Communications of the ACM

Figure 1. An abbreviated taxonomy of the evolution of neural networks shows a progression from simple one-layer Perceptron to multilayered Perceptron (MLP), convolutional and recurrent networks with memory and adaptable reinforcement learning algorithms.

possible terrorists. Some military strat-
egists talk about using them as auto-
matic fire-control systems. How can we
trust the networks?

This is a hot question. We know that a network is quite reliable when its inputs come from its training set. But these critical systems will have inputs corresponding to new, often unanticipated situations. There are numerous examples where a network gives poor responses for untrained inputs. This is called the “fragility” problem. How can we know that the network’s response will not cause a disaster or catastrophe?

The answer is we cannot. The “ programs” learned by neural networks are in effect enormous, incomprehensible matrices of weights connecting millions of neurons. There is no more hope of figuring out what these weights mean or how they will respond to a new input than in looking for a person’s motivations by scanning the brain’s connections. We have no means to “explain” why a medical network reached a particular conclusion. At this point, we can only experiment with the network to see how it performs for untrained inputs, building our trust slowly and deliberately.

Q: Computers were in the news 20 years ago for beating the grandmaster chess players, and today for beating the world’s Go master champion. Do these advances signal a time when machines

Back in favor because of the MLP breakthrough, neural networks advanced rapidly. We have gone well beyond recognizing numbers and handwriting, to networks that recognize and label faces in photographs. New methods have since been added that allow recognition in video moving images; see the figure for some of the keywords.

Q: What propelled the advances?

Two things. The abundance of data, especially from online social networks like Twitter and Facebook, large-scale sensor networks such as smartphones giving positional data for traffic maps, or searches for correlations between previously separate large databases. The questions that could be answered if we could process that data by recognizing and recommending were a very strong motivating force.

The other big factor is the proliferation of low-cost massively parallel hardware such as the Nvidia GPU (Graphics Processing Unit) used in graphics cards. GPUs rapidly process large matrices representing the positions of objects. They are super-fast lin-ear-algebra machines. Training a network involves computing connection matrices and using a network involves evaluations of matrix multiplications. GPUs do these things really well.

Q: These networks are now used for
critical functions such as medical di-
agnosis or crowd surveillance to detect
Advertise with ACM!

Reach the innovators
and thought leaders
working at the
cutting edge
of computing
and information
technology through
ACM’s magazines,
websites
and newsletters.