Does artificial neural network support connectivism’s assumptions?

International Journal of Instructional Technology and Distance Learning
March 2017 Vol. 14 No.3.- 3 -
Editor’ Note: This paper seeks to understand and validate connectivism theory. The paper questions
connectivism's principles, compares it with other learning theories, and validates it in relationship to Artificial
Intelligence and Artificial Neural Networks.
Does Artificial Neural Network support
Connectivism’s assumptions?
Alaa A. AlDahdouh
Portugal
Abstract
Connectivism was presented as a learning theory for the digital age and connectivists claim that
recent developments in Artificial Intelligence (AI) and, more specifically, Artificial Neural
Network (ANN) support their assumptions of knowledge connectivity. Yet, very little has been
done to investigate this brave allegation. Does the advancement in artificial neural network
studies support connectivism’s assumptions? And if yes, to what extent? This paper addresses the
aforementioned question by tackling the core concepts of ANN and matching them with
connectivist's assumptions. The study employed the qualitative content analysis approach where
the researcher started with purposely selected and relatively small content samples in
connectivism and ANN literature. The results revealed that ANN partially supports
connectivism’s assumptions but this does not mean that other learning theories such as
behaviorism and constructivism are not supported as well. The findings enlighten our
understanding of connectivism and where it may be applied.
Keywords: learning theory; connectivism; constructivism; behaviorism; artificial neural network; ANN;
neural network; artificial intelligence; AI; machine learning; e-learning; online learning; distance learning
Introduction
In 2005, George Siemens started his proposed learning theory, connectivism, by asserting the
huge impact of technology on our learning activities. In his words, "technology has reorganized
how we live, how we communicate, and how we learn" (Siemens, 2005, p. 1). It thus follows that
learning theory should reflect these changes. Based on this assumption, he criticized former
learning theories such as behaviorism and constructivism; and advocated new theoretical
framework. The suggested framework, of course, incorporates technology in its principles.
Specifically, one of connectivism’s principles states that "Learning may reside in non-human
appliances" (Siemens, 2005, p. 5). This principle alludes to the ability of technology to learn.
Before long, Stephen Downes embraced the theory and integrated it to the idea of connective
knowledge where knowledge is distributed and it does not exist in a specific and single place
(Downes 2005, 2006). Downes (2012) concentrated on network structure of the internet and how
it may empower online students to do things that were hardly ever possible before, such as
distance collaboration and information searching. More recently, Downes shared a sequence of
posts on artificial intelligence and neural network findings in indication of their relevance to
connectivism.
Aldahdouh, Osório and Caires (2015) explained the idea of networked knowledge thoroughly and
made it clearer in relation to AI. They claimed that connectivism is based on network science
principles. Their step-by-step explanation of knowledge network (neural, conceptual and
external) has led them to argue that connectivism "has been drawn from a long history of
Artificial Intelligence findings" (p. 17).

March 2017 Vol. 14 No.3.- 4 -
On the other hand, these allegations have brought criticisms to connectivism. Verhagen (2006)
argued that machine learning, inductive learning, and fuzzy logic software have nothing to do
with human learning. "Modern cognitive tools are nothing but an extension of the toolkit"
(Verhagen, 2006, p. 4). In Verhagen’s perspective, artificial neural network does not differ from a
pocket calculator. Moreover, Bell (2011) repeatedly mentioned that connectivism’s principles
lack of rigor and are, in most part, untested.
Although these criticisms were presented right after proposing connectivism, very little has been
done to examine the relationship between AI findings and connectivism’s principles. The
question of whether connectivism was built on top of AI findings remained almost intact. In this
article, ANN was selected to represent machine learning (which is a branch of AI) for many
reasons. First, the core idea of ANN is inspired from human brain and, second, connectivists
frequently refer to ANN as if ANN has supported their claims.
This study reviewed literature of both connectivism and ANN and tried to match their principles.
The paper starts with brief description of connectivism. Then it moves to describe ANN concepts
in relation to connectivism’s assumptions. Subsections include artificial neuron, network
architectures and learning algorithm. The paper avoids presenting complex algorithms to improve
text clarity and readability. In addition, it avoids going into ANN details that will not serve the
purpose of this study; the reader should be aware of the extendibility of concepts presented here.
Connectivism
The reasons that make educators keen to develop new learning theory can be summarized in three
broad motives: (1) international growth in internet usage and the gap between learners’ and
school’s activities (Bell 2010; Brabazon 2016); (2) half-life of knowledge becomes shorter and
knowledge changes rapidly (Aldahdouh and Osório 2016; Aldahdouh et al. 2015; Downes 2006;
Siemens 2006); (3) and human-technology interaction where the interaction leads to changes in
both sides of the equation; as technology changes, human also changes (Dirckinck-Holmfield,
Jones, and Lindström 2009; Siemens 2005).
Connectivism started from these premises and integrated principles from different theories
including chaos, network, and self-organization theories. Siemens (2005) introduced
connectivism as an alternative theory which was not built on previous learning theories. In
general, connectivism’s principles are shifting focus from content itself to the connections of
content. Maintaining existing connections, making decision of which connections to add, and
creating new connections between different fields are essential part of learning according to
connectivism.
The core assumption of connectivism is that knowledge has a structure; and this structure is better
to be conceived as a network (Aldahdouh et al. 2015; Downes 2005, 2006; Siemens 2006). A
network is a group of nodes linked to each other by connections. A node can be in one of three
different levels: neural, conceptual, and external (Aldahdouh et al. 2015; Siemens and
Tittenberger 2009). A connection serves as a bridge that conveys information from one node to
another. The more connections a network has, the more robust it will be. Without those
connections, the whole network will fall apart. Thus, Siemens (2006) has concluded that "The
pipe is more important than the content within the pipe" (p. 32). This makes the content less
important in comparison to the connection. A more extreme view toward content sees it as
something from the past; 'Content is a print concept' (Cormier 2016). However, Aldahdouh et al.
(2015) adopted a moderate view toward content in that "The information needs a connection to
reach the target and the connection needs the flow of information to stay alive. Therefore, no flow
of information exists without connection and no connection remains without flow of information"
(p. 11).

