What is AI Really? Want to Try AI?
Understand AI and see how it works. Form your own opinion.
Genesis.
AI has been around since 1956 and has gone through several boom and bust phases. It is only recently that PC based platforms have become powerful enough to support the intensive computational load demanded by AI libraries like TensorFlow, scikit-learn, and Pytorch. With my PC I can now explore AI, form my own opinions, and ignore all the hype and misinformation. It’s an excuse to have fun!
When I morphed from a grumpy systems architect to a grumpy lecturer I put together some introductory AI sessions for my students. This story builds upon and summarises some of my presentations. Being who I am, I added in links to AI tools and libraries including getting stated tutorials and setup guides. Try things out yourself!
1956 Birth of AI
Mystified by AI or perhaps just ignoring it? Maybe you have read about killer robots or totalitarian regimes controlling minds? Misinformation, hype and ignorance abound. Let’s start the cure and look at the original definition of “Artificial Intelligence”.
The term Artificial Intelligence (AI) was coined at a 6 week summer Dartmouth College workshop in 1956. The proposal for the workshop stated that a “study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it”.
“Simulate” and “Precisely Described”.
Simulation is the keyword to overcoming all the hype. AI does not have an objective to create intelligence but rather to “simulate it”. The success of the simulation depends upon the precision of the description of human intelligence and the method adopted for validating the simulation.
The Turing test
The Turing Test dating to 1950 was adopted as an early validation technique. The idea is that a human judge would observe conversations between a human and a machine designed to generate human-like responses. If the judge could not tell the difference between human and machine, then the simulation would be considered valid. The Turing test and its variants are still the subjects of much controversy.
The problem with “Precisely Described”.
The phrase “precisely described” implies a reductionist approach to the study of intelligence. However, modern research shows intelligence emerges from the complex electro-chemical interactions between the brain’s neurons. In this view, the brain is a complex adaptive system. Reductionism fails because the emergent properties such as intelligence cannot be reduced. In other words, the properties at a given scale cannot be described in terms of properties at a lower scale.
Complexity theory explains why after nearly 70 years of intensive research booms and busts the optimism of the 1956 participants has not been realised. Furthermore, the phrase “Artificial Intelligence” has been shown to be a misnomer, but the term has been popularised to such a degree that it’s here to stay.
Despite the failure to explain general intelligence, the mathematical tools developed as subsets of AI are extremely useful and now underpin our modern interconnected society. These “learning” tools are excellent for problems involving pattern matching.
The Tools of AI: Machine Learning and Deep Learning
While “Artificial Intelligence” (AI) is concerned with the simulation of human intelligence, the commercially successful tools of AI research focus on data parsing and pattern matching. Together these tools are referred to as “Machine Learning”. Deep Learning is a subset of Machine Learning.
Machine Learning has been widely deployed in fields including facial recognition; cargo distribution; sales and marketing; climate science; social media; and financial trading. Machine Learning algorithms can be implemented in many applications without the need for changing any code. Instead, standard sets of labelled data are used to train its algorithms to recognise desired patterns. These “trained” algorithms can then be applied to new datasets to discover the desired patterns.
The issues with machine learning stem from the design of the training sets. If the training sets are too small or omit vital patterns then they will have poor accuracy and likely to miss-categorise outcomes.
Artificial Neural Networks
Artificial Neural Networks were designed to mimic the electrical interactions between neurons of the human brain (ignoring the vital role of chemical components). The explanation of the relationship between the mathematical tools and the brain’s structure and function are tenuous at best. However, these tools are useful, valuable and in constant development.
In Machine Learning, the set of ANN algorithms consist of three levels: input, parsing and output. The term Deep Learning is applied when a neural network has more than three layers. Some of the modern Deep Learning neural networks have many thousands of layers (refer to Pytorch and Tensorflow). Every deep learning layer can be optimised through a technique called backpropagation.
Backpropagation gives Deep Learning the ability to independently learn and make decisions on pattern recognition. That is, there is no need to prepare training data and its time-consuming classification (labelling) requirements. Deep Learning does require stringent management of data quality, (fitness for purpose) and careful validation of results (can they be explained?). Incremental Learning is a technique for for preparing Deep Learning algorithms.
The AI technologies of Deep Learning are now main-stream tools in IT applications. The European Union has recognised and mandated benchmarks to prove AI products comply with strict ethical and legal standards extant in Europe.
Getting Started with AI: Learn by Doing.
I only use open-source software. You should make sure your PC, laptop or Mac has enough guts to support AI programming. Also, although it’s not needed when you are starting, give some consideration to GPU hardware acceleration (I like Jason Dsouza’s article — see below).
I am assuming you know enough about Python to do some coding — not at an expert level, but more at a “get the job done” level. If you want some self-learning in python:
If you want to try your hand at coding examples and applications in AI then I suggest that it’s a good idea to install Anaconda Individual Edition. Anaconda is a widely used, open source, data science analytics platform. Anaconda will take care of hair-tearing-out tasks like version control. Comes with Python.
AI-Libraries.
These are the AI libraries that I have experience in using and teaching. I’ve included tutorial links and there is documentation on how to load in Anaconda.
1. Scikit-learn.
I highly recommend this library if you want to just get a basic understanding and see how AI works. scikit-learn comes installed with Anaconda. The exercise I really love is an example of classifying flowers using the Iris introductory exercise in scikit-learn. I use the Iris example in my first class session to show students how useful AI can be. The data is simple enough to allow you to manually sort through the Irises to see why AI is popular.
2. TensorFlow (Deep Learning).
I used TensorFlow in my consulting work running in a dedicated Anaconda environment. I can’t recommend Keras enough, especially when learning TensorFlow.
3. Pytorch (Deep Learning).
I like the looks of PyTorch, but have not had a great deal of experience with it.
