How and what to learn — The modes of machine learning

Sihan Feng, Yong Zhang, Fuming Wang and Hong Zhao*

Department of Physics, Xiamen University, Xiamen 361005, Fujian, China
*[email protected]

Despite their great success, there is still no comprehensive understanding of

machine learning. Here we proposal a new approach, namely the weight pathway
analysis (WPA), to study the mechanism of recognition and classification of
multilayer neural networks. Instead of neurons or weight edges, we consider
weight pathways linking neurons longitudinally from input neurons to output
neurons as the basic units of a neural network. We decompose a neural network
into a series of subnetworks of weight pathways that have either a hidden-layer
neuron or an output-layer neuron as node, and then establish characteristic maps
for these subnetworks. The characteristic map characterizes the enhancement or
suppression of the influence of each component of an input vector over the hidden-
layer neuron or the output-layer neuron, and at the same time reveals the internal
coherence structure of the neural network. The parameters of a characteristic map
can be visualized, providing a “longitudinal” perspective of the network and
making the neural network explainable. Using WPA, we discover that a neural
network stores and utilizes information in a "holographic" way, that is, the
network encodes all training samples in a coherent structure. An input vector
interacts with this "holographic" structure to enhance or suppress each
subnetwork which working together to produce the correct activities in the output
neurons to recognize the input sample. Furthermore, with WPA, we reveal
fundamental learning modes of a neural network: the linear learning mode and
the nonlinear learning mode. The former extracts linearly separable features while
the latter extracts linearly inseparable features. It is found that hidden-layer
neurons self-organize into different classes in the later stages of the learning
process. It is further discovered that the key strategy to improve the performance
of a neural network is to control the ratio of the two learning modes to match that
of the linear and the nonlinear features while avoiding overlearning, and that
increasing the width or the depth of a neural network helps this ratio controlling
process. This provides theoretical ground for the practice of optimizing a neural
network via increasing its width or its depth. WPA gives us a newer and deeper
perspective of neural networks, the knowledge gained with WPA enables us to
understand the fundamental questions such as what to learn, how to learn, and
how can learn well.

INTRODUCTION

In the last decades, learning algorithms have made remarkable progress on

numerous machine learning tasks and dramatically improved the state-of-the-art in
many practical areas [1]. In addition to practical applications, theoretical studies
striving to understand the mechanism of learning and make the “black-box” of learning
models transparent are receiving increased attention [2-4]. Researchers have carried out
studies from various angles, including mutual information [5,6], the hidden manifold
model [7,8], statistical mechanics [9,10], mean-field theory [11], etc.. These
investigations deepened our understanding of the learning mechanism, but we are only
beginning to have some understanding of it [2,3]. In particular, although various
methods and strategies to improve the performance of learning models have been tried,
such as increasing the width and the depth of the network, using the smallest possible
learning rate, etc., it is not fully clear why would these methods work or to what extent
will they work. To advance in this area, researches on interpretability and explainability
of the training process, learned representations, and decisions with human-interpretable
explanations have developed into a subfield of machine learning [12,13]. Recently
there has been a return to simple models, particularly linear models such as Deep Matrix
Factorization [14,15] ， for gaining the better interpretability. These studies have

achieved great success. However, linear models suffer from low predictivity, and more
essentially, the success of machine learning is mainly due to the application of nonlinear
models. Therefore, to develop explanation methods applicable for both linearly
separable and linearly inseparable features and to investigate the difference between the
learning of these two kinds of features are important.

Fig.1 Network of weight pathways, (a) a subnetwork of an output neuron,

(b)-(d) subnetworks of a hidden-layer neuron. Networks connections are
represented by dark solid lines.

We present a framework based on the concept of weight pathways to study neural

networks; we call it weight pathway analysis (WPA). Different from previous layer by
layer analyzing approaches [16,17], we establish a “longitudinal” radiographic view of
the network. In more details, considering a neural network shown in Fig. 1, we see that
there are pathways between input neurons and output neurons that are connected by
weighted edges. We call a path connecting an input neuron and an output neuron via
these weighted edges a weight pathway and the product of the weights along the path
the weight product of the weight pathway. In the case of a monotonic neuron transport
function (which is the common case for machine learning; the idea can be extended to
more complex neuron transfer functions), the product of a weight pathway has a
monotonic relation to the neural network outputs. Weight pathway view point give us
ability to investigate direct relations between inputs and outputs of a network. On the
contrary, a single weight or a single layer of weights does not provide sufficient
information to infer its contribution to the outputs without further calculation. Therefore,
weight pathways instead of single weights or single layers of weights should be treated
as the basic units for understanding a neural network.
With hidden-layer or output-layer neurons as nodes, we decompose a neural
network into subnetworks of weight pathways, as examples shown by Fig. 1. Then, we
integral all of the weight pathways connecting an input neuron and the output neuron
(i.e., sum up the weight products of these weight pathways) and define the value as the
penetration coefficient from the input neuron to the output neuron. The penetration
coefficients of all of the input neurons to the output neuron of a subnetwork are used to
construct a characteristic map of the subnetwork, which characterizes the enhancement
or suppression effect of the subnetwork on each input vector component’s ability of
excite the output neuron of the subnetwork.
We visualize the penetration coefficients of a subnetworks and call it the
radiograph of the subnetwork. In the case that samples are two-dimensional images, the
penetration coefficients and the image pixels have a one-to-one correspondence, by
which one can infer the enhancement or suppression effect of the subnetwork to each
pixel. The radiograph gives a “longitudinal” penetrating view of the neural network,
making the "black-box" less opaque and more interpretable, as any adjustment or
operation to the network can be visualized via the radiograph. The visualization here
reflects the structure of the network itself, in other words, the structure of the system
parameters, which is different from the famous "mask" method [18], which needs
backtracking calculations with the input of specific samples. Such penetrating
radiographs enable us to gain important insights on the learning process of neural
networks. For example, we discover that subnetworks are "holographic", that is, every
training sample is encoded in every subnetwork instead of some training samples
encode in some subnetwork or neurons. This discovery provides important information
on how the information is stored in a neural network and how new objects are identified.
With the help of penetrating radiographs, we show that hidden-layer neurons self-
organize into different classes in order to reduce the cost function. More essentially, we
reveal that the network uses a learning mode, we call linear learning mode, for
extracting linearly separable features, while uses another learning mode, we call
nonlinear learning mode, to extract linearly inseparable features. It is found that linearly
separable features can be extracted by single neurons, while linearly inseparable
features can only be extracted via cooperation of multiple neurons using the
nonlinearity characteristics of their transfer function. As a result, neural networks will
always try to extract linear features with linear learning mode first; this explains an
important previous finding, that is, the earlier learning stage of a deep neural networks
is equivalent to a linear classifier [19,20]. Our study also reveals that the nonlinear
learning mode is initiated when linear features are insufficient to further minimize the
cost function. Therefore, how to maximize the extraction of linearly separable features
and linearly inseparable features is the key to optimize the performance of a neural
network. We will show that increasing the width or the depth of a neural network helps
this optimization process. This finding provides theoretical ground for the practice of
optimizing a neural network via increasing its width or its depth.
This paper is arranged as the following. Section 2 introduces our model and
algorithm. Section 3 demonstrates the basic principle of WPA using toy samples and a
three-layer neural network. The linear and nonlinear learning modes, as well as the self-
organized classification of hidden-layer neurons, are shown in this section. Section 4
applies WPA to the learning of the MINIST set of the handwritten digits, and explains
why increasing the width and depth of a neural network can help maximizing the
extraction of linearly separable and linearly inseparable features. Section 5 introduce
another approach for monitoring activities of training process, i.e., the degree of
attention of learning, as an augmentation to the WPA. The final section gives
conclusions and discussions.

1 MODEL and METHOD

In this paper we consider a multi-layer neural network given by following

equations,
(𝑙) (𝑙)
𝑥𝑖𝑙 = 𝑓(𝛽h𝑖𝑙 )

(𝑙) 𝑙−1𝑁 (𝑙) (𝑙−1)

ℎ𝑖𝑙 = ∑𝑖𝑙−1 =1 𝑤 𝑖𝑙 𝑖𝑙−1 𝑥𝑖𝑙−1 （1）

(𝑙) (𝑙)
where 𝑥𝑖𝑙 and ℎ𝑖𝑙 are the output and the local field of the 𝑖𝑙 th neuron in the lth layer

respectively, 𝑤 (𝑙) 𝑖𝑙𝑖𝑙−1 represents the weight that connects the 𝑖𝑙−1 th neuron in the (l-
1)th layer to the 𝑖𝑙 th neuron in the lth layer, and 𝑓(𝑥) is the neuron transfer function,
𝑁𝑙−1 is the number of the neurons in the (l-1)th layer. The weights in each layer are
bounded in an interval of (-1,1) and are randomly initialized. Note that we do not
involve the neuron bias into the model for sake of the simplicity. Extending the analysis
to cases involving biases is straightforward.
Assuming that the training set is consist of P samples {(𝒙(𝜇), 𝒚(𝜇)), 𝜇 = 1, ⋯ , 𝑃}，

(0)
where 𝒙(𝜇) is the input vector of the 𝜇 th sample with components 𝑥𝑖0 (𝜇) (𝑖0 =

1, … , 𝑁0 )，𝒚(𝜇) defines its expected state in the output layer. Here we set 𝑦𝑖𝐿 (𝜇) = 1

if this sample belongs to the 𝑖𝐿 th class (we call this neuron the label neuron of this
class), and 𝑦𝑖𝐿 (𝜇) = −1 otherwise. For the sake of simplicity, we always apply the
(𝐿) (𝐿)
linear neuron transfer function to the output layer neurons, i.e., set 𝑥𝑖𝐿 = ℎ𝑖𝐿 , while
applying 𝑓(ℎ) = ℎ and 𝑓(ℎ) = tanh⁡(ℎ) to the hidden-layer neurons in linear
neural networks (LNNs) and nonlinear neural networks (NNNs) respectively. We then
define the cost function as
1 2
(𝐿)
𝑆(𝑑) = ∑𝑃𝜇=1 ∑𝑁
𝑖𝐿 =1 (𝑥𝑖𝐿 (𝜇)𝑦𝑖𝐿 (𝜇) − 𝑑) ,
𝐿
(2)
𝑃𝑁𝐿
where 𝑑 is a parameter that controls the gap between the output of the label neuron
and that of the other neurons in the output layer. This cost function is equivalent to the
Margin commonly used by support vector machines. Minimizing the cost function to
(𝐿)
zero gives 𝑥𝑖𝐿 (𝜇)𝑦𝑖𝐿 (𝜇) = 𝑑 for all of the samples (we call this relation the goal of