March 2017 Vol. 14 No.3.- 5 -
The second core assumption of connectivism is that a single connection between two nodes does
not have meaning in its own. The meaning is distributed across group of connections called
pattern (Downes 2006). The pattern "refers to a set of connections appearing together as a single
whole" (Aldahdouh et al., 2015, p. 5). This pattern should be considered as the smallest unit that
has meaning in its own. And hence, the network can be seen as a group of patterns that interact
with each other to give the meaning of the entire network. Since connections may 'die' or 'live',
the patterns and the knowledge are conceived as dynamic objects where some nodes become
isolated and others become connected. The patterns change rapidly which made Aldahdouh et al.
(2015) see knowledge as a "jellied creature" (p. 15).
The aforementioned assumptions can lead us directly to the definition of learning in
connectivism. If knowledge has a network structure and meaning is distributed in dynamic
patterns, then learning should be defined as "a continuous process of network exploration and
patterns finding; it is a process of patterns' recognition" (Aldahdouh et al., 2015, p. 14).
This paper concentrates on one of connectivism’s principles which refers to the ability of
technology to learn. This principle has been criticized and its meaning remains unclear for some
researchers while others have questioned its validity (Bell 2011; Kop and Hill 2008; Verhagen
2006). The paper does not provide comprehensive review of connectivism literature. The reader
is recommended to see the work of Aldahdouh et al. (2015) for clearer explanation of the theory.
Nevertheless, the paper will return to connectivism’s principles as it explains the concepts of
ANN.
Artificial Neural Network
Artificial Intelligence (AI) refers to the art of creating machines that are able to think and act like
humans; or think and act reasonably (Russell and Norvig 2010). In order to build an agent that
can think and act as so, the agent must be able to learn new things. To learn means that the agent
should improve its performance on future tasks taking its past experience into account (Russell
and Norvig 2010). Making an agent able to learn is an area of study called Machine Learning
(ML).
Artificial Neural Network or ANN is a software structure developed and based on concepts
inspired by biological functions of brain; it aims at creating machines able to learn like human
(Goodfellow, Bengio, and Courville 2016; Nielsen 2015; Russell and Norvig 2010). Thus, ANN
is part of ML. Interestingly, ANN has many other names in AI field including parallel distributed
processing, neural computation and connectionism (Russell and Norvig 2010). Most ANN types
are supervised learning network. That is, both an input and the correct output should be given to a
network where the network should learn a function that maps inputs to outputs. There are some
types of ANN such as Deep Belief Network (DBN) which can do unsupervised and semi-
supervised learning (Nielsen 2015). However, research is still conducting on DBN to improve its
performance. This article concentrates on supervised learning networks which showed a very
good performance in wide variety of tasks.
Before proceeding into details, it is important to know that ANN is a vivid research area. Recent
years, namely from 2011 to 2015, have witnessed a sequence of records breaking in the field of
ML driven by ANN (Nielsen 2015). Even more, Goodfellow et al. (2016) indicated that ANN
evolves rapidly so that new best architecture "is announced every few weeks to months" (p. 331).
This makes writing this article a very challenging task.
Artificial Neuron
Since a structure of ANN has been inspired by biological brain, ANN should consist of a
collection of neurons. AI researchers designed artificial neurons called perceptron and sigmoid
which are believed to have similar function to a biological neuron (Goodfellow et al. 2016;

