没有合适的资源?快使用搜索试试~ 我知道了~
给初学者看的神经网络的数学基础 A Beginner's Guide to the Mathematics of Neural ...
需积分: 7 13 下载量 27 浏览量
2018-09-25
14:19:00
上传
评论
收藏 935KB PDF 举报
温馨提示
试读
61页
In this paper I try to describe both the role of mathematics in shaping our understanding of how neural networks operate, and the curious new mathematical concepts generated by our attempts to capture neural networks in equations. My target reader being the non-expert, I will present a biased selection of relatively simple examples of neural network tasks, models and calculations, rather than try to give a full encyclopedic review-like account of the many mathematical developments in this field.
资源推荐
资源详情
资源评论
A Beginner's Guide to the
Mathematics of Neural Networks
A.C.C. Co olen
Department of Mathematics, King's College London
Abstract
In this pap er I try to describe both the role of mathematics in shap-
ing our understanding of how neural networks operate, and the curious
new mathematical concepts generated by our attempts to capture neu-
ral networks in equations. My target reader b eing the non-exp ert, I will
present a biased selection of relatively simple examples of neural network
tasks, models and calculations, rather than try to give a full encyclop edic
review-like account of the many mathematical developments in this eld.
Contents
1 Intro duction: Neural Information Pro cessing 2
2 From Biology to Mathematical Mo dels 6
2.1 From Biological Neurons to Mo del Neurons . . . . . . . . . . . 6
2.2 Universality of Mo del Neurons . . . . . . . . . . . . . . . . . . 9
2.3 Directions and Strategies . . . . . . . . . . . . . . . . . . . . . 12
3 Neural Networks as Asso ciative Memories 14
3.1 Recipes for Storing Patterns and Pattern Sequences . . . . . . 15
3.2 Symmetric Networks: the Energy Picture . . . . . . . . . . . . 19
3.3 Solving Mo dels of Noisy Attractor Networks . . . . . . . . . . . 20
4 Creating Maps of the Outside World 26
4.1 Map Formation Through Competitive Learning . . . . . . . . . 26
4.2 Solving Mo dels of Map Formation . . . . . . . . . . . . . . . . 29
5 Learning a Rule From an Exp ert 35
5.1 Perceptrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.2 Multi-layer Networks . . . . . . . . . . . . . . . . . . . . . . . . 39
5.3 Calculating what is Achievable . . . . . . . . . . . . . . . . . . 43
5.4 Solving the Dynamics of Learning for Perceptrons . . . . . . . 47
6 Puzzling Mathematics 52
6.1 Complexity due to Frustration, Disorder and Plasticity. . . . . 52
6.2 The World of Replica Theory . . . . . . . . . . . . . . . . . . . 55
7 Further Reading 59
1
1 Intro duction: Neural Information Pro cessing
Our brains perform sophisticated information pro cessing tasks, using hardware
and op eration rules which are quite dierent from the ones on which conven-
tional computers are based. The pro cessors in the brain, the neurons (see gure
1), are rather noisy elements
1
which op erate in parallel. They are organised in
dense networks, the structure of which can vary from very regular to almost
amorphous (see gure 2), and they communicate signals through a huge num-
ber of inter-neuron connections (the so-called synapses). These connections
represent the `program' of a network. By continuously up dating the strengths
of the connections, a network as a whole can modify and optimise its `program',
`learn' from exp erience and adapt to changing circumstances.
Figure 1: Left: a Purkinje neuron in the human cerebellum. Right: apyramidal
neuron of the rabbit cortex. The black blobs are the neurons, the trees of wires
fanning out constitute the input channels (or dendrites) through which signals
are received which are sentoby other ring neurons. The lines at the b ottom,
bifurcating only mo destly, are the output channels (or axons).
From an engineering p oint of view neurons are in fact rather p o or processors,
they are slow and unreliable (see the table below). In the brain this is overcome
by ensuring that always a very large numb er of neurons are involved in any task,
and byhaving them operate in parallel, with many connections. This is in sharp
contrast to conventional computers, where op erations are as a rule p erformed
sequentially, so that failure of any part of the chain of operations is usually
fatal. Furthermore, conventional computers execute a detailed sp ecication of
orders, requiring the programmer to know exactly which data can b e exp ected
and how to respond. Subsequentchanges in the actual situation, not foreseen
by the programmer, lead to trouble. Neural networks, on the other hand,
can adapt to changing circumstances. Finally, in our brain large numbers of
neurons end their careers eachday unnoticed. Compare this to what happens
if we randomly cut a few wires in our workstation.
