6
$\begingroup$

After having read something that Elon Musk said about artificial intelligence and how it could affect our lives, I've been reading about artificial intelligence, deep learning, etc. The recurrent topic is neural networks, which are used for "recognition tasks".

Although this is not my main question, can we say the following about neural networks?

Neural networks use the sigmoid function and gradient descent to fine-tune weights.

In any case, neural networks do not seem to really mimic any of the properties that we associate with intelligent beings, like humans, such as thinking outside the box, being able to observe, reflect and come up with innovations.

So, is there any artificially intelligent system that really mimics human intelligence?

$\endgroup$
1
  • $\begingroup$ Artificial Intelligence alone may mean something different to what a "human intelligence" means. Maybe other inputs (and outputs) like emotions, memories, experiences, etc, etc should be part of (or somehow influence) "intelligence" $\endgroup$ – Raul Alvarez Apr 25 at 11:10
4
$\begingroup$

As far as emulating an intelligent being, no. There are a few different potential architectures for possible AGI. Many of these are extremely infantile, as the bulk of AI research is in narrow AI, which focuses on creating algorithms that are highly specialized for a specific task.

With that being said, here is one supervised learning approach to this problem by Andrej Karpathy: Karpathy Method

OpenAI, also has a team who works on the AGI problem.

$\endgroup$
3
  • $\begingroup$ Great keywords that I Googled: "AGI" and "narrow AI". Thanks. Although after a quick read so far, it makes me wonder why Elon Musk is so bleak. From the looks of it, it's way too early to predict that AI poses any threat - I'll keep reading/Googling. Anyway, I will check-mark your answer in the next day or two to give others time to chime-in. Thanks again. $\endgroup$ – Curious G Sep 27 '17 at 1:23
  • $\begingroup$ Yes, I would assume part of Elon's positions stems from the impression things like deep reinforcement learning leave, where unsupervised(for the most part) agents are learning everything from Go to Starcraft 2 to the stock market. But I do agree his position is mostly unfounded. $\endgroup$ – hisairnessag3 Sep 27 '17 at 1:35
  • 1
    $\begingroup$ I suspect some of the alarm derives that the intelligent systems are getting too complex for human analysis, i.e. we're not going to know what they're thinking. This is a serious problem in regard to "value alignment", particularly in light of the symbol grounding problem. $\endgroup$ – DukeZhou Sep 27 '17 at 18:43
2
$\begingroup$

No, we are not even near an algorithm that can be compared to human level general intelligence. You might have heard the claim that a neural network works similar to the neurons in the brain, but that's quite a stretch. Serena Yeung talks about this in the forth lecture of the Stanford course CS231n. Jump to 1:04:30 in the video, that's around where this part starts.

When it comes to the topic of consciousness and qualia, it is highly unlikely that we do possess the technology to create a conscious AI yet. I say highly unlikely because from a philosophical perspective we lack the ability to detect qualia in another being (called the qualia problem), at least based on our current knowledge. So we cannot prove, that we haven't already created artificial consciousness.

I've also gathered from your comments that you are wondering, why we are afraid of strong AI. I can recommend Superintelligence from Nick Bostrom. He discusses many different scenarios in which strong AI could be threatening humanity. Consciousness is usually not required for those scenarios.

$\endgroup$
1
  • 1
    $\begingroup$ Wow, many more things for me to read about, and thank-you very much for the youtube video. As for the "qualia problem", oh boy, it will take me some time to dive into those depths. I'm more math/logic minded. It's a shame I cannot check-mark 2 answers - both have been on point. $\endgroup$ – Curious G Sep 28 '17 at 16:59
2
$\begingroup$

Regarding "mimicking human intelligence" --> No, not even close.

Regarding this:

More specifically, can we say the following about neural networks?

Neural networks use the sigmoid function and gradient descent to fine-tune weights.

No, again. ANN's aren't required to use a sigmoid activation function. There are many other options that are used depending on the circumstances. Also, gradient descent is not strictly required either, even if it is one of the most popular optimization algorithms used with backpropagation. There are other ways to train ANN's - for example, genetic algorithms can be used.

$\endgroup$
1
  • $\begingroup$ Yes, I agree in hindsight that my "one sentence" fails to be generic when I included specific names of functions. I actually already knew that there are many other functions/tweeks/etc. that can be used. But would you agree with my assessment: The whole world should call it "recognition software" instead? I mean, Does the current AI do anything more than pattern recognition that leads to preforming a related task? Thanks. $\endgroup$ – Curious G Sep 29 '17 at 0:43

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.