Article at this link https://www.quantamagazine.org/been-kim-is-building-a-translator-for-artificial-intelligence-20190110/ talks about "A New Approach to Understanding How Machines Think". It says in intro:
"Neural networks are famously incomprehensible — a computer can come up with a good answer, but not be able to explain what led to the conclusion. Been Kim is developing a “translator for humans” so that we can understand when artificial intelligence breaks down."
Are you aware of other research and researchers doing similar work? Can you share links, resources and/or your research on this, please? Thanks!
Please take a look at this paper, it looks like a good approach to this problem:
Conference Paper Toward Explainable Deep Neural Network Based Anomaly Detection
"May not machines carry out something which ought to be described as thinking but which is very different from what a man does?” (Turing, 1950)
https://www.bbvaopenmind.com/en/articles/the-future-of-human-machine-communications-the-turing-test/
Machines do not think; they compute. Computation is a science. So your question is too broad, I'm afraid, to warrant an answer. Instead, a textbook on AI and/or Cognitive Science could be a starting point.
@Dennis Mazur, @Spyridon Revithis
I appreciate that title of question can be misleading a bit. I was hoping that people will read article at link I posted in description of question, first, and it will be clear. Let me explain in more details. Also, I put word "think" under quotation marks now to indicate the difference.
Question is about next issue: AI based system can come up with a good answer, but not be able to explain what led to the conclusion. It is some kind of "black-box" for everyone outside. As an example, imagine AI based system which gives disease diagnosis based on set of data, but without explanation how it came to specific conclusion, vs. medical doctor, who explain it to patient in details. AI will just "fire" an answer with no explanation. It can lead to lack of trust to AI systems even they can be more accurate in diagnostic process than humans do.
In article which I posted link to, it is described how Been Kim, a research scientist at Google Brain, is developing a way to ask a machine learning system how much a specific, high-level concept went into its decision-making process.
So my question is aimed to get info about another research in that area.
Hope this clarifies question much better.
Thanks to everyone for contribution.
I just read the article and interview with Been Kim; I would like to point out that there's a flaw in the reasoning behind vectors with regards to the explanation on how an AI system made a decision.
What it is described is not that, what it is described is an evaluator function of how important a certain set of input features was in the decision. This leads us to a paradox because we only can evaluate what we already know (the doctor will ask whether a certain feature that he would have looked for was considered - he could describe the feature and ask the AI model whether this feature was taken into account). But, what if the AI model is taking into account some new pattern not known to the doctor (either because it's new to him due to his experience or even new to medicine)? He would ask and would get partial answers to what led the system to diagnose that way but just because he wouldn't be asking the right questions.
In my opinion, AI systems shouldn't be 'interpretable', they should be 'self-explaining - they should collect the set(s) of features which balanced their decision towards a certain answer and be able to expose them if asked for. Going on with the doctors/medicine scenario, that's how a group of doctors would behave... let's say one of them diagnoses a rare illness for which the other doctors don't know the symptons and/or test results involved, they could ask once and again to the diagnosing doctor: 'did you look at this feature?' 'did you look at that?' 'was this important?' 'was this not?' ... but let's suppose that the deciding feature (either a sympton or a test result or a combination of both) was so rare as to not being asked for and all other symptons/test results would point to a different variaty of illnesses. Unless the diagnosing doctor would point out the features he took and the ones he dimissed and why they were taken or dimissed, the other doctors won't learn from him how to diagnose the next patient with that illness.
This should be the goal... having 'self-explainable' AI systems that can extract 'hard' and 'very-hard (add as many very-very-verys as you like... ad infinitum)' to detect patterns in the data and help humans learn from those NEW patterns and reach new levels or knowdledge and understanding, which at its own time will lead to new discoveries, development and innovation.
Machine thinks based on defined pattern and logical intelligence with expected and actual result in binary 0 or 1.
Preprint Intellectual activity, knowledge, information, data... An at...
Agree, the digital ones anyway do not think, they calculate.
Thus in advanced mode always beat me at chess. They calculate all the moves. The end result is that you think they think.
You might make some reparation if you have programs that program or alter themselvs, that is some analogue of brain plasticity.
Then you have survivalist programs, where anything that does not work gets discarded. Perhaps that is mimicing the evolutionary process.
So far you cannot get near the number of neurons in the brain, and work more in the analogue mode. Neuronal networks can be simulated on a computer.
- Badges
- Science topic
- Similar topics
- Cognitive Science
- Thinking