March 2017 Vol. 14 No.3.- 6 -
Nielsen 2015). Artificial neuron is hereafter referred to as neuron for short. A neuron is a node
that receives input from preceding neurons and makes a decision to 'fire' to the next neurons. To
make that decision, it should first evaluate each input according to its own perspective and then
sum all inputs up to get a single and holistic view. Finally, a neuron presents the holistic view to
its internal judgment system to make a decision to fire or not.
Fig. 1 Perceptron neuron
This system seems trivial but it turns out to be a complicated decision-making model. For
example, suppose that you are a neuron and you want to make a decision to buy car. You
probably make that decision based on many variables which may include gas price ($200), car
insurance ($150), and parking cost ($100). In your perspective, car insurance and gas price are
more important and more likely to increase in near future than parking cost. In this case, you
weigh up car insurance (1.5) and gas price (1.5) while downplay parking cost (0.5). Then you
sum that up to get the holistic perspective (100*0.5 + 1.5*150 + 1.5*200). Therefore, according
to your own perspective, a car would cost you $575 per month. Then you present this holistic
perspective to your internal judgment system which may have been previously set on a specific
threshold ($480). Therefore, you make a decision not to buy a car because it exceeds the
threshold ($575 > $480). Your own perspectives of inputs, the internal judgment system, and the
threshold are called weights, activation function and bias respectively. By changing weights and
bias you reach a completely different decision. For example, set gas weight to 1 instead of 1.5 and
notice the difference. Searching for weights and bias that generate the desired output is the job of
learning algorithm.
In the previous example, we imagined a neuron as a person. This sounds familiar in connectivism
literature. Connectivists often argue that networks exist everywhere and these networks are
similar in some way or another (Downes 2016). Interestingly, researchers in ANN share the same
assumption with connectivists (Nielsen 2015). They sometimes conceive a single neuron as a
person and in some other times conceive the whole network as a single human brain. Zooming in
and out help them understand and modify both the neuron and the network in very similar way.
We will see in next how this assumption works well in both levels. Another thing that matches
connectivism well is the bias of a neuron. Each neuron has its own bias and, therefore, ANN
contains a variety of neurons in which each neuron works in completely different way. In
connectivism, Downes (2010) identifies four principles for applying democracy in education:
autonomy, diversity, openness and interactivity. Siemens (2005, 2006) states that knowledge and
learning rests in diversity of options and Aldahdouh et al. (2015) argues that educational systems
should foster the learners' diversity, not their similarity.
Previously given example seemed as though a neuron works well in simulating human decision-
making system. However, with little thinking, one can figure out that it does not. Suppose for
instance that you kept all variable values as they were in the previous example except your
perspective of (weight of) gas price. If you set the weight to be (1), then your holistic perspective
becomes ($475). This is below your bias value ($480), thus you decide to buy a car. Now, try to

March 2017 Vol. 14 No.3.- 7 -
set your perspective of gas to be (1.05). Your holistic perspective becomes ($485). That is greater
than your bias value ($480), thus you decide not to buy a car. This is really a naive system. Our
internal judgment systems do not do that. In real world, 5 dollars below or above a predefined
threshold may not make that difference. This is called perceptron neuron which has a hard
activation function (Russell and Norvig 2010). It would be better if a neuron has a soft activation
function. Soft so that it goes gradually from (Yes, I will absolutely buy a car) to (No, I will
absolutely not buy a car). Note that we have just concentrated on the meaning of the output itself
but it is even not logical that a very tiny shift in a single weight (from 1 to 1.05) makes that big
difference in the output. If ANN is going to learn the right weights and biases, a small change in
one weight or bias should produce small change in network output (Nielsen 2015). In order to be
able to learn, ANN should move slowly and smoothly from one decision to another. For these
reasons, ANN researchers have examined many alternative soft activation functions such as
sigmoid, tanh and rectified linear neuron.
Fig. 2 Sigmoid neuron
In comparison, connectivism has been criticized for its oversimplification of interaction between
nodes as the connection can be either active or inactive (Clarà and Barberà 2014). However,
connectivism proponents (Aldahdouh et al. 2015) have shown that a connection is graded and not
necessarily sharp. The graded view of a connection is congruent with sigmoid neuron function.
Other issue presented here is the speed of learning. Connectivism puts the bulk of its attention on
the rapid change of knowledge but it does not describe how exactly a learning process is in this
dynamic environment. However, there are signs in connectivism literature that a learning process
should cope with this rapid change. For example, one of the connectivism’s principles states that:
Decision-making is itself a learning process. Choosing what to learn and the meaning of
incoming information is seen through the lens of a shifting reality. While there is a right
answer now, it may be wrong tomorrow due to alterations in the information climate
affecting the decision (Siemens, 2005, p. 5).
Aldahdouh et al. (2015) emphasized the same concept and criticized the education system as they
said, "Sciences are developing very rapidly and the (reluctant) derivers' decisions are coming too
late" (p. 13).
Thus, the output of soft neuron goes smoothly from 0 to 1. The output is now a real number. Even
though this soft neuron seems to work better than hard neuron, it still has its own problems. Take
for example a sigmoid neuron.