(𝐿)
training), but as long as 𝑥𝑖𝐿 (𝜇) is the largest on the label neuron the sample is
correctly classified (we call this condition the goal of classification).
The WPA approach is applicable to all learning algorithms in principle. However,
in order to clearly demonstrate how to control the transition from linear learning mode
to nonlinear learning mode, we use a simple gradient-free algorithm, namely the Monte
Carlo (MC) algorithm [21,22]. The algorithm is quite simple：select a weight randomly

with equal probability, then change it randomly to a new value; accept the change if it
reduces the cost-function and discard it otherwise. The operation is repeated until the
minimization of the cost function is achieved. Because each update is judged by all
samples, the cost-function will be reduced monotonously. Besides, since the changed
neurons induced by the mutation involves only those connected by weight paths
involving the mutated weight, the computation time is acceptable for neural networks
with few hidden layers [21].
Since there is no need of differentiation and back-propagation in this algorithm,
we can limit the range of weights, for example, by setting |𝑤 (𝑙) 𝑖𝑙𝑖𝑙−1 | ≤ 1. Then, with

an appropriate transfer function coefficient β ， we can control the input-output

sensitivity of the network to avoid the overlearning. Due to the restriction of the norm
of the weights, linearly separable feature of a sample cannot be amplified infinitely, and
thus the cross-over transition from the linear to nonlinear learning modes can be clearly
shown by increasing the parameter d. Furthermore, with the limitation on the input-
output sensitivity of the network, the width of the network can be greatly increased
without overlearning. On the contrary, traditional algorithms, such as the back-
propagation (BP) algorithm, does not have these properties. Therefore, although its
training speed is not as fast as other traditional algorithms, the MC algorithm is more
suitable for studying the mechanism of the machine learning due to its flexibility.

2 RESULTS

2.1 Weight pathways and visualization of penetration coefficients

(1) (2) (3) (𝐿)

We call weight edges w𝑖1 𝑖0 → w𝑖2 𝑖1 → w𝑖3 𝑖2 → ⋯ → w𝑖𝐿𝑖𝐿−1 that connects the
𝑖0 th neuron in the input layer to the 𝑖𝐿 th neuron in the output layer a weight pathway.
(1) (2) (3) (𝐿)
We use the product of the weights w𝑖1 𝑖0 w𝑖2 𝑖1 w𝑖3 𝑖2 ⋯ w𝑖𝐿𝑖𝐿−1 to characterize a weight
pathway and call it the weight product. We call a weight pathway a positive pathway
(p-pathway) if its weight product is positive, otherwise a negative pathway (n-pathway).
We divide a neural network into subnetworks. Each subnetwork connects all of the input
neurons to an output neuron, as illustrated in Fig. 1.
We then define
(𝑖 ,𝑙) (1) (2) (3) (𝑙) (𝑙+1) (𝐿)
𝑐𝑖0𝑙𝑖𝐿 = [∑𝑖1 ,𝑖2 ,⋯𝑖𝑙−1 w𝑖1 𝑖0 w𝑖2 𝑖1 w𝑖3 𝑖2 ⋯ w𝑖𝑙 𝑖𝑙−1 ]w𝑖𝑙+1 𝑖𝑙 ⋯ w𝑖𝐿𝑖𝐿−1
as the penetration coefficient of a subnetwork from the 𝑖0 th input neuron to the 𝑖𝐿 th
output neuron, where the summation is over all of the weight pathways that come out
from the 𝑖0 th input neuron, go through the 𝑖𝑙 th neuron in the lth layer, and end with
the 𝑖𝐿 th output neuron. Note that the summation covers all hidden-layer neurons
between the 𝑖0 th input neuron and the 𝑖𝑙 th neuron in the lth layer. Depending on the
selection of l, we can have summation over weight pathways that pass through all the
hidden-layer neurons (l=L) as in Fig. 1(a), over a single weight pathway from an input
to an output (l=1) as in Fig. 1(b) and 1(c), or over weight pathways through a portion
of the hidden neurons (1<l<L) as in Fig. 1(d).
The most important feature of a weight pathway is that it connects the input and
the output of a neural network. In the case of monotonic neuron transport functions, the
weight products of the weight pathways monotonously determine the direction of the
evolution of the local field of an output neuron that shared by the weight pathways, and
thus determines the evolution of the cost function. As a result, when we input the μth
(0) (𝐿)
sample, the contribution of its 𝑖0 th component 𝑥𝑖0 (𝜇) to the local field ℎ𝑖𝐿 (𝜇) of

(𝑖 ,𝑙) (0)
the 𝑖𝐿 th neuron in the output layer is proportional to 𝑐𝑖0𝑙𝑖𝐿 𝑥𝑖0 (𝜇), and the map,

(𝑖 ,𝑙) 𝑁 (𝑖 ,𝑙⁡ )
𝐻𝑖𝐿𝑙 (𝜇) = ∑𝑖00=1 𝑐𝑖0𝑙𝑖𝐿 𝑥𝑖0 (𝜇), (4)

(𝐿)
characterizes the contribution of the whole subnetwork to the local field ℎ𝑖𝐿 (𝜇). Note
that in the particular case of 𝑙 = 𝐿 (as in the case of Fig. 1(a)), for a LNN,
(𝑖 ,𝐿) (𝐿) (𝑖 ,𝐿) (𝐿)
𝐻𝑖𝐿𝐿 (𝜇) = ℎ𝑖𝐿 (𝜇) , and for a NNN 𝐻𝑖𝐿𝐿 (𝜇) ∝ ℎ𝑖𝐿 (𝜇) . Therefore, penetration
coefficients characterize the influence of each component of an input vector on an
output neuron by a specific subnetwork of weight pathways; it helps us to infer whether
such an input component’s influence over that output neuron is positive or negative by
the subnetwork. By decomposing a neural network into subnetworks, we can
investigate how each part of the neural network works individually and cooperatively
to achieve the goal of classification and the goal of training.
Suppose samples are given by images (with single grey channel) that consist of
𝑀 × 𝑀 bitmap pixels, and the representation vector of the μth sample is coded as
(0)
(𝑥𝑖0 (𝜇), 𝑖0 = 𝑀𝛼 + 𝛽, 𝛼 = 1, ⋯ , 𝑀, 𝛽 = 1, ⋯ , 𝑀 ). By plotting a two-dimensional
heat map as the following
(𝑖 ,𝑙) (𝑖 ,𝑙)
𝑙𝛼𝛽𝑙 (𝑖𝐿 ) = 𝑐𝑖0𝑙𝑖𝐿 （5）,
we have a visualization of the penetration coefficients. Note that penetration coefficient
(𝑖 ,𝑙)
𝑐𝑖0𝑙𝑖𝐿 can either be positive or negative, corresponding to p-pathway dominant or n-

(𝑖 ,𝑙)
pathway dominant respectively, and the corresponding heat map 𝑙𝛼𝛽𝑙 (𝑖𝐿 ) pixels are
positively displayed (in shades of red pseudo colors) or negatively displayed (in shades
of blue pseudo colors) respectively. The resulting heat map image will have patterns of
positively or negatively displayed regions, corresponding to regions of input pixels that
would be connected to the output neuron with positive or negative penetration
coefficients. With these patterns, one can infer the enhancement or suppression effect
to a given sample image at the pixel level following the characteristic map. The
visualization can be considered a view of the internal structure of a subnetwork. We
(𝑖 ,𝑙)
thus call heat map 𝑙𝛼𝛽𝑙 (𝑖𝐿 ) the radiograph of the subnetwork pertaining to the 𝑖𝑙 th
neuron in the 𝑙 th layer, and call the patterns in the radiograph the mode of the
subnetwork or the mode of this neuron.

2.2 WPA approach

We illustrate the basic principles of the WPA approach with three sets of toy
samples. The input samples are 100x100 bitmaps, and they all contain an identical circle
at different position. The pixels inside the circle (the face zone) are assigned a value of
e = 1, 2, or 3, while the pixels outside of the circle (the ground zone) are always assigned
a value of -1. Fig. 2(a) shows the three samples of the first training set, where circles
with value e=1, 2, and 3 are shown in yellow, light brown and dark brown colors,
respectively. There is no overlap between the face regions in first training set. Fig. 2(b)
shows the second training set. The difference from the first one is that the face regions
of these three samples completely overlap with each other. Fig.2 (c) shows the third set
of training samples. The difference from the first set is that the face regions of the three
samples are partially overlapped. Each sample can be vectorized as a 𝑁 = 100 ×
100 = 104 dimensional vector. We train a 104 − 200 − 3 neural network to perform
the task of classification; the cost function is defined by (2). Each sample represents
one class; the first to the third output neurons are the label neurons of the first to the
(2)
third sample or class, respectively. The achievement of classification means ℎ𝑖2 (𝜇) is
the largest if 𝜇 = 𝑖2 , while the achievement of the training goal means
(2)
ℎ𝑖2 (𝜇)𝑦𝑖2 (𝜇) = 𝑑. It will be seen that the second sample of the second set is linearly
inseparable, and the second sample of the third set has both linearly separable and
inseparable features. All the other samples are linearly separable.
Fig. 2 The visualization of penetrating coefficients. (a)-(c) show the first, second, and
third sample sets. The first line gives overlap images of the three samples. (d)-(f)
show scenographs of penetrating coefficients of subnetworks for the three sample
sets, respectively. The first to the third columns are for the first to the third output
neurons. The first to third rows are for the LNN30, the NNN30 and NNN200,
respectively. (g) shows scenographs of penetrating coefficients of subnetworks of
hidden-layer neurons classified by the LNI for the NNN30. The plots are managed
flowing the order of the matrix Π(𝜇|𝑖2 ) with 𝑖2 , 𝜇 = 1,2,3. (h) is same as (g) but for
the NNN200. (i) shows the scenographs of penetrating coefficients of subnetworks
of hidden-layer neurons of the ⁡ 𝐶1 to 𝐶6 classes for the second sample set and
the NNN30. Here the first to third columns are for the first to third output neurons.
(j) and (k) are same as (g) and (h) but for the third sample sets.
 Fig. 3 The transition from the linear learning mode to the nonlinear learning mode. (a)-(c)
(2)
show 〈𝑥𝑖𝐿 (𝜇)𝑦𝑖𝐿 (𝜇)〉 as a function of training time (MC steps) for the three sample sets,

respectively. The solid(green), dot lines(blue) and dashed(red) lines for the LNN30, the
NNN30 and the NNN200. (d)-(f) show the LNI for the three sample sets, respectively. The
first, second and third columns are for the LNN30, the NNN30 and the NNN200,
respectively. The parameter 𝑖2 indicates the order of the output neuron. (g) The transition
from the phase of the linear learning mode to the mixture phase of linear and nonlinear
learning modes. The triangles, circles, and stars represent Π(1|2) , Π(2|2) and Π(3|2)
obtained after these quantities become stationary after a sufficient long training times.

