(a series of posts for each chapter)
<<< Previous post (“Rebooting AI” notes-1)
>>> Next post (“Rebooting AI” notes-3)
In this post I would like to list some quotes from the book which I found them either important or interesting. (Note: some parts of the quotes are paraphrased by myself and some of them are my own insights)
2. What’s at Stake
A lot can go wrong when we put blind faith in big data (Cathy O’neil, TED Talk, 2017)
Despite AI is giving very promising results in various field, from face detection to realistic and high resolution face generation and many more, we can never know when it going to surprise us with unacceptable errors that are nonsensical or even dangerous (recall AI being utilized in self-driving car projects).
However, the above concerns does not imply that we should worry about AI all the time. The point here is to be aware of what AI can do (at least for now) and can’t do, and to see the current state of the AI development. This is important for both misleading the people about AI development and to see what we have to do next to further improve current AI to be more robust and reliable.
Trending AI Articles:
1. AI, Machine Learning, & Deep Learning Explained in 5 Minutes
2. How to train a neural network to code by itself ?
3. Are you using the term ‘AI’ incorrectly?
4. Artificial Intelligence Conference
The following technologies could give some intuition;
- AlphaGo: simply doesn't care about questions like "Is there life outside the Go board?, "Is it fair that my masters leave me to do nothing but play Go all day?". AlphaGo literally has no life or curiosity at all beyond the board
- Tay chatbot: released by Microsoft in 2016. Less than a day later the project was canceled, simply because a nasty group of users tried to drown the chatbot in racist, sexist, and anti-Semitic hate.
- Face recognition system in China: sent a jaywalking ticket to an innocent person, who was a well-known entrepreneur, when it saw her picture on the side of a bus. Not realized that a larger-than-real-life photo on a moving bus can not be real person.
A catastrophic misinterpretation could cause a troublesome result when your robot takes your requests too literally. Assume you have been told to your robot to take everything left in the living room and put it away in the closet. And you may see, when you come back to your room, everything — the TV, the furniture, the carpet — broken into tiny pieces to fit into closet.
- We want our robots and AI to take us seriously, but not always literally and we need to be careful in trusting AI too much until it gets radically better.
Machine Learning depends on the precise details of large training sets and thus often fail when they are applied to new problems
- Machine-translation system trained on legal documents do poorly when applied to medical articles…
Yet another challenge for current AI is that it may end up with the wrong goals, i.e. finds a shortcuts to achieve a predefined goal:
- Robot which supposed to grasp objects by training on images eventually finds out to solve the problem “just” by putting its hand between the camera and the object.
- AI software which tasked with playing Tetris decided to pause the game indefinitely rather than risk losing.
- These kind of failures show us the importance of finding out a proper evaluation metrics for different types of tasks, so that we can expect an AI to solve the problems without shortcut solutions which are useless for us.
None of above mentioned challenges don’t imply that AI can’t do better. However, we really need a new approach altogether.
The new approach might consist of:
- feedforward models (like Deep CNNs) for feature extraction
- feedback mechanism (like RNN, LSTM, etc) for processing sequential data
- generative models (like VAE, GAN) to understand the data distribution, thus understand the nature of the data
- knowledge transfer (like neuro-symbolic computation) to transfer the previously learnt knowledge into new tasks — commonsense. Note that this is different from so-called Transfer Learning.
Don’t forget to give us your 👏 !
“Rebooting AI” notes-2 was originally published in Becoming Human: Artificial Intelligence Magazine on Medium, where people are continuing the conversation by highlighting and responding to this story.