March 2017 Vol. 14 No.3.- 8 -
Fig. 3 Sigmoid activation function
Suppose that your initial thought was to buy a car only if it costs you less than $600 per month.
That is, your bias was set to $600. Suppose further that you were wrong and you should learn that
$475 is the right threshold using a sigmoid function as shown in Fig. 3. Since you learn according
to a sigmoid function, you should go slowly from $600 down a curve to $475. Note that as your
bias changes from $600 to $550, your decision remains nearly constant; it is almost 1. The same
has to be said if you should go from $350 to $400. In these periods, the output of the neuron is
saturated (Nielsen 2015). It seems as though a neuron does not learn anything; bias changes but
that does not produce a change in a decision. This is one of the problems of sigmoid neuron. To
solve this issue, researchers make learning speed independent on activation function (Nielsen
2015). Sigmoid neuron is still one of the most used neuron types in ANN. Other common soft
neuron types show similar and, in some cases, better performance than sigmoid (Goodfellow et
al. 2016; Nielsen 2015). Different soft neuron types and how they reflect to educational context
deserve a separate work.
An obvious and legitimate question to ask is whether these functions reflect human internal
judgment system. For example, suppose a learner in an online learning setting has internal
judgment system like a sigmoid. In this case, when a student decides to share and comment all the
time or decides to be silent all the time, could that be a sign that he is at saturation level? Could
that be a sign that he is not learning new things? Wang, Chen and Anderson (2014) indicated that
"the interaction of connectivist learning is divided into four levels: operation interaction,
wayfinding interaction, sensemaking interaction, and innovation interaction" (p. 121). These
levels grade from concrete to abstract and from surface to deep cognitive engagement. Pattern
recognition, decision making, aggregation and sharing appear on sensemaking level. This means
that sharing activity resides in the same level as learning process in connectivism. Downes (2009)
asserted the same concept and suggested four steps for network creation: aggregation, remix,
repurpose and feed forward. Thus, educating students to share information frequently is one of
the connectivism aims and it is part of student's learning process. This is not congruent with at
least a flat part of sigmoid function.
Nielsen (2015) has shown that a neuron can also implement any logic gate (e.g. NOT, AND, OR
and NAND). This means a single neuron has some sort of logic in its own and a network of

March 2017 Vol. 14 No.3.- 9 -
neurons is actually a network of logic gates. For example, if you want to design a network to
compute a sum of two bits, you may need a network of (n) neurons. And if you want to multiply
two bits, you may need a network of (m) neurons, and so on.
From this, we have a hunch that meaning exists in pattern of connections which is one of the
main connectivism’s assumptions. But for now, a thing we are sure about is that one node in a
network has trivial meaning in comparison to the meaning of a group of nodes. Actually, we are
usually not interested in a single node meaning (AND, OR, NAND); we are interested in the
meaning of the group of nodes as a whole (summation, multiplication). In connectivism, this
matches the concept of emergent property (Downes 2016) where "a compounded node is larger
than the sum of its inner nodes" (Aldahdouh et al., 2015, p. 12).
Artificial neural network architectures
The previous section describes a single neuron function. It is time now to see how researchers
arrange group of neurons to form a learnable network. In this article, the way of arranging
neurons in certain order is called network architecture. Recall that ANN may refer to two levels
of abstraction: (1) ANN as a person’s brain and (2) ANN as a group of learners. Thus, network
architecture refers first to a learner's inner abilities and mental capacities and; second, refers to a
way in which designers of learning-environment arrange a network of learners.
It is worth noting that ANN is a universal modeling system. Universality means that ANN can
learn any given function no matter what neuron type is used. It has been proved that with few
neurons and by changing biases and weights only, ANN can compute any zigzag-shaped function
(Nielsen 2015). The question now is how we arrange neurons in ANN to make it easier for a
learning algorithm to find those biases and weights. For clarity and simplicity, the paper divides
the most common ANN architectures based on three criteria: (1) number of layers, (2) flow of
information and (3) neuron connectivity.
Number of layers:
By looking on how many layers a network has, ANN can be divided into (1) shallow and (2) deep
networks.
Fig. 4 Shallow neural network
A shallow neural network consists of three layers ordered from left to right: (1) input, (2) hidden
and (3) output layer. The input layer does not really consist of neurons. Actually, it carries the