As a reference, we first train these three training sets with linear neurons of
f(h)=0.002h and with d=30. We shorthand this neural network as LNN30. In Fig. 3(a)-
(2)
3(c), we show the evolution of local fields of output neurons 〈𝑥𝑖𝐿 (𝜇)𝑦𝑖𝐿 (𝜇)〉 with
training time as blue dotted lines. Here, 〈∙〉 is averaged over the three samples and
three output neurons. We see that the goal of training is reached for both the first and
(2)
the third sets. For the second set, it has 〈𝑥𝑖𝐿 (𝜇)𝑦𝑖𝐿 (𝜇)〉 ≈ 20. Averaging only over the

(2) (2)
three samples, we found that 〈𝑥1 (𝜇)𝑦1 (𝜇)〉 ≈ 〈𝑥3 (𝜇)𝑦3 (𝜇)〉 ≈ 30 for the first and

(2)
the third output neurons, but 〈𝑥2 (𝜇)𝑦2 (𝜇)〉 ≈ 0 for the second output neuron. It can
be checked that the output of the second sample is never the largest.
We then train these training sets with nonlinear neurons of f (h)= tanh (0.002h)
and with d=30. We shorthand this neural network as NNN30. The local fields of output
(2)
neurons 〈𝑥𝑖𝐿 (𝜇)𝑦𝑖𝐿 (𝜇)〉 are shown as green solid lines in Fig. 3(a)-3(c). It can be seen
that the goal of training is also completely fulfilled for the first and third training sets.
(2)
For the second training set, it has 〈𝑥𝑖𝐿 (𝜇)𝑦𝑖𝐿 (𝜇)〉 ≈ 28. Again, averaging only over

(2) (2)
the three samples we find that 〈𝑥1 (𝜇)𝑦1 (𝜇)〉 ≈ 〈𝑥3 (𝜇)𝑦3 (𝜇)〉 ≈ 30 , and

〈𝑥2(2) (𝜇)𝑦2 (𝜇)〉 ≈ 25. Thus, both the goal of classification and the goal of training are
fully achieved for the first and the third samples. The goal of training is not fully
(2)
achieved for the second sample; however, it can be checked that 𝑥2 (2) is larger than

(2) (2)
𝑥2 (1) and 𝑥2 (3), i.e., the goal of classification is achieved. When d is increased to
d=200 (We shorthand this neural network as NNN200), the goal of classification is
achieved for all training sets, but the goal of training is not fully achieved by any of the
training sets., as shown by the red dashed lines in Fig.3(a)-(c). For the first and third
(2) (2)
sets, 〈𝑥𝑖𝐿 (𝜇)𝑦𝑖𝐿 (𝜇)〉 ≈ 100 . In detail, it can be checked that 〈𝑥1 (𝜇)𝑦1 (𝜇)〉 ≈

〈𝑥2(2) (𝜇)𝑦2 (𝜇)〉 ≈ 〈𝑥3(2) (𝜇)𝑦3 (𝜇)〉 ≈ 100 for these sets. For the second set,

〈𝑥𝑖(2) (𝜇)𝑦𝑖𝐿 (𝜇)〉 ≈ 70 and it can be checked that 〈𝑥1(2) (𝜇)𝑦1 (𝜇)〉 ≈ 90 ,
𝐿

〈𝑥2(2) (𝜇)𝑦2 (𝜇)〉 ≈ 50 and 〈𝑥3(2) (𝜇)𝑦3 (𝜇)〉 ≈ 75, where 〈∙〉 is averaged over the three
samples.

2.2.1 Holographic structure

After all the training have finished, we show radiographs of subnetworks of the
(1,2) (2,2) (3,2)
three output neurons 𝑙𝛼𝛽 (1), 𝑙𝛼𝛽 (2), and 𝑙𝛼𝛽 (3) for the LNN30, the NNN30,
and the NNN200 in Fig.2 (d)-2(f), where the columns correspond to the output neurons
and the rows correspond to the three types of neural network. For better visualization,
we calculate an ensemble average of penetration coefficients of 16 replicas of neural
networks for each training set. We see that each visualization image clearly shows the
patterns of all the samples in the training set, indicating that each subnetwork
"holographically" encodes all samples of a training set. Some zones are positively
displayed while others are negatively displayed, implying that the subnetwork of weight
pathways establishes coherent structures for enhancement or suppression of different
features. The "holographic" structure reveals how a neural network stores information,
and indicates that classifications are achieved through the interaction between the input
samples and the "holographic" structure. In the following, we will explain how the
"holographic" structures are formed and in what a way they work.

2.2.2 The extraction of linearly separable feature and the classification of hidden-

layer neurons.

We study the first training set in this section. For this set, the characteristic map
(𝑖 ,𝑙)
can be decomposed into four parts, 𝐻𝑖2 𝑙 (𝜇) = c(G)x(G) + c(𝐹1 )x(𝐹1 ) +
c(𝐹2 )x(𝐹2 ) + c(𝐹3 )x(𝐹3 ) , which respectively represent the contributions of the
common ground zone (zone G), the face zone of the first sample (zone 𝐹1 ), the face
zone of the second sample (zone 𝐹2 ), and the face zone of the third sample (zone 𝐹3 ).
Here c(G) , c(𝐹1 ) , c(𝐹2 ) and c(𝐹3 ) is obtained by summarizing the penetration
(𝑖 ,𝑙)
coefficients within the zone, i.e., ∑𝑖0 𝑐𝑖0𝑙𝑖𝐿 , over the zones of G, 𝐹1 , 𝐹2 and 𝐹3 zones,
respectively. From Fig 2(d) we see that the polarity of each zone stays the same for all
three rows, meaning that the modes of output neurons of either the LNN or the NNNs
with a small d and a large d are almost same.
How the goal of classification is achieved can be revealed by the patterns of the
radiographs shown in Fig. 2(d). These patterns indicate the modes of the three output
neurons. As the positive and negative penetration coefficients represent the
enhancement or the suppression effect, the signs of c(G), c(𝐹1 ), c(𝐹2 ) and c(𝐹3 )
thus are essential and are employed to represent the mode of a neuron. From Fig. 2(d)
we see that (sign(c(G)), sign(c(𝐹1 )), sign(c(𝐹2 )), sign(c(𝐹3 ))) = (−, +, −, −)
(+, −, +, −), and (+, −, −, +) for the three output neurons, respectively. Combining
the input values in the four zones (x(G), x(𝐹1 ), x(𝐹2 ), x(𝐹3 )) as (−1,1, −1, −1)，

(−1, −1,2, −1)，and (−1, −1, −1,3) for the first, the second, and the third samples

respectively, we know either positive or negative contributions to each output neuron

from each zone of an input sample, i.e., we know the signs of the contribution of each
(𝑖,2)
zone to 𝐻𝑖2 (𝜇). For example, since the first output neuron has the mode (−, +, −, −)，
the signs of the contribution from each zone to the characteristic map are
(1,2)
𝐻1 (1)⁡ ~(+, +, +, +), 𝐻1(1,2) (2)⁡ ~(+, −, −, +), and 𝐻1(1,2) (3)⁡ ~(+, −, +, −), for
the first, the second, and the third samples. We see that the contribution from the 𝐹1
(1,2)
zone to 𝐻1 (1) is positive, while that to 𝐻1(1,2) (2) and 𝐻1(1,2) (3) are both negative.

(1,2)
So as long as c(𝐹1 ) is large enough, the condition of 𝐻1 (1) > 0, 𝐻1(1,2) (2) < 0

