1
$\begingroup$

I heard that your ML model's quality depends directly on the quality and the quantity of data you use.

So I was thinking that can question answers be used as data to train an algorithm which can solve any high school science problems? Because we do have a gazillion number of high school books with millions of Question-Answers which are both high in quality as well as quantity.

P.S: I don't have any in-depth knowledge in any of the AI fields, so please answer accordingly!

$\endgroup$
  • 1
    $\begingroup$ This is a very vague question..try to specify what type of problems ... level of human input..level of abstraction,etc $\endgroup$ – DuttaA Dec 11 '17 at 14:42
  • $\begingroup$ Probably OP heard that AI makes human obsolete so he wants to obtain PhD in AI without learning anything. It is much in agreement with popular version of AI presented in media and by Elon Musk. Maybe even it is Elon who ask? Who knows? $\endgroup$ – kakaz Dec 11 '17 at 19:24
  • 1
    $\begingroup$ @kakaz, I would love to do PhD, but given I am a high school student, I have some more important short-term goals :) $\endgroup$ – Richardo Martinez Dec 12 '17 at 4:53
  • $\begingroup$ Focus on these and everything will be all right! $\endgroup$ – kakaz Dec 12 '17 at 5:08
0
$\begingroup$

The short answer, I guess, is yes. In theory, you could have a massive, insane network that could take as input any high school science textbook question and output an answer. But think about what that entails:

Take in a question, in any of the various languages that high school science textbooks can be written in. Get all the relevant information out of that. This includes questions like "List four noble gases" as well as ones about applying Kirchoff's laws to a circuit as well as explaining how the ribosome changes its shape to produce proteins. There is soooooooo much information that is contained in just the wording of those questions and so many topics. Then, let's say you have converted the question into a representation that can be "understood" by the algorithm. I want to emphasize that ML models do not "know", "understand", "think", or any other of the words we use to personify them. They are basically just a mathematical function of some level of complexity (usually very complex). That being said, now the algorithm needs to relate this tractable representation of the question to the correct answer.

Since everything before this created some representation of the question, you could view that as a function mapping the input to this representation. Then, everything between intermediate representation and output answer is another function. So, your model needs to learn a single function that is the sum of all human high school level science understanding.

These algorithms you want to build would end up being staggeringly complex. In addition, as you add complexity to an ML model, you need significantly more data to train it. How will you take every high school science book in the world and convert it into digital pairs of questions and answers? How will you handle that some of the answers are incorrect? What is the benefit of this system?

The long answer, I think, is no. It makes no sense to do this. Instead of learning this complex mapping, just build some solvers with equations built in that the user inputs some numbers to and selects the equations. It will take so much less time to build and deploy and I guarantee work better. There are people building ML systems that are trying to get the SAT entirely right (see page 33 of this report). First, I don't think this is what you're asking and second, it still makes more sense, in my mind, to build some solvers. Hopefully, this all makes sense and, if anyone sees things I misstated, let me know.

$\endgroup$
  • $\begingroup$ First of all, thanks for the answer. It is amazing! Second, you asked what is the benefit of this system? Let us consider subjects like Inorganic chemistry, Zoology, Botany, History, Civics and Political Science ( pick up one subject out of these ). In these subjects, ( at least till the high school level ), there is no IN DEPTH understanding required, as they're mostly facts. Now, if a software has mastered any of the subjects, we could use it as a kind of teacher for the students. Think of it like " Siri for education ". $\endgroup$ – Richardo Martinez Dec 12 '17 at 5:07
  • $\begingroup$ Even in theory it is completely impossible. $\endgroup$ – kakaz Dec 12 '17 at 5:10
  • $\begingroup$ And students would be much more comfortable in asking questions to a software rather than a teacher because they won't ever be judged. No matter how silly their questions are, they could always ask them and get an appropriate answer, which is better both for their curiosity and their education! $\endgroup$ – Richardo Martinez Dec 12 '17 at 5:10
  • $\begingroup$ @kakaz yeah I do understand. But yeah, it feels nice to think about such stuff you know ;) $\endgroup$ – Richardo Martinez Dec 12 '17 at 5:11
  • $\begingroup$ Things you name as school problems today, yesterday was serious scientific problems, used to build technology, knowledge and the things you use in real life. Thinking it is possible to train AI to solve it is just equivalent to think it is possible to build machine solving all problems of XIX century science say. It is impossible. If you think other way, just check something about just integration and implementation full algorithm of it. Oner side, clearly to may build just big database to check if given problem is routine one, and check given book-like solution. But no AI is needed there... $\endgroup$ – kakaz Dec 12 '17 at 5:19
1
$\begingroup$

Its true that your AI model's performance depends on the quality of data that you use. However, high quality data alone is insufficient to guarantee that your model will learn effectively and score well on a particular dataset. Other factors such as smarter algorithms and the use of high performance computing infrastructure must be factored in for your AI system to perform well.

Although A.I research has made massive progress in the past decade, ML engineers are yet to build a system that can match the general scope and generalization ability of the human mind. Upto the first decade of the 2000's AI was dominated by expert systems that emulated the decision making ability of an expert. AI at this point couldn't process unstructured data and therefore it lacked the capacity to sit for and pass high school exams.

This was until 2011 when IBM Watson a question answering computer system competed against two former Jeopardy quiz show winners and placed first. IBM Watson was built on top of Deep QA (a computer system that could answer natural language questions) and UIMA (a software achitecture to process and analyse unstructured information). Below is a link to a paper giving an overview of how IBM's Watson works https://www.aaai.org/Magazine/Watson/watson.php

In 2012 a team led by Geofrey Hinton won the ImageNet competition by exploiting deep convolution networks. This was soon followed by Dahl's team winning the Merck Molecular Activity Challenge using deep neural network architecture. Yann LeCun's work in CNN's, Geoff Hinton's back propagation and Stochastic Gradient Descent aproach to training datasets alongside Andrew Ng's large scale use of GPU's ignited accelerated progress in ML. This was frequently referred to as unreasonable effectiveness of Deep learning.

Following recent advances in fields such as image captioning, natural language processing, information retrieval and computer vision it is highly probable that current generation AI systems can pass high school exams such as SAT.

The Allen AI Institute has made significant progress in developing AI systems that can read, learn and express that understanding through question answering and explanation. Founded by Paul Allen Microsoft's co-founder, the Allen AI Institute's singular focus according to their mission is to conduct high impact research in the field of AI. Below is a news link covering their cognitive system passing high school exams fortune.com/2015/09/21/computer-artificial-intelligence-math/

So far Allen AI Institute has demonstrated a cognitive platform called Geos that is capable of answering geometry questions as well as the average high school student. While another system called Aristo can answer high school science exam questions by leveraging information extraction alongside knowledge representation and reasoning models. You can access AAI's GeoS service here http://allenai.org/euclid/ and Aristo here http://allenai.org/aristo/

Meanwhile researchers working on the Todai project in Japan have demonstrated a cognitive system that is capable of passing the Tokyo University Mathematics entrance exam. My conclusion from the above examples is that possibly we already have AI that can sit for and pass high school exams.

$\endgroup$
0
$\begingroup$

I'm thinking that you could write an AI that takes the question as input, weights it, and googles info based on the first layer of neurons, then takes the first two to three pages of results and spits out an answer. It would be a crapshoot, but maybe you could take the list of results, choose one using another layer, choose the info from the page using a third layer, then answering the question using the info.

$\endgroup$

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.