March 2017 Vol. 14 No.3.- 10 -
input values to the network. For example, a value that passes from X1,2 to next neurons is 25,
which is the input value. The second layer is named 'hidden' because it resides in the middle and
does not appear in either the input or the output of the network. Other than that, it is a normal
neural layer which contains normal neurons (Nielsen 2015). The output layer also contains
normal neurons and its output represents the output of the network.
Fig. 5 Deep neural network
A deep neural network is the same as shallow neural network but it has two or more hidden
layers. This architecture is also called Multilayer Perceptron (MLP). The original thought of
presenting deep network stems from the idea of complex problem defragmentation. ANN
researchers first noted that people are usually splitting the problem into sub-problems, solving
each sub-problem alone and then reconstructing them to solve the entire problem (Nielsen 2015).
They inferred that if the first hidden layer is going to handle the first level of the problem, then
there should be second, third and more hidden layers to handle next levels of the problem. The
initial steps of training deep network were frustrating because the network took long time to train
and didn't show a big difference from shallow network results.
In general, the terms shallow and deep are somehow misleading because they are not in line with
educational terminology of surface and deep learning (Vermunt and Vermetten 2004). Actually,
there is nothing special in deep neural network except it gives more accurate results, if it was
trained well. Moreover, the concept of layers is completely incompatible with connectivism’s
assumptions. The idea of that a network consists of a sequence of layers contradicts with chaos
theory which is one of the underpinning theories of connectivism. One can argue that organizing
neurons in layers is a matter of rearranging neurons positions spatially and this does not impose
any constraint on neurons connectivity. This is not true, even though, because by arranging
neurons in layers a neuron output is not allowed to connect to neurons in any layer other than the
next layer. It should be understandable, however, that ANN researchers thought to arrange ANN
in layers to facilitate the computational model of a network where each layer is represented by
two mathematical vectors, one for biases and another for weights.
Flow of information:
By looking on how information flows through a network, ANN can be divided into (1)
feedforward and (2) recurrent networks.

March 2017 Vol. 14 No.3.- 11 -
Fig. 6 Feedforward neural network
In feedforward networks, the output of a layer is used as an input for the next layer. There are no
loops in feedforward networks; information flows in one direction where the output of a neuron
can never return to its input. Feedforward network is one of the most used network structures.
The value of this structure is self-explanatory since it significantly reduces the network
complexity.
Fig. 7 Recurrent neural network
Recurrent network is a family of neural networks that processes the input sequentially and allows
feedback connections (Goodfellow et al. 2016). Feedforward network structure assumes that all
inputs are independent of each other (Britz 2015). It assumes that inputs order has no meaning.
This, however, turns out to be false assumption for some tasks. For example, in natural language
processing, the order of words makes a significant difference in meaning. Recurrent network tries
to recover this issue by allowing feedback in a network. The feedback is allowed but with a delay
constraint. That is, if the inputs are a sequence of A, B and C; then the output of hidden layer in
step A can only be passed to the input of the hidden layer in step B, not the hidden layer in step A
itself. To make a network simple, ANN researchers usually unfold the loop to see what it looks

March 2017 Vol. 14 No.3.- 12 -
like on each step of the inputs. In Fig. 8Fig. 8, one can see that a loop allows information to flow
from one step to another, and, therefore, acts as a memory (Britz 2015; Goodfellow et al. 2016;
Olah 2015).
Fig. 8 Unfolded recurrent neural network
Note that at t0, the input of neuron (X1,1) is 12 and the output of neuron (X3,1) is 9. At t2, the same
neuron (X1,1) receives the same input (12) but the neuron (X3,1) shows 11 as an output. This is
mainly because the hidden neuron (X2,1) receives feedback from its previous state at t1.
Flow of information and connection directionality are some of subjects discussed in connectivism
literature. Aldahdouh et al. (2015) showed that some connections in knowledge network are
bidirectional while others unidirectional. They also showed that "The node can connect to itself"
(p. 5). A latter concept is congruent with recurrent but not with feedforward network. However,
caution should be taken when comparing ANN architecture with connectivism. Researchers
restrict the flow of information in feedforward network and delay the feedback in recurrent
network because it is the only way they can control and compute the output; not because they
believe it is the right way of controlling the flow of information. With that said, one can consider
ANN structure as a special case of connectivism network structure. The second and very
important point to make here is the inclusion of time in network design. The time makes a
significant difference in network output and that is one of the common points with connectivism.
Connectivism’s principle of shifting reality matches the example given above. "While there is a