(1,2)
and 𝐻1 (3) < 0 can always be reached. Similar argument can be applied to the
second and third output neurons, and one can thus get positive local field for label
neurons (𝜇 = 𝑖𝐿 ) while keeping all the non-label neurons (𝜇 ≠ 𝑖𝐿 ) negative. Hence
the goal of classification can always be achieved.
Following the modes of the output neurons, one can construct a linear
classification hyperplane for each sample. For example, a vector in the input vector
space can be selected as the classification vector, whose components that lie within the
G, 𝐹1 , 𝐹2 and 𝐹3 zones have signs of (−, +, −, −) , as the signs of
(c(G), c(𝐹1 ), c(𝐹2 ), c(𝐹3 ) have. Then, the projection of the first sample to this vector
must be the largest among the three input samples. The existence of such classification
vectors indicates that samples in this set are all linearly separable. This conclusion is
also confirmed by the fact that samples in this training set can be classified by the LNN.
 (2)
However, the goal of training, i.e., the requirement that ℎ𝑖2 (𝜇)𝑦𝑖2 (𝜇) = d for
each sample as well as for each output neuron, generates additional tasks to the hidden-
layer neurons; they will accomplish this goal by working cooperatively. Specifically,
the hidden-layer neurons self-organize into classes that are specific to each sample and
to each output neuron. One way to capture such self-organization is to count the number
of hidden-layer neurons that give the largest contribution to each output neuron when a
specific sample is fed to the neural network and study the distribution of such counts.
We denote the number of the largest-contribution hidden-layer neurons from input
sample 𝜇 for the 𝑖2 th output neuron as Π(𝜇|𝑖2 ) , and call it largest contribution
hidden-layer neuron classification index, or simply the largest neuron index (LNI).
Without hidden-layer neuron self-organization, LNI of the label neuron would equal to
the total number of hidden-layer neurons while that of the other output neurons would
equal to zero. Deviation from such a behavior is an indication of the occurrence of
classification of the hidden-layer neurons.
Fig. 3(d) shows the training time evolution of the LNIs of the first, the second, and
the third output neurons in the LNN30 for the first training set. In each sub-plot, the
solid, dashed and dot lines represents the number of the largest contribution neurons of
the first, second and third samples. Similar LNI training time evolutions of the NNN30
and NNN200 are also shown in Fig.3(d). We see that although the label neuron has
biggest LNI, i.e.,Π(𝜇|𝑖2 ) has the biggest value when 𝜇 = 𝑖2 , there are still a lot of
neurons give their largest contribution to non-label output neurons. This fact reveals the
differentiation and self-organization of the hidden-layer neurons.
The classification of hidden-layer neurons can provide more information on how
the cost function is minimized. In Fig. 2(g)-2(h), we show radiographs of subnetworks
of hidden-layer neurons classified by Π(𝜇|𝑖2 ) for the NNN30 (Fig. 2(g)) and the
(Π(𝜇 |𝑖 ),1)
NNN200 (Fig.2(h)). We denote the radiographs obtained in this way as 𝑙𝛼𝛽 2 (𝜇),
where Π(𝜇|𝑖2 ) specifies the class of hidden-layer neurons. The radiographs for the
LNN are similar to that of NN30 and are not shown here. Taking the first output neuron
of the NNN30 as an example (the first column of Fig.2(g)), we see that the radiograph
(Π(1|1),1)
𝑙𝛼𝛽 (1) (the first row and the first column of Fig.2(g)) shows mode of

(1,2)
(c(G), c(𝐹1 ), c(𝐹2 ), c(𝐹3 ))~(−, +, −, −), which is the same as that of 𝑙𝛼𝛽 (1). With

(Π(1|1),1)
this mode the characteristic map gives 𝐻1 (1)⁡ ~(+, +, +, +) ,

(Π(1|1),1)
𝐻1 (2)⁡ ~(+, −, −, +) , and 𝐻1(Π(1|1),1) (3)⁡ ~(+, −, +, −) for the first, the
second, and the third samples. Since this class of neurons has the largest population
(Fig. 3(d)), it can achieve the goal of classification, as discussed early. We call these
(Π(2|1),1)
hidden-layer neurons the dominant neurons. Meanwhile, the radiograph 𝑙𝛼𝛽 (1)
(the second row and the first column of Fig.2(g)) shows the mode of
(c(G), c(𝐹1 ), c(𝐹2 ), c(𝐹3 ))~(+, +, +, −) . The characteristic map gives
(Π(2|1),1) (Π(2|1),1)
𝐻1 (1)⁡ ~(−, +, −, +) , 𝐻1 (2)⁡ ~(−, −, +, +) , and
(Π(2|1),1) (Π(3|1),1)
𝐻1 (3)⁡ ~(−, −, −, −) . The radiograph 𝑙𝛼𝛽 (1) (the third row and the
first column of Fig.2(g)) shows the mode of (c(G), c(𝐹1 ), c(𝐹2 ), c(𝐹3 ))~(+, +, −, +).
(Π(3|1),1)
The characteristic map gives 𝐻1 (1)⁡ ~(−, +, −, +) ,

(Π(3|1),1)
𝐻1 (2)⁡ ~(−, −, +, +) , and 𝐻1(Π(3|1),1) (3)⁡ ~(−, −, −, −) . Obviously, the
differentiation of hidden-layer neurons provides more freedom to adjust the outputs.
For example, neurons of Π(2|1) class decreases the output of the third sample, and
(2)
neurons of Π(3|1) class decreases the output of the second sample, leading ℎ1 (2)

(2) (2) (2)

and ℎ1 (3) move towards ℎ1 (2) = ℎ1 (3) = −d . Therefore, hidden-layer
neurons with different modes provide more freedom to adjust the outputs to fulfill the
goal of training. We call these classes the auxiliary neurons.
(1,2) (1,2)
For the NNN200, we see the same radiographs of 𝑙𝛼𝛽 (1) , 𝑙𝛼𝛽 (2) and

(1,2)
𝑙𝛼𝛽 (3) (see the third row of Fig.2(d)) as those for the NNN30; the time evolutions of
LNIs are also similar (see Fig.3(d)), but the radiographs of subnetworks of neurons
classified by the LNI show certain significant differences. The G zones of radiographs
almost all reverse their sign relative to those of the NNN30. This phenomenon is a result
of self-organization for the purpose of maximum minimization of the cost function.
From the patterns one can also find out how the self-organization is performed. For
example, Π(2|3) class almost disappears, so the G zone of Π(1|3) class has to
become highly positive in order to decrease the local field of the third output neuron
when inputting the first and the second samples. Note that due to the limits put on the
norm of the weights, the cost function has not been fully minimized to zero.

2.2.3 The extraction of linearly inseparable feature

For the second training set, each sample has only a ground zone (G) and a face
zone (F)，and we thus can perform a detail analysis about the modes of the neurons

before report the results. In this case, a neuron has four possible modes,
(c(G), c(𝐹))~(+, −), (−, +),(+, +), (−, −). Since x(𝐺) = −1 and x(𝐹)= 1, 2, 3 for

(𝑖 ,𝑙)
the three samples, characteristic map 𝐻𝑖2 𝑙 (𝜇) = c(G)x(G) + c(𝐹)x(𝐹) can be
determined directly by neuron modes. For example, mode (+, −) leads to
 (𝑖 ,𝑙) (𝑖,𝑙)
𝐻𝑖2 𝑙 (𝜇) <0 and mode (−, +) leads to 𝐻𝑖2 (𝜇)>0 for all three samples. While mode

(+, +) has c(G)x(G) <0 and c(F)x(F) > 0 , and thus 𝐻𝑖(𝑖,𝑙) (𝜇) may have different
2

signs for different samples depending on the amplitudes of c(G) and c(F). In more
details, for a fixed value of c(G)x(G)<0, depending on the amplitude of c(F), it has
(𝑖,𝑙)
four possible combinations of 𝐻𝑖2 (𝜇) when inputting the three samples, i.e.,

(𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙)

𝐻𝑖2 (1) < 𝐻𝑖2 (2) < 𝐻𝑖2 (3) < 0 ; 𝐻𝑖2 (1) < 𝐻𝑖2 (2) < 0 and 𝐻𝑖2 (3) > 0;

(𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙)

𝐻𝑖2 (1) < 0 and 0 < 𝐻𝑖2 (2) < 𝐻𝑖2 (3); 0 < 𝐻𝑖2 (1) < 𝐻𝑖2 (2) < 𝐻𝑖2 (3) .
We denote them as (+, +)1, (+, +)2 , (+, +)3 , and (+, +)4 , respectively. Similarly,
(𝑖,𝑙)
for the mode (−, −) it has c(G)x(G)>0 and c(F)x(F) < 0, and therefore 𝐻𝑖2 (𝜇)

(𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙)

may have combinations 0 < 𝐻𝑖2 (1) < 𝐻𝑖2 (2) < 𝐻𝑖2 (3) ; 0 < 𝐻𝑖2 (1) <

(𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙)

𝐻𝑖2 (2) and 𝐻𝑖2 (3) < 0; 0 < 𝐻𝑖2 (1) and 𝐻𝑖2 (2) < 𝐻𝑖2 (3) < 0 ;

(𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙)

𝐻𝑖2 (1) < 𝐻𝑖2 (2) < 𝐻𝑖2 (3) < 0. We denote them as (−, −)1, (−, −)2 , (−, −)3,
and (−, −)4 , respectively. Therefore, there are a total of 10 neuron modes for the
second training set and that these results are applicable for both linear and nonlinear
neurons if the neuron transport function is monotonous.
Therefore, neurons or subnetworks with mode (−, −)3 can achieve the goal of
(𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙)
classification for the first sample because 0 < 𝐻𝑖2 (1) and 𝐻𝑖2 (2) < 𝐻𝑖2 (3) <
0 , and neurons or subnetworks with mode (+, +)2 can achieve the goal of
(𝑖,𝑙) (𝑖,𝑙) (𝑖,𝑙)
classification for the third sample because 𝐻𝑖2 (1) < 𝐻𝑖2 (2) < 0 and 𝐻𝑖2 (3) >
0. Indeed, these two samples are linearly separable. For example, one can construct a
classification vector, whose components that lie within the G and F zones would have
signs as sign(x(G), x(𝐹)) = sign(c(G), c(𝐹)) = (−, −) , then the projection of the
first sample to this vector must be the largest among the three input samples because it
has the smallest x(𝐹)=1.
However，we see that there is no mode that can convert the output of the second

sample to the largest value. This is because its face zone is identical to those of the other
samples and its pixel value is between that of the first and third ones; therefore, the
monotonic transformation cannot make its output largest in any way. In other words, a
single neuron or a single subnetwork alone cannot realize the classification of the
second sample. This sample is thus linearly inseparable, which is confirmed by the fact
that a LNN cannot classify this sample.
 Fig.4 The formation of the nonlinear learning mode. The right-hand and left-hand
plots of (a) show the modes of (+, +)3 and (−, −)2 of hidden-layer neurons. (b)
and (c) show the number of neurons in the six classes (see text), for the NNN30 and
the NNN200, respectively.

