Childish augmented intelligence
Scientists are researching on how young children learn naturally.
They hope this will help develop Augmented Intelligence.
Children acquire a great deal of knowledge, without the help of parents or teachers.
Even the most powerful computers, cannot learn as well as a five year old.
One way to solve problems with computers, is the bottom-up approach.
In a basic way, this approach looks at the signals received by the brain,
looks for patterns, and then tries to make sense of these signals.
For example, photons of light are converted into electrical signals, in the retina,
and sent to the visual cortex, in the brain, for processing.
"Connectionist" or "Neural network", systems take inspiration, from the way that the brain,
converts light patterns in the retina, into representations of the world around you.
A neural network does something similar.
It uses interconnected processing elements, to transform pixels,
into increasingly abstract representations, at a higher level.
Patterns of light, shade and colour are abstracted to represent, a nose or mouth.
These are further abstracted to recognise a face.
Neural network concepts, termed as "deep learning", are used by some technology corporations,
including Google and Facebook.
This concept will get better with faster computers, and larger data sets.
Top-down approaches, leverage what a system already knows,
and help it to learn something new.
Such approaches are used in probabilistic, or Bayesian modelling.
Top-down systems start with formulated abstract hypothesis.
The systems make predictions on what the data, should look like if the hypothesis is correct.
The revise the hypothesis, depending on the outcome of the predictions.
Let us take an example, of a bottom-up approach.
Mail needs to be filtered, for spam.
Spam mails tend to have certain distinguishing characteristics.
For example, a long list of addresses, prizes, misspellings, etc..
A pattern of these characteristics can be used to identify spam mail.
Let us take another example.
Identification of the letter 'A'.
In a bottom-up, or deep learning approach, the system is trained,
by different types of 'A', say of different fonts and there corresponding pixel pattern.
When a system receives a new input, it assesses, whether the pixels match the data patterns, that it knows.
The top-down approach for the same objective works in a different way.
The system builds a model of the letter 'A', from its internal library of parts.
It uses this model to identify variants of 'A', like 'A' in italics, or 'A' in a different font.
A neural network, can inspect images, in the internet, and label them as 'Cat', 'Car','Tree' etc..
It looks for features that cluster together, which gives it an identity.
The ability to use augmented intelligence, for image recognition, from large data sets,
like millions of Instagram images, is now possible.
Speech recognition also uses similar principles.
The Bayesian model is a prime example of the top-down approach.
They combine generative models with probability theory,
using a technique called Bayesian inference.
A probabilistic generative model, can tell how likely it is,
that we can see a specific pattern of data, for specific hypothesis is true.
For example, a Bayesian model combines the knowledge we already have about potential hypothesis,
with the data, and calculates quite precisely, how likely, a hypothesis is true.
For example, it can tell us, how likely a mail is legitimate or spam.
In one example, scientists used the top-down method, to teach a computer to identify a character,
based on how it is drawn.
For example, a stroke that goes up or down, or right or left.
When the program was given a character, it could infer the sequence of strokes needed to draw it.
This model guessed whether a particular stroke sequence, was likely to produce a desire character.
In this case, the top-down method worked particularly well, compared to deep learning.
The bottom-up and top-down approach have complimentary strengths and weaknesses.
The bottom-up does not understand the concept, but it requires a large amount of data to learn.
The top down approach or the bayesian system, can learn from a few examples, and generalise it widely.
However, it requires a lot of work, to articulate the right set of hypothesis.
The two approaches work only on narrow and well defined problems.
They cannot be termed as general intelligence.
They cannot be compared with a child's intelligence,
which spontaneously learns many different concepts.
Though there is a lot of speculation, about intelligent computers taking over the world,
computers software is no where near the mysterious, complex and sophisticated learning,
that a child is capable of.
Scientists are trying to understand, what method children use to learn?
May be a combination of bottom-up and top-down methods.
May be a entirely a new method, which we currently have no idea about.
Some scientists are trying to understand, how children learn.
This they hope will help them design more intelligent software.