60 Scientific American, May 2019
Input: Thousands
of cat photographs
Each layer of the network
learns to identify progressively
more complex features
Training
Images
broken
into pixels
Result: Ability to recognize a cat
Output: Image label
Cat
Input: Sets of
different defined
groupings
Training
Pretraining
Input: A few cat photographs
Result: Ability to recognize a cat faster
Cat
Result: Ability to generate convincing cat image
Discriminator
is randomly given
either a real or
a fake cat image
Training
Discriminator
judges whether the
image is real. If not,
in what ways is it not
real? Feedback is fed
to the generator.
Real
Fa ke
Fake (generated)
cat image
Noise
Discriminator
Generator
Input: Random
noise and a class
Cat
Result: Ability to isolate and reconstruct elements
Training
Input: Primitive elements with multiple variables
Bottleneck is gradually loosened
Many AI researchers got into the field because they want to
understand, reproduce and ultimately surpass human intelli-
gence. Yet even those with more practical interests think that ma-
chine systems should be more like us. A social media company
training its image recognizers, for example, will have no trouble
finding cat or celebrity pictures. But other categories of data are
harder to come by, and machines could solve a wider range of
problems if they were quicker-witted. Data are especially limited
if they involve the physical world. If a robot has to learn to manip-
ulate blocks on a table, it can’t realistically be shown every single
arrangement it might encounter. Like a human, it needs to ac-
quire general skills rather than memorizing by rote.
In getting by with less input, machines also need to be more
forthcoming with output. Just the answer isn’t enough; people
also want to know the reasoning, especially when algorithms pass
judgment on bank loans or jail sentences. You can interrogate hu-
man bureaucrats about their biases and conflicts of interest; good
luck doing that with today’s AI systems. In 2018 the European
Union gave its citizens a limited right to an explanation for any
judgment made by automated processing. In the U.S., the Defense
Advanced Research Projects Agency funds an “Explainable AI” re-
search program because military commanders would rather not
send soldiers into battle without knowing why.
A huge research community tackles these problems. Ideas
abound, and people debate whether a more humanlike intelli-
gence will require radical retooling. Yet it’s remarkable how far re-
searchers have gone with fairly incremental improvements. Self-
improvement, imagination, common sense: these seemingly quint-
essential human qualities are being incorporated into machines,
at least in a limited way. The key is clever coaching. Guided by hu-
man trainers, the machines take the biggest steps themselves.
DEEP NETWORKS
More than Most fields of science and engineering, AI is highly cy-
clical. It goes through waves of infatuation and neglect, and meth -
ods come in and out of fashion. Neural networks are the as cen-
dant technology. Such a network is a web of basic com put ing
units: “neurons.” Each can be as simple as a switch that tog gles on
or off depending on the state of the neurons it is con nect ed to.
The neurons typically are arrayed in layers. An ini tial layer ac -
cepts the input (such as image pixels), a final layer pro duces the
output (such as a high-level description of image con tent), and
the intermediate, or “hidden,” layers create arith me tic com bina-
tions of the input. Some networks, especially those used for prob-
lems that unfold over time, such as language recog nition, have
loops that reconnect the output or the hidden layers to the input.
Illustration by Brown Bird Design
Network
Effects
For all their immense power,
neural networks still have
frustrating limitations. For
classifying images, the net
work takes in the image pix
els, processes them through
multiple stages, and outputs
the probabilities of the vari
ous labels the image might be
given. Finetuning the inter
connections typi cally takes
thousands of sample images.
How exactly the network
performs the classifications is
lost in the tangle of wiring.
Several new techniques fix
these shortcomings
Meta-Learning
To reduce the amount of training data, researchers can
prime the network by giving it practice exercises of the
same general type. The network does not retain any of the
information but gradually gets better at solving whatever
new tasks it is given. It learns how to learn.
George Musser is a contributing editor to Scientific American
and author of Spooky Action at a Distance (Farrar, Straus and Giroux,
2015) and The Complete Idiot’s Guide to String Theory (Alpha, 2008).
© 2019 Scientific American