Вы находитесь на странице: 1из 1

NEWS & VIEWS doi:10.

1038/nature19477

ART IFICIAL INTELLIGENCE mathematical, numerical and run-time issues.


Only in the past few years has machine-learn-

Deep neural reasoning ing research overcome these obstacles, through


a compendium of techniques that have become
branded as deep learning5. The authors train-
ing of a DNC is a splendid demonstration of
The human brain can solve highly abstract reasoning problems using a neural the power of deep learning.
network that is entirely physical. The underlying mechanisms are only partially Graves et al. steer clear of grand claims
understood, but an artificial network provides valuable insight. about their works implications for the neural
symbolic integration problem and, with due
caution, suggest possible mappings of DNC
HERBERT JAEGER modules, all of which are entirely non-sym- structures to those of biological brains. This
bolic and operate by exchanging streams of is wise, because the debates fought out in this

A
classic example of logical reasoning purely analog activation patterns just like arena are fierce and without winners. Instead,
is the syllogism, All men are mortal. those recorded from biological brains. There the authors establish an undeniable techni-
Socrates is a man. Therefore, Socrates are two main modules: a memory comprised cal anchor point that will help to ground the
is mortal. According to both ancient and of a large grid of memory cells, each of which debatesthey have shown that certain non-
modern views1, reasoning amounts to a rule- can have a particular numerical value that is trivial, central aspects of symbolic reasoning
based mental manipulation of symbolsin akin to a voltage; and a controller, which is an can be learnt by artificial neural systems.
this example, the words All, men, and so on. artificial neural network. The controller can With regard to practical exploits, deep-
But human brains are made of neurons that access selected locations on the memory grid, learning methods have so far excelled in tasks
operate by exchanging jittery electrical pulses, read what it finds there, combine that with that require limited or no working memory,
rather than word-like symbols. This differ- input data and write numerical values back to such as image recognition6 and sentence-wise
ence encapsulates a notorious scientific and selected memory locations. The two modules language translation7. Whether or not DNCs
philosophical enigma, sometimes referred to interact in many respects like the RAM and will bring about practical advances in big-data
as the neuralsymbolic integration problem2, central processing technologies remains to be seen. The authors
which remains unsolved. In a paper online in unit of a digital com- demonstrations are not particularly complex
Nature, Graves etal.3 use the machine-learn-
The authors puter. as demands on rational reasoning go, and
ing methods of deep learning to impart some neural system Graves and col- could be solved by the algorithms of symbolic
crucial symbolic-reasoning mechanisms to an cannot and leagues demonstrate artificial intelligence of the 1970s. However,
artificial neural system. Their system can solve need not be the capabilities of those programs were handcrafted by humans
complex tasks by learning symbolic-reasoning programmed their system by put- and do not learn from examples.
rules from examples, an achievement that has instead, it is ting it through sev- For the time being, the DNC by itself
potential implications for the neuralsymbolic trained. eral tasks that require cannot compete with state-of-the-art methods
integration problem. rational reasoning, in digital computing when it comes to logical
A key requirement for reasoning is a work- such as planning a multi-stage journey using data mining8. But a flexible, extensible DNC-
ing memory. In digital computers, this role is public transport. Such tasks are fairly easy to style working memory might allow deep learn-
served by the random-access memory (RAM). solve using the symbolic computer programs ing to expand into big-data applications that
When a computer reasons when it executes of artificial intelligence, but have so far been have a rational reasoning component, such
a program information is bundled together rather out of reach of artificial neural networks. as generating video commentaries or seman-
in working memory in ever-changing com- A digital computer solves a given task by tic text analysis. A precursor to the DNC, the
binations. Comparing human reasoning to executing a program that has been written neural Turing machine9, certainly sent thrills
the running of computer programs is not a for that purpose. By contrast, the authors through the deep-learning community.
far-fetched metaphor. In fact, a venerable his- neural system cannot and need not be pro-
torical alley leads from Aristotles definition of grammed instead, it is trained. During Herbert Jaeger is at Jacobs University
syllogisms to the modern model of a program- training, the system is presented with a large Bremen, 28759 Bremen, Germany.
mable computer (the Turing machine). Alan number of solved examples of the task at e-mail: h.jaeger@jacobs-university.de
Turing himself used mind language in his hand. With each new presentation, the system
1. Newell, A. Cogn. Sci. 4, 135183 (1980).
groundbreaking work4: The behaviour of the slightly adapts its internal neural wiring so 2. Hammer, B. & Hitzler, P. (eds) Perspectives of
computer at any moment is determined by the that its response moves gradually closer to the NeuralSymbolic Integration http://doi.org/fsrb8m
symbols which he is observing and his state of given tasks solution. (Springer, 2007).
3. Graves, A. et al. Nature http://dx.doi.org/10.1038/
mind at that moment. The analog, smoothly adaptable nature of nature20101 (2016).
Although there are clear parallels between the authors neural system is the key to its abil- 4. Turing, A. M. J. Math. 58, 345363 (1936).
human reasoning and the running of com- ity to be trained. Mathematically speaking, the 5. LeCun, Y., Bengio, Y. & Hinton, G. Nature 521,
436444 (2015).
puter programs, we lack an understanding of system is a differentiable function, which has 6. Szegedy, C. et al. Proc. IEEE Conf. Computer Vision
how either of them could be implemented in led to the authors calling it a differentiable neu- Pattern Recognition http://dx.doi.org/10.1109/
biological or artificial neural networks. Graves ral computer (DNC). A digital computer is not CVPR.2015.7298594 (2015).
7. Bahdanau, D., Cho, K. & Bengio, Y. Int. Conf.
and colleagues take a substantial step forward differentiable and could not be trained in any Learning Representations Preprint at http://arxiv.
in this quest by presenting a neuro-computa- similar fashion. org/abs/1409.0473 (2014).
tional system that shows striking similarities A DNC is a mathematical object that boasts 8. De Raedt, L. & Kimmig, A. Machine Learn. 100, 547
(2015).
to a digital computer. tens of thousands of adjustable parameters. 9. Graves, A., Wayne, G. & Danihelka, I. Preprint at
The authors system consists of several Training such a monster raises a plethora of http://arxiv.org/abs/1410.5401 (2014).

| NAT U R E | 1

2
0
1
6
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.

Вам также может понравиться