March 2017 Vol. 14 No.3.- 13 -
right answer now, it may be wrong tomorrow due to alterations in the information climate
affecting the decision" (Siemens, 2005, p. 5). Moreover, Aldahdouh et al. (2015) clearly called
for considering the time as one of knowledge dimensions.
Neuron connectivity:
By looking on how each neuron connects to other neurons, ANN can be divided into (1) fully
connected and (2) convolutional networks.
Fig. 9 Fully connected neural network
In fully connected network, each neuron in a specific layer is connected to all neurons in the next
layer. The idea of this connectivity is to allow maximum interactions between neurons. It is also
logical to think of fully connected network since we don't know in advance which connections
should be removed and which ones should be remained. It is the job of learning algorithm to
detect those connections. For example, if the connection between X1,2 and X2,1 should be removed
in order to generate the desired output, the learning algorithm should figure out that the weight of
this connection is 0. In other words, the learning algorithm should kill this link. One may wonder
why would killing a specific connection generate a desired output? Recall the car example and
how you downplayed the weight of parking cost. In some cases, you may even need to ignore the
input at all; for example, you may need to ignore traffic fine as a monthly cost of a car. Full
connectivity may add potentiality to the network but it adds severe difficulty on learning
algorithm as well. Adding tens of neurons to fully connected network increases the number of
weights and biases to be learned dramatically. Try to add two neurons to the hidden layer in Fig.
9 and notice how many new weights are added.
Fig. 10 Convolutional neural network

March 2017 Vol. 14 No.3.- 14 -
A convolutional network limits the connectivity between neurons so that a neuron in specific
layer is connected only to a set of spatially adjacent neurons in the previous layer (Goodfellow et
al. 2016; Olah 2014). Moreover, neurons in the convolutional layer should weigh up the
corresponding input neurons with the same values. In Fig. 10Error! Reference source not
found., the same color connections between the input and convolutional layer should have the
same value. Those connections are called shared weights. The output of convolutional layer is
often called a feature map. It is called so because when you arrange a layer as described, the
output would be detecting a single feature in the input (see Goodfellow et al., 2016 for details).
For example, if the input layer represents an image, a convolutional layer may detect a vertical
line in that image. A convolutional layer is usually followed by a pooling layer. A pooling layer
takes a feature map and tries to summarize it. For example, if a feature map detected a vertical
line in a tiny spot of the image, the pooling layer would summarize that in a larger region and
says: there is a vertical line in this region. The assumptions of convolutional network sound weird
and complicated. From where did those assumptions come? Actually, "Convolutional networks
are perhaps the greatest success story of biologically inspired artificial intelligence" (Goodfellow
et al., 2016, p. 364). A convolutional network was designed to capture the same functionality of
the primary visual cortex in the brain.
Connectivism appreciates network connectivity and seeks to increase it as much as possible.
Three out of eight connectivism’s principles refer directly to the value of the connection (Siemens
2005). Actually, connectivism defines learning as the process of connecting nodes in a network
(Aldahdouh et al. 2015; Siemens 2005). This may indicate that connectivism aims to make a
learner as a node in the fully connected network. However, it has been proved that increasing
connectivity adds complexity to ANN. This complexity makes learning harder and slower. A
convolutional network, on the other hand, decreases the connectivity and achieves better results.
Connectivists should pay attention to this because it disagrees with their main network designs
(see Downes 2010a work). In short, one can argue that connectivism agrees with fully connected
network but disagrees with convolutional network.
It is important to note that the classification shown above is superficial and AI researchers are
used to mixing network architectures together. For example, a network could be deep fully-
connected feedforward network or deep convolutional network. Sometimes, a network
architecture and its opposite can be mixed together. For example, a deep network may consist of
two convolutional layers followed by one fully-connected layer. In general, mixing different
network architectures shows better result and accuracy. In connectivism context, this may
indicate that mixing connectivist's network structure (fully connected network) with other limited
and loosely-connected structures would give us better educational results.
Learning algorithm
Designing network architectures is a difficult task but training and teaching these networks are
surely more difficult. To understand how ANN has been trained, it is better to start with a very
simple one neuron example (Goodfellow et al. 2016; Nielsen 2015). The principles which are
used to teach a single neuron are also used to teach a whole network. However, a network level
adds extra complexity which requires an additional step.
Suppose you have a very simple neuron with one input and one output. You want to teach this
neuron to do a certain task (for example to memorize a multiplication table for number 5). To
teach this neuron, ANN researchers usually give it a so-called training set. A training set contains
a number of different input values (1, 2, 3, 4, 5, 6 ...) paired with the correct output (5, 10, 15, 20,
25, 30 ...).

March 2017 Vol. 14 No.3.- 15 -
Fig. 11 Labeled training data
In the beginning, the neuron receives input and generates output according to its own
weight and bias which were randomly selected. This means, the output of the neuron (as)
would most probably differ from the correct output (ys).
Fig. 12 Single neuron training
The difference between the neuron output and the desired output presents something useful. The
function which measures the difference is often called a cost or loss function. There are many
ways to calculate the cost function. One of the simplest cost functions is the Mean Squared Error
(MSE):
å -=
x
xx ay
n
bwC 2
||||
2
1
),( (1)
MSE is the average of square of differences between the correct output (yx) and the output of the
neuron (ax) for each given input x in the training set. One may simply understand this function as
a way to measure the difference between all as and corresponding ys. It is also important to note
that a cost function is written as a function of weight (w) and bias (b). That is to say, this is a cost
of setting the weight and bias of the neuron in specific values. If we change the values of w and b,
then we should re-calculate the cost again. If the new cost was lower, this means that the
difference between the desired output (ys) and the neuron output (as) became smaller. That is, if
we found w and b that make C(w,b) approaching to 0, then we in fact have found the right value
of w and b. The job of learning algorithm is now to search for weights and biases that reduce the
cost to minimum.
Before going further in a learning algorithm, it is better to stop a while on some of the ideas
presented so far and match them to educational concepts. First, the labeled training data which
contains the input values along with correct output assumes knowledge as something static and
something we know in advance. Learning algorithm is not allowed to manipulate inputs or correct