1
By this we mean that their output signals are to some degree sub ject to random variation;
they exhibit so-called spontaneous activity which appears not to b e related to the information
processing task they are involved in.
2
Figure 2: Left: a section of the human cerebellum. Right: a section of the
human cortex. Note that the staining metho d used to produce such pictures
colours only a reasonably mo dest fraction of the neurons present, so in reality
these networks are far more dense.
Roughly sp eaking, conventional computers can b e seen as the appropriate
tools for p erforming well-dened and rule-based information pro cessing tasks,
in stable and safe environments, where all p ossible situations, as well as howto
respond in every situation, are known beforehand. Typical tasks tting these
criteria are e.g brute-force chess playing, word processing, keeping accounts
and rule-based (civil servant) decision making. Neural information processing
systems, on the other hand, are sup erior to conventional computers in dealing
with real-world tasks, such as e.g. communication (vision, sp eech recognition),
movement co ordination (rob otics) and experience-based decision making (clas-
sication, prediction, system control), where data are often messy, uncertain or
even inconsistent, where the number of p ossible situations is innite and where
perfect solutions are for all practical purposes non-existent.
3
One can distinguish three types of motivation for studying neural networks.
Biologists, physiologists, psychologists and to some degree also philosophers aim
at understanding information pro cessing in real biological nervous tissue. They
study mo dels, mathematically and through computer simulations, which are
preferably close to what is b eing observed experimentally, and try to understand
the global prop erties and functioning of brain regions.
conventional computers biological neural networks
processors neurons
operation speed
10
8
Hz operation speed
10
2
Hz
sig nal=noise
1
sig nal=noise
1
sig nal v el ocity
10
8
m=sec sig nal v el ocity
1
m=sec
connections
10
connections
10
4
sequential op eration parallel op eration
program & data connections, neuron thresholds
external programming self-programming & adaptation
hardware failure: fatal robust against hardware failure
no unforseen data messy, unforseen data
Engineers and computer scientists would like to understand the princi-
ples behind neural information pro cessing in order to use these for designing
adaptive software and articial information pro cessing systems which can also
`learn'. They use highly simplied neuron models, which are again arranged
in networks. As their biological counterparts, these articial systems are not
programmed, their inter-neuron connections are not prescribed, but they are
`trained'. They gradually `learn' to perform tasks by b eing presented with ex-
amples of what they are supp osed to do. The key question then is to understand
the relationships b etween the network performance for a given type of task, the
choice of `learning rule' (the recip e for the mo dication of the connections) and
the network architecture. Secondly, engineers and computer scientists exploit
the emerging insightinto the way real (biological) neural networks manage to
process information eciently in parallel, by building articial neural networks
in hardware, which also op erate in parallel. These systems, in principle, have
the p otential of b eing incredibly fast information processing machines.
Finally, it will be clear that, due to their complex structure, the large num-
bers of elements involved, and their dynamic nature, neural network models
exhibit a highly non-trivial and rich behaviour. This is why also theoretical
physicists and mathematicians have b ecome involved, challenged as they are
by the many fundamental new mathematical problems p osed by neural net-
work models. Studying neural networks as a mathematician is rewarding in
twoways. The rst reward is to nd nice applications for one's to ols in biology
and engineering. It is fairly easy to come up with ideas ab out how certain in-
formation pro cessing tasks could be p erformed by (either natural or synthetic)
neural networks; byworking out the mathematics, however, one can actually
4
quantify the potential and restrictions of such ideas. Mathematical analysis fur-
ther allows for a systematic design of new networks, and the discovery of new
mechanisms. The second reward is to discover that one's to ols, when applied
to neural network mo dels, create quite novel and funny mathematical puzzles.
The reason for this is the `messy' nature of these systems. Neurons are not at
all well-behaved: they are microscopic elements which do not live on a regular
lattice, they are noisy, they change their mutual interactions all the time, etc.
Since this pap er aims at no more than sketching a biased impression of a
research eld, I will not give references to research pap ers along the way, but
mention textbo oks and review pap ers in the nal section, for those interested.
5
剩余60页未读,继续阅读
资源评论
wumoomowu
- 粉丝: 0
- 资源: 4
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功