Dumb Machine That Thinks
In the 21st century, with increasing computing power, the field of Machine learning has flourished vastly that few could think in the past century. Artificial Intelligence has become the buzz word now. This might be a common question for an AI enthusiast, “how much time do we need to get a complete AI?”
You may have been overwhelmed with the recent scam of Sophia, so-called AI-powered robot that can interact with human like a real person proclaimed by its developers. But according to many researchers, it was just a chatbot with a human face. According to many AI scientists, still, we have to go miles.
CURRENT LIMITATION
According to many experts, our works haven’t reached the real Artificial Intelligence. What we see today, we can refer to it as Artificial General Intelligence (AGI).
“We still don’t have any real AI in the sense of original vision of the founders of the field, of what I think you might refer to as AGI…” — Joshua Tenenbaum
To create real AI, we have to know about the impediments. Two major problems for the goal are:
- The computer still can’t think “why”
- Single Model can’t learn everything
In the past century, maybe one of the biggest events in the field of computer science was the chess match between Deep Blue and Grand Master Garry Kasparov. Deep Blue was a chess-playing computer developed by IBM and it won a chess match against the world champion Garry Kasparov. According to Jeff Hawkins, Deep Blue didn’t win by being intelligent; it won by a million times faster than a human.
Deep blue has no intelligence, but still, it managed to beat a world champion. Even though it was a dumb machine, it won by the brute force approach. An intelligent human plays chess by intuition. Can we mimic intuition in the machine just like a biological brain?
Intelligence is important not only in playing chess, but it is also necessary to answer some questions after reading a book. Google tried to replicate this sense in the machine and but results were not something we can tell it has come from any intelligent being.
Today’s AGI can’t understand by reading a story. It can, however, tell if a story is a fiction or real. This is a kind of classification problem that can be solved by fitting a neural network with an adequate amount of labeled data. However, this does not mean that it is giving the computer the ability to read like a human.
HOW FAR IT MIMICS BRAIN
We recognize objects around us in real-time. Nowadays the computer is also able to do that in real-time. With the YOLO algorithm computer can process images in 45 FPS and recognize objects from real-time video data. Another development of this model called Fast YOLO can do that in 155 FPS. YOLO is based on Convolutional Neural Network (CNN) which is inspired by the brain’s visual cortex.
The computer can also recognize document type automatically. A computer’s ability to classify a document or recognize an object in a picture or video doesn’t require any type of intuition. To understand how a computer can do this we just need some superficial idea about training a neural network.
Generally speaking, training a neural network model is involved with tuning the weight matrix or we can say, finding a perfect matrix for a specific task. If our model is multilayered then each layer will contain a weight matrix. In the beginning, we generate those matrices randomly.
And the model actually learns through backpropagation. Backpropagation involves updating the weight matrices. It does the job by calculating error at each iteration by a loss function. If you are new to this concept then you can consider it as a way of telling the neural network to correct itself after each iteration. This is how a neural network model learns from its mistake at every iteration. In deep learning backpropagation has bought tremendous success.
Now the question is, does the biological brain do anything like backpropagation? Researchers are somewhat skeptic about that. According to Yoshua Bengio, maybe brain is doing something similar — not exactly the same, but with the same function.
It is also true, in the way of creating an intelligent machine our goal is not just mimicking the brain. In some cases, machines can show impressive benefits in its way.
I have experience in automatic document classification task with the deep neural network. I had created a neural network model and trained it with forty thousands of news articles of twelve different classes. I saw a clear difference between biological and artificial neural networks. Can you read an article at once? If you try to classify an article by yourself, maybe you have to read a few lines sequentially, then you will understand whether it belongs to politics or sports. Our brain works in a sequential fashion. No matter how fast we can read, we can’t read it at once. But the trained artificial neural network can read the article seemingly at once and assign a class accurately.
TEST OF INTELLIGENCE
There are some tests for human-level AGI. Turing test is widely cited and known test. The Generalized statement of Turing test: we would know machines can think, when we couldn’t tell if we were talking to a computer or a man, provided a wall or a door prevented us from seeing which was which. To put it simply, if we think the machine is intelligent, then it is intelligent. Chatbot developers showing have shown considerable progress in this work in recent times towards passing Turing test, but many researchers and AI philosophers argues that the test is not enough to label a machine as intelligent.
“The whole concept of the Turing Test itself is kind of a joke. While it’s fun to think about, creating a chatbot that can fool humans is not really the same thing as creating artificial intelligence. “ — Mike Masnick
The Robot College Student test
This Test is proposed by Ben Goertzel, which can be identified as a more challenging operational test for AGI.
“As opposed to the Turing test of imitating human chat, I prefer the Robot College Student test: when a robot can enroll in a human university and take classes in the same way as humans, and get its degree, then I’ll consider we’ve created a human-level artificial general intelligence: a conscious robot.” — Ben Goertzel
My personal idea to solve this problem to create a bot that will enroll in an online course in Coursera, say, Machine Learning Course by Andrew Ng. Then it will go through all video lectures, take the weekly quiz and submit programming assignments. After finishing the course it will earn a certificate then we can say it is no more a dumb machine, rather an intelligent one.
CONCLUDING REMARKS
This article is a kind of personal voyage in studying artificial intelligence. I’ve found that many bright researchers are working with passion and patience to solve interesting problems of Artificial Intelligence, like, Yoshua Bengio is trying to make computer think about “why”. Many models are proposed that are performing very close to the human brain, like, siamese neural network can learn to differentiate face with a low amount of data and paves the way of solving one-shot learning. We are just searching for the right methodology that will make a dumb machine to think and act intuitively.