March 2017 Vol. 14 No.3.- 16 -
outputs in any case (Nielsen 2015). This limits the ability of ANN to learn something previously
known, not to discover something new. The idea of static knowledge contradicts with
connectivism’s principle of dynamic knowledge. The second important point here is how we
could interpret this algorithm in an educational context. Let us continue with the connectivism’s
assumption that a single neuron represents a person. Thus, the inputs of the neuron would
represent the training material or the current learning experiences. The correct outputs represent
the reality (ontology) and neuron outputs represent a person's perceptions about the reality
(epistemology). The difference between neuron outputs and correct outputs represents the gap
between learner's perceptions and the reality. That is to say, learning is the process of minimizing
the gap between learner's perceptions and the reality. Of course, this definition perfectly fits
constructivist theory of learning. Jean Piaget (2013) interprets learning process mainly using two
sub-processes: assimilation and accommodation. Assimilation refers to learner's tendency to use
his current knowledge to deal with new situations while accommodation refers to learner's
tendency to change his current knowledge when it conflicts with reality. This theory clearly
matches the way used to teach a neuron using labeled training data. ANN researchers,
furthermore, insist that, in the learning stage, one should look at the gap between epistemology
and ontology not at the correctness of epistemology (Nielsen 2015). The reason for this claim is
that the number of correct answers is not smoothly related to the changes in weights and biases.
That is, in learning stage, a teacher should not count how many times a learner gives correct
answers and try to increase them. Instead, a teacher should focus on the gap between what a
learner believes and the reality and try to decrease it. In other words, if a learner gives two wrong
answers (123, 24) for a given question (5x5=?), these answers should not be treated equal.
Because when a learner says 24, it seems he learned something closer to the reality even though
the answer is not correct.
The time has come to see how a learning algorithm finds weight and bias that minimize the
output of a cost function. To understand what the algorithm does, it is better to plot a cost
function in relation to the variation of weight and bias. Since the output of a cost function (C)
depends on weight (w) and bias (b) of the neuron, then we may plot C in three-dimensional space
where each one of w, b and C represents one dimension.
Fig. 13 Cost function in relation to weight and bias

March 2017 Vol. 14 No.3.- 17 -
The variation of weight and bias pair in relation to a cost function may constitute any terrain
forms. Suppose it looks like a valley as shown in Fig. 13Fig. 13. Since we selected the value of
weight and bias randomly at the beginning, the initial values of weight and bias can represent any
point located on the surface. Suppose the point is located as shown. A learning algorithm should
find a way to roll the point down the hill and make it settle at the bottom. Finding a right direction
in three-dimensional space is not an easy task because it comprises watching the variation of
three variables at once. It is better to split the task so we watch every two variables alone. To do
so, we should pretend as if the third variable is constant.
Fig. 14 (1) Cost function in relation to bias. (2) Cost function in relation to weight
In the first plot in Fig. 14, we are interested in finding a tiny change in bias that makes the cost
smaller while keeping the weight constant (the same has to be said for the weight change in the
second plot). To find this tiny change, we have to find a ratio (∂c/∂b) that relates the change of
bias (∆b) to the change of cost while keeping the weight constant (∆C1). This ratio is known as
partial derivative.
b
b
c
c D
¶
¶
=D 1 , w
w
c
c D
¶
¶
=D 2 (2)
A total change in cost is the summation of ∆C1 and ∆C2.
w
w
c
b
b
c
c D
¶
¶
+D
¶
¶
=D (3)
Note that a change of cost ∆C also means the difference in the cost value before Cold and after
Cnew the change occurs.
oldnew ccc -=D (4)
Since we need a new cost Cnew to be smaller, this means ∆C should be negative. But how to
guarantee that ∆C is always negative? If we choose ∆w and ∆b as following, this would
guarantee ∆C to be always negative.

