12
$\begingroup$

I have a background in Computer Engineering and have been working on developing better algorithms to mimic human thought. (One of my favorites is Analogical Modeling as applied to language processing and decision making.) However, the more I research, the more I realize just how complicated AI is.

I have tried to tackle many problems in this field, but sometimes I find that I am reinventing the wheel or am trying to solve a problem that has already been proven to be unsolvable (ie. the halting problem). So, to help in furthering AI, I want to better understand the current obstacles that are hindering our progress in this field.

For example, time and space complexity of some machine learning algorithms is super-polynomial which means that even with fast computers, it can take a while for the program to complete. Even still, some algorithms may be fast on a desktop or other computer while dealing with a small data set, but when increasing the size of the data, the algorithm becomes intractable.

What are other issues currently facing AI development?

$\endgroup$

3 Answers 3

7
$\begingroup$
  1. we don't really know what intelligence is.

  2. we don't truly understand the best model of intelligence we have available (human intelligence) works.

  3. we're trying to replicate human intelligence (to some extent) on hardware which is quite different from the hardware it runs on in reality.

  4. the human brain (our best model of intelligence) is mostly a black-box to us, and it's difficult to probe/introspect its operation without killing the test subject. This is, of course, unethical and illegal. So progress in understanding the brain is very slow.

Combine those factors and you can understand why it's difficult to make progress in AI. In many ways, you can argue that we're shooting in the dark. Of course, we have made some progress, so we know we're getting some things right. But without a real comprehensive theory about how AI should/will work, we are reduced to a lot of trial and error and iteration to move forward.

$\endgroup$
3
$\begingroup$

I am assuming by AI you mean AG(eneral)I, not machine learning or expert systems tuned for specific tasks.

In addition to @mindcrime's answer, sometimes we run out of samples to train and sometimes computers became so slow to process enough samples to work in manageable timescales. @bpachev mentioned memory but on the surface, our supercomputers have more than enough memory to store a human brain matrix. But we lack the ability to simulate it real time. After we are able to do that, we also need to connect external input, even more processing power is required for that. Even that would not be enough to simulate a human brain fully as biochemistry plays an important role.

One final note would be there is little incentive to develop AGI other than understanding how the human mind works. There are classification algorithms, expert systems, knowledge engines that can out-perform even the best humans for specific tasks.

$\endgroup$
4
  • $\begingroup$ Regarding your last note, you seem to imply that understanding the human mind and human cognition is of little value outside of itself. Couldn't AGI systems be used to help treat patients with considerable brain damage and memory loss or with interacting the brain with mechanical appendages? It seems that if we can create a truly genuine AGI, we will have answered one of the questions that has plagued mankind for millennia and would have countless applications outside of just an intellectual pursuit. $\endgroup$ Aug 9, 2016 at 18:58
  • $\begingroup$ But I do agree with you that specialized expert systems are invaluable resources to us apart from creating AGI. $\endgroup$ Aug 9, 2016 at 18:59
  • 1
    $\begingroup$ Understanding how human mind works is not only intellectual pursuit, it has medical implications as you stated. But, not just brain damage and memory loss, but psychological as well. Solving PTSD or OCD will have great positive effect on society. On the other hand, it will have military applications. Why interrogate someone when you can extract the info from their brain, right? $\endgroup$ Aug 9, 2016 at 20:34
  • $\begingroup$ This seems like a great topic for another question: What are the ethical implications for creating a genuine AGI? $\endgroup$ Aug 9, 2016 at 20:35
-2
$\begingroup$

One obstacle to the development of AI is the fundamental limitations of computer memory. Computers, at a fundamental level, can only work with bits. This limits the type of information that they can describe.

EDIT:

The precise nature and complexity of human memory isn't fully understood, but I would argue that at the very least, human memory is well adapted for the types of tasks that humans perform. Thus, computer memory, even if theoretically capable of representing everything that human memory can, is probably inefficient and poorly structured for such a task.

$\endgroup$
3
  • $\begingroup$ Neoromorphic memory indexing schemes, such as Kanerva's 'Sparse Distributed Memory', have been around for a while. $\endgroup$ Aug 6, 2016 at 17:35
  • 1
    $\begingroup$ You'd have to be more specific about what you mean with "The human brain encodes information with chemicals". Information is encoded in the brain through changes in the neurons connections to one another and not through the neurotransmitters themselves. However in the end it can be loosely modeled as being a single "weight" value per connection, thus the idea that chemicals encode information in a more complicated fashion than bits would appear to be incorrect. $\endgroup$
    – tomzx
    Aug 7, 2016 at 13:37
  • $\begingroup$ @tomzx Thanks for your feedback. It was my understanding that human memory isn't well understood, but that due to the chemical nature of the brain, memory was represented chemically. I'll edit my answer to be less speculative. $\endgroup$
    – bpachev
    Aug 8, 2016 at 17:27

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .