another decade we will come up with a new idea but the past is indicative that it is more likely we will find we hit a roadblock.

The difference is that AI has become useful at the level it is today therefore it won't go away but progress may be very slow.

A critical problem is the 2 abstraction level limit which seems to be hard to break. Our best Deep Learning AI’s can do 2 abstraction layers. That is they can abstract up an eyebrow from a set of pixels. Another layer can abstract combinations of eyebrows, lips, noses, ears to see a happy face, an asian face, even Robert’s face but going another level up to abstract at higher and higher levels has been impossible.

The AI’s become worse and lose their learning if you push them with too much data or try to make the learning too high level. Different algorithms work better in different domains. The models don’t cross domains making higher level abstractions hard.

We have no easy answers to these problems. It will require another step forward in AI similar to the jump in 2010 when the latest neural models advanced to solve any one of the problems I have talked about so we are not on the verge of suddenly making a leap to human level intelligence.

It also means we may find that we simply can't get beyond a certain point recognizing voices, text, objects which limits the practicality of the technology.

I don't want to end on a negative note. This technology is making improvements in a more consistent way than the past. The applicability of the technology has reached a crossing point where it is useful for lots of tasks. It seems quite likely that cars will be able to drive autonomously with the level of AI we have today.

So, the usefulness of this technology is not in doubt. The problems I have surfaced around how to prevent it from causing problems in our society or to insure its quality when it is deployed to life and death situations are daunting but not insurmountable.

There are big differences and a lot of reason to believe computers will not achieve truly human level intelligence in the next several decades.