Therefore, the fact that the NNN can achieve the goal of classification implies that
it makes use of combinations of neurons with different modes. As shown in Fig. 4(a), a
pair of neurons with (+, +)3 and (−, −)2 modes can convert the output of the second
sample to be the largest, as long as the outputs of the first and third samples lie on the
nonlinear region of tanh⁡(ℎ) ≈ ±1 . In this case, the outputs of the first and third
samples cancel approximately, while that of the second sample appears as a large
positive quantity since both neurons contribute a positive term. It is easy to check that
the conversion can also be realized by combinations of (+, +)3 and (−, −)3 , (+, +)2
and (−, −)2 , though they are less efficient at converting. We emphasize that the
nonlinearity of the transfer function plays a key role here. With a linear neural transfer
function, f(ℎ) = 𝑘ℎ, (+, +)3 and (−, −)2 must lead to the combined outputs of the
second sample vanishing if one wants to offset the outputs of the first and the third
samples.
It is obvious that, for the first output neuron, for example, the goal of training
cannot be achieved by neurons only with modes of (−, −)3 , since it gives 0 >
(2) (2) (2) (2)
ℎ2 (2) > ℎ2 (3) and thus fails to approach the condition of ℎ2 (2) = ℎ2 (3) =
−d. Therefore, auxiliary neurons with different mode must be utilized. In Fig.3(e) we
show the LNI for this training set. It clearly confirms the differentiation of hidden-layer
neurons, as the non-vanishing Π(1|𝑖2 ) and Π(3|𝑖2 ) indicate that each output neuron
has hidden-layer neurons in different modes. The fact that Π(2|𝑖2 ) vanishes indicates
that there is no single neuron can convert the output of the second sample to be the
largest. Note that neurons in a class of Π(𝜇|𝑖2 ) may have several neuron modes. For
example, neurons of Π(1|1) may involve modes of (−, −)2 , (−, −)3 and (+, +)4 ,
since all of them have the largest contribution to the first output neuron from the first
input sample. For a more detailed investigation, we can check the distribution of
neurons for all of the possible modes. Note that neuron modes of this training set can
be divided into six classes: (+, −), (+, +)1 and (−, −)4 leading to negative outputs
for all samples and are classified into one class (designated as C1 class); (−, +)，
(+, +)4 and (−, −)1 leading to positive outputs for all samples and are classified into
another class (designated as C2 class); For the rest modes: (+, +)2 , (+, +)3 ,
(−, −)2 , and (−, −)3, each represents a different class (designated as C3 , C4 ,⁡ C5 , and
C6 classes respectively). For the sake of simplicity, we calculate the number of hidden-
layer neurons of the six classes instead of the 10 modes for the NNN30 and NNN200,
and show them in Fig. 4(b)-4(c). This approach provides another classification standard
for checking the differentiation. The results reveal more details of the classification of
hidden-layer neurons than LNI does.
With these details, we can understand how the classification of hidden-layer
neurons is essential for minimizing the cost function. Using the first output neuron as
an example again，it can be seen that although (−, −)3 accounts for a large proportion,

neurons of all the six classes exist. Neurons with mode (+, +)2 raises the output of
(2) (2)
the third sample, leading the local fields evolve towards ℎ2 (2) = ℎ2 (3) = −𝑑 .
The 𝐶1 class decreases the outputs globally, but can also adjust the amplitude of
contributions for different samples. In such a way, the local fields evolve towards the
goal of minimizing the cost function to zero. For this purpose, neurons with mode
(−, −)3 and (+, +)2 , as well as the modes in the 𝐶1 class are more essential;
therefore, in the case of NNN200, these modes remain but others are suppressed
dramatically. Due to the insufficient number of hidden-layer neurons, minimization of
the cost function can only give 〈𝑥𝑖𝐿 (𝜇)𝑦𝑖𝐿 (𝜇)〉 ≈ 70 , and due to the insufficient
number of auxiliary neurons, it has 〈𝑥1 (𝜇)𝑦1 (𝜇)〉 ≈ 90 , 〈𝑥2 (𝜇)𝑦2 (𝜇)〉 ≈ 50 and
〈𝑥3 (𝜇)𝑦3 (𝜇)〉 ≈ 75.
(𝐶 ,1)
In Fig 2 (i), we show radiographs 𝑙𝛼𝛽𝑖 (𝑖𝐿 ) of subnetworks of the 6 classes for
the NNN30 as examples. We see that they are in well agreement with the patterns of
(𝑖,2)
modes of that class. Note that radiograph 𝑙𝛼𝛽 (𝑖𝐿 ) of Fig.2(e) is the result of
superposition of radiographs of the 6 classes. For the first and the third output neurons,
since hidden-layer neurons with modes (−, −)3 and (+, +)2 have the largest
(1,2) (3,2)
populations respectively, 𝑙𝛼𝛽 (1) and 𝑙𝛼𝛽 (3) thus show patterns of

(c(G), c(𝐹))~(−, −) and ~(+, +) . For the second output neuron, neurons with
modes of (+, +)3 and (−, −)2 have roughly equal numbers, (see Fig. 4(b)), patterns
(2,2)
resulted by them thus are almost canceled with each other. The pattern of 𝑙𝛼𝛽 (2)，

(c(G), c(𝐹)) = (+, −), should be caused by neurons of mode (+, −) in the 𝐶1 class.
Note that according to Fig. 4(b) this class has a large number of hidden-layer neurons.
Therefore, extracting the linearly separable and inseparable features has an
essential difference. For the linearly separable feature, neurons can independently
achieve the goal of classification. The differentiation of hidden-layer neurons is for the
purpose of achieving the goal of training. Specifically, for the first training set, a LNN
can achieve the goal of classification with similar modes or radiographs as those of a
NNN, see Fig.2(d) and Fig3(d). We call this way of learning the linear learning mode.
For linearly inseparable features, for example, the second sample of the second training
set, the neural network must use a combination of neurons with different mode and
relay on the nonlinearity of the neuron transfer function to convert the output to be the
largest. Linear networks cannot achieve such a goal. We call this way of learning the
nonlinear learning mode. In principle, the nonlinear learning mode can extract both
linearly separable and inseparable features; however, the nonlinear learning mode
requires more network resources and hence is reserved for the extraction of the
nonlinear features or for the purpose of reaching the difficult goal of training. As shown
(2)
in Fig. 3(b), even in the case of d =30, the goal of ℎ2 (2) = 𝑑 has not been strictly
reached because the number of auxiliary neurons is not sufficient (Fig. 4(c)).

3.2.4 The extraction of co-existing linearly separable and inseparable features.

In the third sample set, the second sample has both linearly separable features (the
part of the face zone that do not overlap with any of the face zones of the other samples,
which shall be called the unit linear features of the sample, and the part of the face zone
that only overlap with some but not all of the face zones of the other samples.) and
linearly inseparable features (the part of the face zone that overlap with all of the face
zones of the other samples). Therefore, it can be used to investigate how the neural
network works in the coexistence of linearly separable and inseparable features. The
first and third samples have only linearly separable features. For the convenience of
discussion below, we denote zones of the unit linear feature of the three samples by 𝐹10 ,
𝐹20 , and 𝐹30 , respectively.
Fig. 3 (f) shows the training time evolution of the LNI in the LNN30, the NNN30,
and the NNN200 in the first to the third column, respectively. We see that the LNI of
the NNN30 behaves almost exactly as that of the LNN30. Furthermore, Fig. 2 (f) shows
that the radiographs of these two neural networks are also qualitatively similar.
Therefore, the NNN30 behaves almost identically to that of the LNN30, indicating that
the nonlinear network invokes only the linear learning mode and extracts only the
linearly separable features.
Radiographs of output neurons can reveal how the linear learning mode works.
(2,2)
For example, from 𝑙𝛼𝛽 (2) in Fig. 2(f) we see that 𝐶(𝐹20 ) is positive, and 𝐶(𝐹10 )
and 𝐶(𝐹30 ) are negative. Following the characteristic map, the signs of contributions
(2,2)
of these zones to the outputs should be 𝐻2 (1) ∝ x(𝐹10 )𝐶(𝐹10 ) + x(𝐺)𝐶(𝐹20 ) +

(2,2)
x(𝐺)𝐶(𝐹30 ) ~(−, −, +) , 𝐻2 (2) ∝ x(𝐺)𝐶(𝐹10 ) + x(𝐹20 )𝐶(𝐹20 ) + x(𝐺)𝐶(𝐹30 )

(2,2)
~(+, +, +) , and 𝐻2 (3) ∝ x(𝐺)𝐶(𝐹10 ) + x(𝐺)𝐶(𝐹20 ) + x(𝐹30 )𝐶(𝐹30 )~(+, −, −),
respectively. Similar to the case of the first sample set, these modes provide the freedom
to fulfill the goal of classification, by amplifying only the amplitude of p-pathways in
𝐹20 zone for example.
The subnetworks of hidden-layer neurons classified by the LNI reveal more details.
In second column of Fig.2(j) we show the radiographs of the three class, Π(1|2) ,
Π(2|2) and Π(3|2), of hidden-layer neurons for NNN30 respect to the second output
neuron, respectively. The non-vanishing of Π(1|2) and Π(3|2) reveal the
phenomenon of differentiation. The Π(2|2) class takes the dominant population of
hidden-layer neurons (see Fig.3(f)), which confirms that the classification information
extracted by single neurons and thus should be linearly separable feature. As a result,
(2,2) (Π(2|2),1)
the 𝑙𝛼𝛽 (2) is qualitatively identical to 𝑙𝛼𝛽 (2) since the former is also
dominated by the class of Π(2|2). The role of classes Π(1|2) and Π(3|2) is to assist
the Π(2|2) class to fulfil the goal of training. For example, the center zones of
(Π(2|2),1) (Π(3|2),1)
𝑙𝛼𝛽 (2) and 𝑙𝛼𝛽 (2) are positive, which give positive contributions to the
second output from all the samples, of which that of the third sample is the largest since
the pixel value of its face zone is the largest. To offset this positive contribution from
(2) (Π(1|2),1)
the third sample, i.e., ℎ2 (3), the face zone of the third sample in 𝑙𝛼𝛽 (2) is
negative, and thus the Π(1|2) class can decrease the output of the third sample, which
(2) (2)
provide the freedom to help ℎ2 (3) move towards ℎ2 (3) = −𝑑.
These facts indicate that applying a NNN does not necessarily mean the extraction
of the linearly inseparable features when both linearly separable and linearly
inseparable features exist in the samples. The reason is that the linear learning mode is
performed by hidden-layer neurons independently, while the nonlinear learning mode
needs to have multiple neurons working cooperatively and thus is more difficult to be
manifested. So, if the goal of training can be achieved using only the linear separable
features, the neural network would avoid to activate the nonlinear learning mode.
Therefore, one may be at risk of losing the information of linear inseparable feature in
the case of coexistence of linearly separable and inseparable features even using a NNN.
However, since we limit the norm of weights, and thus limit the products of weight
pathways. Once the parameter d exceeds a threshold, weight pathways connecting only
the linearly separable zones will not be able to fulfil the goal of training. In this case,
the neural network has to initiate the nonlinear learning mode to extract the linearly
inseparable features to further drive the local fields towards the goal of training. For the
NNN200, we see from Fig.2(f) and Fig. 3(f) that there is no substantial change for the
first and the third output neurons for both the radiographs and the LNI’s, indicating that
the neural network still extracts information with the linear learning mode, since in
these cases all the information are the linearly separable feature. However, remarkable
(2,2)
changes can be seen in the case of second output neuron. The center zone of 𝑙𝛼𝛽 (2)

