Neural Turing Machines.pdf
免费下载
Neural Turing Machines
Alex Graves gravesa@google.com
Greg Wayne gregwayne@google.com
Ivo Danihelka danihelka@google.com
Google DeepMind, London, UK
Abstract
We extend the capabilities of neural networks by coupling them to external memory re-
sources, which they can interact with by attentional processes. The combined system is
analogous to a Turing Machine or Von Neumann architecture but is differentiable end-to-
end, allowing it to be efficiently trained with gradient descent. Preliminary results demon-
strate that Neural Turing Machines can infer simple algorithms such as copying, sorting,
and associative recall from input and output examples.
1 Introduction
Computer programs make use of three fundamental mechanisms: elementary operations
(e.g., arithmetic operations), logical flow control (branching), and external memory, which
can be written to and read from in the course of computation (Von Neumann, 1945). De-
spite its wide-ranging success in modelling complicated data, modern machine learning
has largely neglected the use of logical flow control and external memory.
Recurrent neural networks (RNNs) stand out from other machine learning methods
for their ability to learn and carry out complicated transformations of data over extended
periods of time. Moreover, it is known that RNNs are Turing-Complete (Siegelmann and
Sontag, 1995), and therefore have the capacity to simulate arbitrary procedures, if properly
wired. Yet what is possible in principle is not always what is simple in practice. We
therefore enrich the capabilities of standard recurrent networks to simplify the solution of
algorithmic tasks. This enrichment is primarily via a large, addressable memory, so, by
analogy to Turing’s enrichment of finite-state machines by an infinite memory tape, we
1
arXiv:1410.5401v2 [cs.NE] 10 Dec 2014
dub our device a “Neural Turing Machine” (NTM). Unlike a Turing machine, an NTM
is a differentiable computer that can be trained by gradient descent, yielding a practical
mechanism for learning programs.
In human cognition, the process that shares the most similarity to algorithmic operation
is known as “working memory.” While the mechanisms of working memory remain some-
what obscure at the level of neurophysiology, the verbal definition is understood to mean
a capacity for short-term storage of information and its rule-based manipulation (Badde-
ley et al., 2009). In computational terms, these rules are simple programs, and the stored
information constitutes the arguments of these programs. Therefore, an NTM resembles
a working memory system, as it is designed to solve tasks that require the application of
approximate rules to “rapidly-created variables.” Rapidly-created variables (Hadley, 2009)
are data that are quickly bound to memory slots, in the same way that the number 3 and the
number 4 are put inside registers in a conventional computer and added to make 7 (Minsky,
1967). An NTM bears another close resemblance to models of working memory since the
NTM architecture uses an attentional process to read from and write to memory selectively.
In contrast to most models of working memory, our architecture can learn to use its working
memory instead of deploying a fixed set of procedures over symbolic data.
The organisation of this report begins with a brief review of germane research on work-
ing memory in psychology, linguistics, and neuroscience, along with related research in
artificial intelligence and neural networks. We then describe our basic contribution, a mem-
ory architecture and attentional controller that we believe is well-suited to the performance
of tasks that require the induction and execution of simple programs. To test this architec-
ture, we have constructed a battery of problems, and we present their precise descriptions
along with our results. We conclude by summarising the strengths of the architecture.
2 Foundational Research
2.1 Psychology and Neuroscience
The concept of working memory has been most heavily developed in psychology to explain
the performance of tasks involving the short-term manipulation of information. The broad
picture is that a “central executive” focuses attention and performs operations on data in a
memory buffer (Baddeley et al., 2009). Psychologists have extensively studied the capacity
limitations of working memory, which is often quantified by the number of “chunks” of
information that can be readily recalled (Miller, 1956).
1
These capacity limitations lead
toward an understanding of structural constraints in the human working memory system,
but in our own work we are happy to exceed them.
In neuroscience, the working memory process has been ascribed to the functioning of a
system composed of the prefrontal cortex and basal ganglia (Goldman-Rakic, 1995). Typ-
1
There remains vigorous debate about how best to characterise capacity limitations (Barrouillet et al.,
2004).
2
of 26
免费下载
【版权声明】本文为墨天轮用户原创内容，转载时必须标注文档的来源（墨天轮），文档链接，文档作者等基本信息，否则作者和墨天轮有权追究责任。如果您发现墨天轮中有涉嫌抄袭或者侵权的内容，欢迎发送邮件至：contact@modb.pro进行举报，并提供相关证据，一经查实，墨天轮将立刻删除相关内容。
下载排行榜
评论