What is Human Intelligence, Artificial Intelligence, Machine Learning, Artificial Neural Networks, Deep Learning and Data Science?
The hardware of the human intelligence is the human brain. Neuroscience is trying to figure out the processes for information processing and learning taking place in the brain. The brain is the paradigm for a lot of machine learning algorithms and the capabilities and the complexity of artificial intelligence is compared to human intelligence.
If you want to dig deeper into information processing in the brain and the comparison to artificial neural networks, I recommend the blog of Tim Dettmers.
Algorithms who solve problems for which humans need some kind of “intelligence” to solve them are called artificial intelligent. Examples for problems for which some kind of intelligence is necessary are playing chess, go or Jeopardy. Also recognizing objects on photos or in videos, understanding speech and natural language or translating from one human language into another language is considered intelligent. Algorithms who solve these problems are called artificial intelligent.
Two different kinds of artificial intelligence are distinguished:
Specialized AI or Artificial Narrow Intelligence (ANI): are algorithms designed for special problems or tasks (like pattern recognition). These algorithms cannot be used for other tasks. All AI algorithms today belong to the specialized AI.
Generalized AI or Artificial General Intelligence (AGI): are algorithms that can solve a high number of different intelligent tasks. They can generalize knowledge learned from one task and use it to solve other tasks. Humans possess generalized Intelligence, but for algorithms generalized AI is still in an early research state.
When algorithms capable of a generalized AI are developed and these algorithms have access to the huge amount of data on the internet, it might be possible that these algorithms could become more intelligent than humans. This kind of intelligence is called Artificial Super Intelligence (ASI). This point in time in the development of AI algorithms is called “singularity”. A lot of people are warning of the possible dangers of super-intelligent algorithms to humanity (e.g. Stephen Hawking, Bill Gates, Elon Musk). Nick Bostrom, Professor at the University of Oxford, has written a book about the possible dangers of Superintelligence (see material). Tim Urban has written an excellent blog on his site whaitbutwhy.com about the the possible development and dangers of superintelligence as described by Ray Kurzweil and Nick Bostrom (part 1 and part 2). An excellent article describing what everybody can do against the potential risks of a superintelligence can be found here. I also recommend to watch the TED Talk by Sam Harris, a neuroscientist and philosopher. He is talking about the inevitable advent of superintelligence (link).
Futurism.com descries four different levels of AI from purely reactive to self-aware: see link.
For a comprehensive introduction into the field of artificial intelligence, see the book of Stuart Russell and Peter Norvig: Artificial Intelligence, A Modern Approach (see material).
Algorithms are called machine learning algorithms, if they optimize their performance by the processing of data. These algorithms are not programmed by a humans to solve a problem, but “learn” to solve a problem with training data.
The formal but precise definition of machine learning is from Tom M. Mitchell, a researcher from CMU:
“A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E.”
Artificial Neural Networks
Artificial Neural Networks are machine learning algorithms which use some ideas found in the human brains. Artificial neurons, which are called units have a certain state or activity. Special input- or output-units can are used to feed some input data into the model or get the output of the model.
The units are connected by artificial synapses, which are directed connections between units. These connections have a certain “weight”. The weight of a connection determines how the sending unit influences the receiving unit.
Training algorithms optimize the values of the weights to optimize the performance of the network by processing a high amount of training data. The weights of the network can be considered as the “long-term memory” of the the network.
Because the units are usually ordered in layers and the activities of the units are computed layer after layer, these networks are called Multi-Layer Perceptrons (MLP) or Feedforward Networks.
Feedforward Networks with a high number of layers are called “deep neural networks” or deep learning algorithms”. The high number of layers allows the network to learn different representations of the input data in the different layers. Lower layers are learning to recognize simple features and higher layers are learning to recognize more complex and abstract features based on the simple features recognized in the lower layers.
To learn these different layers of feature recognition properly, a high number of training data is necessary. As a rule of thumb, for an acceptable performance of the network a minimum of 5.000 labeled examples are necessary per learned category. Training sets of 10 million or more lableled examples can even exceed human performance (see Ian Goodfellow, Yoshua Bengio, Aaron Courville: Deep Learning, page 20).
Actually, deep learning algorithms show superior performance compared to other machine learning or Artificial Intelligence algorithms.
There exists also critical views on Deep Learning and the Backpropagation algorithm in particular. Gary Marcus has summarized these views in a recent article (Deep Learning: A Critical Appraisal).
Data science is an interdisciplinary field about scientific and engineering methods, processes and systems to extract useful information or insights from data in various forms, either structured or unstructured. Machine learning can be viewed as a subset of data science. To solve a machine learning problem, a significant part of time and resources is usually spent with data science tasks to prepare and clean the training and test data and to find a proper representation of the data.
The Harvard Business Review called data science „The Sexiest Job of the 21st Century“ (link).