(2,2)
changes to the same pseudo-color as that of the second sample set (𝑙𝛼𝛽 (2) for the two
NNNs, see Fig. 2(e)). Fig. 3(f) shows that the Π(2|2) is no longer the largest, instead,
Π(1|2) turns to be the largest as in the case of the second sample set (see Fig. 3(e) for
the two NNNs).
These facts indicate that the nonlinear learning mode is initiated for the second
sample, i.e., the features of the center zone of this sample begins to contribute to the
(2)
local field ℎ2 (2), which can be observed more clearly by studying radiographs of
subnetworks of hidden-layer neurons classified by the LNI. Fig.2 (k) shows these
(2,2) (Π(1|2),1)
images for the NNN200. It is seen that 𝑙𝛼𝛽 (2) and 𝑙𝛼𝛽 (2) have the same
patterns, as a result of Π(1|2) class being the dominant population of neurons. The
radiograph of Π(2|2) shows non-substantial changes to that of the NNN30, indicating
that the main role of this class is still for extracting the linearly separable features of the
(Π(1|2),1)
second sample. But radiographs of Π(1|2) and Π(3|2) , i.e., 𝑙𝛼𝛽 (2) and

(Π(3|2),1)
𝑙𝛼𝛽 (2), have changed substantially from that of Fig. 2(j), indicating a different
learning mode is involved. For the Π(1|2) class, by definition, each neuron gives the
largest contribution to the second output neuron with the input of the first sample. The
(Π(1|2),1)
center zone of its radiograph 𝑙𝛼𝛽 (2) is negative, making contribution from the
third sample to the same output neuron the most negative or the smallest in value. In
addition, from the radiograph we see that the unit linear feature zone 𝐹20 is positive,
and thus making the contribution from the second sample larger. As a result, for this
(Π(1|2),1) (Π(1|2),1) (Π(1|2),1)
class we have 𝐻2 (1) > 𝐻2 (2) > 𝐻2 (3) . Similar analysis can

(Π(3|2),1) (Π(3|2),1) (Π(3|2),1)

lead to 𝐻2 (1) < 𝐻2 (2) < 𝐻2 (3) for the Π(3|2) class.
Therefore, these two classes of neurons play the role of modes (+, +)3 (−, −)2 in Fig.
4 and perform the extraction of the linearly inseparable features of the second sample
in this sample set.
Therefore, in the case of the NNN30, the majority of the hidden-layer neurons is
employed to extracting the linearly separable feature. The goal of training is achieved
by this kind of neurons with the help of some auxiliary neurons, and the excitation of
the nonlinear learning mode is not necessary. In the case the NNN200, the goal of
training cannot be achieved by learning just the linearly separable features, and the
nonlinear learning mode has to be initiated. In Fig 3(g), we plot Π(1|2), Π(2|2) and
Π(3|2) of the NNN as a function of control parameter d. Each data point is obtained
after the training becomes stationary. It can be seen that Π(2|2) begins to decrease
around d=60, and Π(1|2) begins to increase and becomes dominant after d=140. This
plot confirms that the neural network tends to use linear learning mode to extract linear
features first, and starts to invoke the nonlinear learning mode only when linear features
are insufficient to approach the goal of training.
 Fig. 5. The visualization of penetrating coefficients of neural networks of MINIST data
set. The first line: ten digits picking up from the data set. Two to four lines: the
radiographs of NNNs with d=30 trained by the first three, first five and all of the ten
digits. Five line: the radiographs of the NNN with d=75 trained by the ten digits. Six to
eight lines: the radiographs of the LNN and with d=30, the NNNs with d=30 and d=70,
trained by 600 MINIST samples.

3 Application.

3.1 Extracting the linearly separable and inseparable features of

handwritten digits.

The first row of Fig. 5 shows 10 handwritten digits selected from the MNIST set.
Since each sample is composed of a 28x28 bitmap, we design a 784-600-P network
with P =3, 5, 10 to classify the first 3, first 5, and all 10 samples, respectively, with
(𝑖 ,2)
control parameters d=30 and β=0.15. Radiographs of label neurons, i.e., 𝑙𝛼𝛽𝐿 (𝑖𝐿 ) for
𝑖𝐿 = 1,2, … , 𝑃, are shown in the second to the fourth rows in Fig. 5, respectively. The
radiographs are obtained after the neural network has been trained to have its cost
function less than 0.01. It can be seen that in the case of only a few samples, radiographs
are similar to those of the toy sample sets, i.e., patterns of all the digits in the training
set distinctly appear in every radiograph, indicating that the “holographic” nature of the
neural network remains. With the increase of the number of samples, the pattern of the
digit corresponding to a label neuron can still positively recognized, while patterns of
other digits gradually become less distinguishable. By studying the progressive trend
of radiographs from the cases of P=3 to P=5 and to P=10, one can realize that even for
the case of P=10, the patterns of all the digits are indeed still there; they overlap with
each other to form the negatively displayed region. In other words, the “holographic”
structure exists always.
From the radiographs we see that the zones containing unit features of the labeled
digit (the zones of a digit that do not overlap with other digits), which represent the
main part of linearly separable feature of a digit, is positively highlighted, while the
zones containing unit features of other non-labeled digits are negatively explored. The
zones of the labeled digit that overlap with other digits are also positively displayed,
but to a lesser degree than that of the unit features. Following the characteristic map,
these kinds of patterns enhance the output of a sample on its label neuron, while
suppress the outputs of samples of other classes on this output neuron. These patterns
thus indicate how linearly separable feature is carried out by the linear learning mode.

Fig. 6 The LNI for the NNNs with (a) d=30 and (b) d=75. The lines in each
panel represent Π(𝜇|𝑖2 ) with 𝜇 = 1, … ,10 for the specific output neuron 𝑖2 ;
the dark line represents Π(𝜇|𝑖2 ) with 𝜇 = 𝑖2 . The time length in (a) is 8 ×
106 and in (b) is 2 × 107 MC steps in each panel. .

In principle, overlapping among patterns of different classes is unavoidable, and

thus there should be linearly inseparable features in every sample. Note that the zones
of each digit correspond to the face zones in the toy sample sets. The fifth row of Fig.
5 shows the radiographs of the NNN obtained by learning the ten digits with d=75.
Comparing to the case of d=30, we see that they are quite similar. In other words, it is
difficult to verify whether the nonlinear linear mode is initiated by studying the
radiographs alone. The reason is that even if there are linearly inseparable features, they
will be fragmented and difficult to identify. This is different from the case of the second
sample in the third toy sample set, where the linearly inseparable feature zone is large
and distinctive. However, we can still see an increasing degree of fragmentation in
radiographs of d=75 comparing to the case of d=30, which may be a hint of the
appearance of the nonlinear learning mode.
Solid evidences of the existence of the nonlinear learning mode can, however, be
provided by the LNI. In Fig. 6, we show the evolution of the LNI of the NNN trained
by the 10 handwritten digits as a function of time for d=30 and d=75. We see that in the
case of d=30, the learning process is dominated by the linear learning mode, since the
LNI satisfies the condition that Π(𝜇|𝑖2 ) is the largest for 𝜇 = 𝑖2 . In the case of d =75,
however, the LNI of five digits no longer meet this condition, indicating that the
nonlinear learning mode has been initiated. This result confirms that using the cost
function with a large d is an effective strategy to excite the nonlinear learning mode.
The benefit of the extraction of linearly inseparable feature can be revealed by the
improvement of the accuracy of recognition. To demonstrate this effect, we train a LNN
and a NNN using the first 600 samples from MNIST. Optimal parameters d and β are
searched and used to generate Fig. 7(a) which shows the evolution of the accuracy rate
on the supplied test set of MNIST. We can see that the test accuracy of the nonlinear
network is noticeably higher than that of the linear network, although it is found that
the training accuracy for both the LNN and the NNN reach 100%. These facts indicate
that each training sample contains sufficient linearly separable features for successful
classification, and that the training samples also contain linearly inseparable features
that only the NNN is able to learn and such learnings help it to achieve a noticeable
higher accuracy on the test set.
The 9th to 11th rows of Fig. 5 show radiographs of the ten output neurons for the
LNN, the NNN with d=30，and the NNN with d=70, respectively. We see that patterns

of digits emerge in the radiographs. The positively highlighted parts of a digit represent
the zones of common linearly separable feature of a class of samples. Again, these
radiographs show less distinctive feature, and we see no significant changes from the
9th row to the 11th row. As revealed by Fig.87(a), the extra improvement of the
accuracy by the NNN is about 8%. Therefore, in the case of the MINSIT data set, the
linear separable feature is the dominant feature, so radiographs of MNIST are
dominated by linearly separable features, and thus no obvious difference can be
discerned between the radiographs with or without the nonlinear learning mode.
How to maximally extract the linearly separable and inseparable features should
be an essential problem for obtaining the optimal neural network. Fig. 7(b) shows the
accuracy of the three-layer nonlinear network for several values of β as a function of d.
Here, the accuracy at a given value of d is similar to final accuracy in Fig 7(a), that is,
with each value of d, we train the neural network to find the maximum value of the
accuracy appearing in the training process. We see that, for a given β, increasing d
increases the accuracy initially till it reaches the maximum value, and further increasing
d actually decreases the accuracy, which is a sign of overlearning. Similarly, at a fixed
d, increasing β increases the accuracy initially, but further increasing causes
overlearning and the accuracy to decrease. It seems that the understanding and
controlling of overlearning is the key to the optimal solution.

Fig. 7 Characterizing the overlearning by the distribution of outputs of hidden-layer

neurons. (a) The accuracy as a function of training time for the linear (circles) and
nonlinear (triangles) neural networks. (b) The maximum accuracy as a function of d
for bata=0.05(triangles), 0.15(circles) and 0.3(stars). (c) The distribution of outputs of
hidden-layer neurons for d=40(solid line),70(dash line) and 100(dot line) at bata=0.15.
(d) The distribution of outputs of hidden-layer neurons for bata=0.05(solid
line),0.15(dash line) and 0.3 (dot line) at d=70.

By study of the distribution of outputs from hidden-layer neurons, we find that

the overlearning induced by excessive d or β is all due to the presence of excessive
number of extremely nonlinear neurons, i.e. neurons with output in the extreme
nonlinearity region. In Fig.7(c) and 7(d) we show the distribution of outputs of hidden-
layer neurons at d=40,70,100 with β=0.15, and at β=0.05,0.15,0.3 with d=70,
respectively. In each plot, the distribution evolves from the state with less extremely
nonlinear neurons to the state with more extreme nonlinear neurons. Note that the
extremely nonlinear neurons are characterized by the peaks around x=±1, and linear
neurons are characterized by the distribution around x=0. We see that the maximum
accuracy appears with a proper balance between the extremely nonlinear neurons and
linear neurons. Too much or too little nonlinear neurons would both decrease the test
accuracy.
Extremely nonlinear neurons are indispensable in the extracting of linearly
inseparable features by the nonlinear learning mode. As shown in Fig.4, for the goal of
converting the output of the second sample to have the largest value requires those of
the first and third samples to cancel each other, i.e., 𝑓(ℎ) ≈ ±1 with these two
samples. However, extremely nonlinear neurons do have their drawbacks; they decrease
the robustness of the neural network since they behave like the step neurons, 𝑓(ℎ) =
𝑠𝑖𝑔𝑛(ℎ), and too many extremely nonlinear neurons would lead to the out-of-balance
between the extraction of linearly separable and inseparable features. In more detail,
for a given training set, the proportion of linearly separable and inseparable features of
samples is fixed. With a fixed number of hidden-layer neurons, there are two possible
scenarios that may lead to a less effective neural network. One is that the total number
of neurons is insufficient for extracting the complete information. Another is that the
ratio of neurons with linear and nonlinear learning modes does not match that of the
linearly separable and inseparable features.
With this understanding we can explain the results of Fig. 7(b). Because there is a
plenty of linearly separable features, the neural network tends to invoke the linear
learning mode to extract linearly separable features. If the training goal is achieved by
the linearly separable features alone, the training is stopped and the test accuracy would
be low because of losing the information contained in linearly inseparable features.
With the increase of β, d and the training time, nonlinear neurons with extreme
nonlinearity increase, leading to more nonlinear linear mode triggered and linearly
inseparable feature extracted, and thus the accuracy is improved. However, excessive
extremely nonlinear neurons would lead to the out-of-balance between the linear and
nonlinear learning modes, and results would be overlearning.

3.2 Increasing the width or depth of neural networks to balance the linearly

separable and inseparable features.

Then the remaining questions are what is the optimal distribution and what factors
determine the optimal distribution. We will show that increasing the width and the depth
of neural networks provides two effective ways to obtain the optimal distribution.
To show the width effect, we study a 784- N1-10 neural network trained by the first
600 MINIST samples with β=0.15. Here, the value of β is the optimal value from the
last section with N1=600, and it is checked this value remains optimal for N1s used in
this section. In Fig 8(a), the circles and triangles show the accuracy as a function of d
with N1=1200 and N1=1800, where each data point is the highest accuracy achieved at
each value of d. Fig. 8 (b) shows the optimal accuracy as a function of the width where
an optimal d is searched and used in its calculation. We see that the accuracy increases
with the increase of width and gradually tends to saturation around N1=1500.
In Figure 8(c), we show the distribution of outputs of hidden-layer neurons for
N1=1200 and 1800. Fig. 8(d) shows the height of peaks ℎ̅ (averaged with the values
of x=±1) as a function of the width. It can be seen that ℎ̅ decreases with N1.
 Fig. 8. The width and depth caused improvement of the network performance. (a)
The accuracy as a function of d for N1=1200 and N1=1800. (b) The maximum
accuracy as a function of N1. (c) The distribution of outputs of hidden-layer neurons
for N1=1200 and N1=1800 at the maximum accuracies. (d) The height of the picks
of the distribution as a function of N1 with the log-log scale. (e) The accuracy as a
function of d for the four- and five- layer neural networks. (f) The distribution of
outputs of first hidden-layer neurons for the four- and five-layer neural networks.

The reason that bigger widths improve test accuracies can be understood on the
basis of balancing the extraction of linearly separable and inseparable features. The
amounts of linearly separable and inseparable features in the training set are fixed. With
a small-size neural network, it does not have sufficient neurons to extract completely
both linear and nonlinear features. As the linearly separable features are dominant, the
network is inclined to use more neurons to extract this kind of information, leading to
an insufficient number of nonlinear neurons and a suboptimal small ratio of extremely
nonlinear neurons. With the increase of the width, there are more neurons available and
thus some of them can be spared to extract the linearly inseparable features, and thus
the ratio of extremely nonlinear neurons can increase to an optimal value. However, we
can expect that the total number of the nonlinear neurons used for extracting the linearly
inseparable feature should to stay approximately constant for even larger width; we
denoted this number as 𝑁𝑛𝑜𝑛. The reason is that, as shown in Fig.4, the nonlinear
learning mode requires the combination of different neurons and thus occupies more
resources of the neural network. Therefore, further increasing of the width would not
increase the number of neurons that extract linearly inseparable features. As a result,
one can expect that the ratio of extremely nonlinear neurons, characterized by the
amplitude of peaks around x=±1 to decrease as 𝑁𝑛𝑜𝑛. /(𝑁1 − 𝑁𝑛𝑜𝑛. ) or scale as 1/𝑁1
if 𝑁1 is large enough; the extra neurons should concentrate towards the region of x=0.
These predictions are confirmed by Fig. 8(c) and 8(d). In these situations, neurons for
extracting both linearly separable and inseparable features are approximately saturated;
therefore, continuing increase of the width does not increase the test accuracy
significantly as can be seen in Fig. 8(b). In summary, with a sufficient width, the neural
network has the possibility to extract both linearly separable and inseparable features
while keeping the balance between the linear and nonlinear learning modes.
To show the depth effect, we study a four-layer 784-600-600-10 neural network
and a five-layer 784-600-600-600-10 neural network. Fig. 8 (e) shows the accuracy of
these two neural networks as a function of d, and Fig. 8(f) shows the output distribution
of neurons in the first hidden layer. Together with the result of the three-layer neural
networks shown in Fig.7(b), we see that the test accuracy increases with the depth of
the neural network, while the number of extremely nonlinear neurons decreases with it.
Comparing Fig. 8(e) and Fig. 8(a), we see an important advantage of a deeper-
layer neural network over a wider-layer one for practical applications. As shown in Fig.
8(a), the optimal value of d increases with the increasing width quite significantly. To
obtain the optimal test accuracy of a wider neural network, one would have to search
in a wide range for the optimal d. In contrast, as shown in Fig. 8(e), the optimal value
of d does not change with the increasing depth of the neural network. In fact, there is a
wide range of d that would give approximately optimal test accuracy for five-layer
neural network (triangles). It is difficult for us to apply our MC algorithm to training
neural networks with more layers, but from the trend exhibited by the three to five-layer
neural networks, we can speculate that, with the further increase of the depth, the
optimum solution region should appear as a perfect plateau over a large range of d
values. As a result, it would be much easier to obtain the optimal solution by sparsely
sampling the control parameter space. Furthermore, once an optimal control parameter
is found, one would not have to search it again for networks with deeper depths. This
property is very favorable for neural networks using the softmax cost function. We
suspect that this property may be at least one of the superiorities of the deep-layer neural
networks.
The mechanism that the accuracy increases and the number of extremely nonlinear
neurons decreases with the increasing the depth is similar to that of increasing the width,
since increasing layers also increases the total number of neurons in the network. In the
last row of Fig. 5, we show the radiographs of the ten output neurons of the four-layer
neural network, which indicate that the “holographic” structure of the network is
remained. From the point of view of weight pathways, a deep-layer neural network has
much more weight pathways comparing to a three-layer neural network with the same
number of hidden-layer neurons. For example, a 784-1200-10 network has
784x1200x10 ~ 9x106 weight pathways, while a 784-600-600-10 network has
784x600x600x10 ~ 3x109 weight pathways. Therefore, deep-layer networks have much
more freedom to establish subnetworks, and should provide more freedom to construct
learning models. This should be the reason that a deep-layer neural network can have
the nonlinear linear modes with a relatively small d. The details of this mechanism is
still to be investigated in the future.

5 Degree of attention of learning process.

Fig.9 Visualization of the attention degree for the NNN with d=30. The first,
second and third lines are for the first, second and third training sample sets, and
the first, second and third columns are for the first, the second and the third
samples. .

As an augmentation to the WPA, we introduce the attention degree to monitor

activities of weight mutations in this section. This concept is benefited by the MC
training algorithm we adopted. The attention degree can reveal and study the
contributions or weights of different features. It is defined as follows: for the mutation
(1)
of weight wi1 i0 , if it maximizes the increment of local fields of all output neurons in
the direction of iL -class, then the mutation is defined as a successful attention of feature
bitmap i0 on class- iL , and then let P(i0 |iL ) → P(i0 |iL ) + 1 . During or after the
training, we plot the heatmap
latt
αβ (iL ) = P(Mα + β|iL ),

to visualize the degree of attention. Fig. 9 shows the heatmaps of the degree of attention
of the NNN30 for the three sample sets . The information involved in other plots can
be easily interpreted.
The visualized degree of attention can explore many details of learning and
learning process. Particularly, we find that there exist two modes of attention, namely,
the face attention mode and the ground attention one. In Fig 9, we see that the degree
of attention for different region is different. For the first sample of the first sample set,
the background has the largest degree of attention, while the face regions of the first
sample has a attention that is larger than that of the third but smaller than that of the
second. Therefore, to establish the subnetwork for the first sample to the first output
neuron, the attention is indeed focused on the background. We call this mode the
background attention mode. For the third sample, on the contrary, we see that the face
zone of the third sample has the largest attention, while the other regions have less
degree of attention. We call this mode the face attention mode. It can be find that the
training performs the background attention for the second sample of the first set, the
first and the second samples of the second set, and the first sample of the third set, while
performs the face attention model for the others. Indeed, it can be realized that the
degree attention has a complex structure, particularly for samples with complex features.
For example, for the second sample of the third set, an extremely high degree of
attention is focused on the cross area of the face zones of the first and the second
samples.

Fig. 10. The visualization of the degree of attention. First to third lines: the
visualization images of the attention degree for the NNN with d=30 trained by the first
three, the first five, and all of the ten digit samples shown in the first line of Fig. (7).
The last line shows the results of the NNN trained by 600 MINIST samples.

Usually, the identifiable information lies in regions of both the target and the
background. In the limit when it has 𝑥𝑖0 (𝜇) → 0 in the face zone, the neural network

have to recognize the samples completely through the geometry of the background,
while in the opposite case of 𝑥𝑖0 (𝜇) → 0 in the ground region, the geometry of the

face becomes the only recognizable feature of the sample, and the recognition should
be performed by the face attention mode. Generally, these two modes of attention are
coexisted, which is similar to the way that the brain recognizes objects.
As can be seen, for example, from the heat maps of a specific sample that the
value of the degree of attention of the face zone of other samples is not zero, which
means that to identify a sample needs also the information of others. This proves the
fact revealed by the weight-path analysis that the subnetworks of each sample stores
also the information of others, i.e., holographic feature of the neural network.
The degree of attention is also helpful for analysis the learning process for the
MINIST data set. In the first to third lines of Fig. 10 we show the visualized degree of
attention for the first three, first five and all of the ten digit samples shown in the first
line of Fig. 7. In the last line of Fig. 10 we show the results for the first 600 samples
respect to the ten output neurons. We can see that there exist also background attention
mode (the second plot in the second line) and the face attention mode (all others). In
addition, the unitary features of a sample, as well as those features having high degree
of overlap with other samples, are usually payed more attentions.

6 SUMMARY AND DISCUSSION

Weight pathways connect inputs to outputs of a neural network through

subnetworks that are characterized by penetration coefficients; any internal change of
the network can be qualitatively displayed by the changes of such penetration
coefficients. Therefore, the WPA approach provides an effective means to detect the
internal structure of the network. In more details, with neurons as nodes, the network
can be decomposed into a series of subnetworks, each of which is characterized by a
characteristic map with a set of penetration coefficients. The training process can be
interpreted as the organization of weight pathways to produce the minimum cost
function, creating coherence structures for the penetration coefficients that represents
the enhancement or suppression on the corresponding component of the input vector.
With the visualization of the penetration coefficients, we gain a penetrating view of the
inside of the "black box" of a neural network, making its learning and recognition
mechanisms interpretable. One of the important findings through the visualizations is
that each subnetwork uses a "holographic" structure to encode all the training samples
instead of just one class of samples. The "holographic" structure reveals how
recognition is performed, that is, when an input vector is presented, every subnetwork
reacts to the whole information in it with its "holographic" structure and decides
collectively what the input vector is or is not. The "holographic" structure is also the
basis of the generalization capability of the neural network. Our findings support the
latest neurobiological understanding of the biological neural networks in the brain
[22,23], and confirms that information is stored in the networks of weight pathways.
The WPA approach reveals the self-organizing classification of the hidden-layer
neurons. After the self-organization, hidden-layer neurons can be generally divided into
dominant-mode neurons and auxiliary neurons. The former produce the largest output
in the label neuron of an input vector, while the latter adjust the local fields of the other
output neurons to approach their goal of training values. The analysis of the
classification of the hidden-layer neurons leads to the finding of the linear and the
nonlinear learning modes. The former extracts linearly separable features and can be
performed by single neurons independently. The latter can extract linearly inseparable
feature and requires the cooperation of neurons and the help of the nonlinearity of the
neuron transfer function.
The WPA approach not only reveals what to learn and how to learn, but also
reveals how to learn better. What to learn and how to learn? Extract the linearly
separable features with linear learning mode, and extract linearly inseparable features
with nonlinear learning mode. How to learn better? Maximize the extraction of linearly
separable and inseparable features and avoid the overlearning. Our study shows that a
neural network tends to first extract linearly separable features with the linear learning
mode. If each sample contain sufficient linearly separable features, the network may
reach the training goal relying only on the linear learning mode, thus wasting linearly
inseparable feature.
In order to extract linearly inseparable feature, the network needs invoke the
nonlinear learning mode with a cost function that is difficult to minimize with only the
linearly separable features. To maximum utilize the information of the samples, the
neural network must extract both features completely and with the right balance. We
demonstrate that increasing the width and depth of the network are effective strategies
for this purpose. When the width of a network is large enough, there are enough neurons
to support both the linear and nonlinear learning modes, and thus one has a good chance
to extract both features completely. Increasing the depth of the network can also
increase the probability of nonlinear learning mode, and improve the network
performance. More importantly, we find that increasing the depth of a network can get
the optimum performance relatively easier than increasing the width. In more details,
the optimum control parameters stay approximately constant for different depths, hence
the careful search for the control parameters is not necessary. This property is very
beneficial for applications. We have not studied in detail neural networks with deeper
layers, other types of neuron transfer function, other updating algorithms, or other types
of networks. We hope such studies can be shown in future.

ACKNOWLEDGEMENTS

This work is supported by the NSFC (Grants No. 11975189, No. 11975190)

Author contributions:

H.Z. proposed the idea of WPA; F.S. performed all the simulations and proposed the
idea of the degree of attention of learning; all authors contributed significantly to
analysis with constructive discussions and manuscript preparation; H.Z., F.M. and
F.S. wrote the manuscript.

Reference

[1] L. Yann，Y. Bengio，G. Hinton, Deep Learning, Nature, 521, 436-444 (2015).
[2] T. Poggio, A. Banburski, and Q. Liao, Theoretical issues in deep networks, PNAS,
117(48), 30039(2020).
[3] C. Zhang, S. Bengio, M. Hardt, B. Recht, and O. Vinyals, , Understanding deep
learning (still) requires rethinking generalization. Communications of the ACM, 64(3),
107(2021).
[4] G. Alain, Y. Bengio. Understanding intermediate layers using linear classifier probes,
arXiv:1610.01644, 2016.
[5] M. Gabri´e, A. Manoel, C. Luneau, J. Barbier, N. Macris, F. Krzakala, and L.
Zdeborova´, Journal of Statistical Mechanics: Theory and Experiment 2019, 124014
(2019).
[6] R. Shwartz-Ziv, N. Tishby, Opening the black box of deep neural networks via
information, arXiv:1703.00810, 2017.
[7] S. Chung, D. D. Lee, and H. Sompolinsky, Physical Review X 8, 031003 (2018)
[8] F. Gerace, B. Loureiro, F. Krzakala, et al. Generalisation error in learning with
random features and the hidden manifold model, arXiv:2002. 09339, 2020.
[9] Y. Yoshida, R. Karakida, M. Okada, and S.-I. Amari, Journal of Physics A:
Mathematical and Theoretical 52, 184002 (2019).
[10] T. Hou and H. Huang, Statistical physics of unsupervised learning with prior
knowledge in neural networks, Physical Review Letter 124, 248302(2020).
[11] S. Mei, A. Montanari, and P.-M. Nguyen, PNAS 115, E7665 (2018).
[12] D. Bau, J.-Y. Zhu, H. Strobelt, A. Lapedriza, B. Zhou, and A. Torralba,
“Understanding the role of individual units in a deep neural network,” PNAS, 117(48),
30071 (2020).
[13] W. Samek, G. montavon, S. Lapuschkin, CJ. Anders, and K. Mueller, Explaining
deep neural networks and beyond: A review of methods and applications,
PROCEEDINGS OF THE IEEE 247 109(3), 2021.
[14] AM. Saxe, JL. McClelland, S. Ganguli, Exact solutions to the nonlinear dynamics
of learning in deep linear neural networks, arXiv:1312.6120, 2013.
[15] S. Arora, N. Cohen, W. Hu, Y. Luo, Implicit regularization in deep matrix
factorization, Advances in Neural Information Processing Systems 32, 7413(2017).
[16] C. Olah, A. Satyanarayan, I. Johnson, S. Carter, L. Schubert, K. Ye, and A.
Mordvintsev, The building blocks of interpretability, Distill 3, no. 3 (2018): e10.
[17] D. Bau, B. Zhou, A. Khosla, A. Oliva, and A. Torralba, Network dissection:
Quantifying interpretability of deep visual representations, In Proceedings of the IEEE
conference on computer vision and pattern recognition, pp. 6541-6549 (2017).
[18] L. Phillips, G. Goh, and N. Hodas, Explanatory Masks for Neural Network
Interpretability, arXiv:1911.06876,2019.
[19] D. Kalimeris, G. Kaplun, P. Nakkiran, B. Edelman, T. Yang, B. Barak, H. Zhang.
"Sgd on neural networks learns functions of increasing complexity." Advances in
Neural Information Processing Systems 32, 3496(2019).
[20] W. Hu, L. Xiao, B. Adlam, and J. Pennington, The surprising simplicity of the
early-time learning dynamics of neural networks, arXiv:2006.14599, 2020.
[20] H. Zhao, A general theory for training learning machine, arXiv:1704.06885 (2017).
[21] H. Zhao, Inferring the dynamics of “black-box” systems using a learning machine,
Science China Physics, Mechanics & Astronomy 64, 270511(2021).
[22] L. F. Barrett, Seven and a half lessons about the brain, Houghton Mifflin Harcourt
Publishing, NY, 2020
[23] J. Cepelewicz, The Brain Doesn’t Think the Way You Think It Does, Quanta
Magazine, 24 Aug 2021 (https://www.quantamagazine.org/mental-phenomena-dont-
map-into-the-brain-as-expected-20210824/)