March 2017 Vol. 14 No.3.- 18 -
w
c
¶
¶
-=D hw ,
b
c
¶
¶
-=D hb (5)
Why choosing these values of ∆w and ∆b would guarantee ∆C to be always negative? Because by
substituting these choices into ∆C equation (3), one may easily find that these choices make sure
∆C negative.
ú
ú
û
ù
ê
ê
ë
é
÷
ø
ö
ç
è
æ
¶
¶
+÷
ø
ö
ç
è
æ
¶
¶
-=D
22
b
c
w
c
c h (6)
Selecting ∆w and ∆b repeatedly as such will roll the point down the curve slowly and keep it
settle at the bottom. This process is called gradient descent (see gradient in mathematics for
general perspective and Nielsen, 2015 for specific discussion). So far so good but what we
haven't mentioned yet what a factor (η) which appears in ∆w, ∆b and ∆C equations is called.
They call it learning rate.
To understand why ANN researchers called η factor a learning rate, it is better to
concentrate on one equation, take for example the equation of ∆b:
b
c
¶
¶
-=D hb (7)
Note first that a sign of η refers to the direction in which we want to go to. A negative sign means
we want to go down the curve. Now, if we choose η large, the step of ∆b becomes wide. And if
we choose η small, the step of ∆b becomes tiny. Therefore, η controls the speed of ∆b learning.
Since η appears on ∆w and ∆C equations as well, then we can say that η controls the speed of a
neuron learning and it is logical to call it learning rate. It seems tempting to increase η so the step
becomes wider and, hence, the point reaches the bottom faster. However, this is a false
conclusion because a slope (∂C/∂b) is only valid for a tiny shift of b. To understand why, look at
Fig. 15 below:
Fig. 15 Effect of selecting different value of bias on learning

March 2017 Vol. 14 No.3.- 19 -
Suppose you have a cost function in relation to bias as shown in the figure. The initial point is in
(x). The red line represents the slope (∂C/∂b) at this point. First, if we choose η very small, the
result would be a very small step like ∆b1. This is not a good strategy because it requires many
steps before we reach the bottom of the curve. Therefore, learning becomes very slow. On the
other hand, if we choose η very large, the result would be a very wide step like ∆b3. This is not a
good strategy as well. It makes the point jump to higher cost (y). In this case, the right choice of η
should be an intermediate value that produces step like ∆b2. How do ANN researchers tune the
value of η? So far, there is no rule and they depend merely on try-and-error strategy (Nielsen
2015)!
It is obvious that extracting an understandable educational interpretation out of this part of the
algorithm is not an easy task. This paper does not also claim that it will offer a comprehensive
interpretation. Instead, the interpretation coming shortly should be seen as an initial step toward
understanding machine learning algorithms in a humanitarian learning context. Cumulative
efforts from concerned researchers may eventually lead us to better understanding human and
machine learning.
Recall that the cost represents a gap between learner's epistemology and ontology. Thus, one may
argue that Cold represents the gap before passing through a learning experience. Likewise, Cnew
represents the gap after passing a learning experience. As a learner passes through a learning
experience, the gap reduces from Cold to Cnew. The gap shrinkage is ∆C. Therefore, ∆C represents
the learning outcome. The learning outcome stems from the change in a bias of a learner's internal
judgment system (∆b) and his own perspective (∆w). Changing student's perspectives ∆w and his
bias ∆b toward smaller gap between epistemology and ontology (Cnew < Cold) represents a
learning outcome ∆C. Or in short, the learning outcome refers to the extent of progress a learner
makes in bridging the gap between his epistemology and the ontology after passing learning
experience.
A learning rate refers simply to the speed of learning outcome. Or how fast a learner should learn.
The learning rate should not be too fast that makes a learner jump from point to point; long jumps
disrupt learning. A learning rate should not be very slow too; it makes a learner crawl in details
that would not serve him to achieve his goal. Finding the right pace of learning is a difficult task
that depends on the initial state of the learner’s perspectives and bias. The determinant factor of
learner’s speed of learning is where his epistemology is located in relation to the ontology. This
may interpret why each learner has his own learning rate and why the same learner may change
his rate from task to task.
One note in ANN model of learning is how AI researchers are setting the value of learning rate.
Actually, learning rate is one of many other parameters which are left free for human and outside
of ANN’s control. For example, (1) the number of layers, (2) the number of neurons in each
layer, (3) the size of training set, (4) the activation function type, and (5) regularization parameter
as well as (6) the learning rate are some of those free parameters which are called hyper-
parameters (Nielsen 2015). Choosing the right values of hyper-parameters is left for a person who
manages the ANN (see Fig. 16).

March 2017 Vol. 14 No.3.- 20 -
Fig. 16 Artificial Neural Network and human management
As we assumed in the beginning that ANN may represent a learner, this makes us wonder what
does this person who is playing with hyper-parameters represent in a human’s mind? Arguably,
hyper-parameters are a way in which the learner exercises control over his thoughts and learning
and, therefore, this person represents human’ agency or consciousness. Bandura (2006) contends
that human agency has four core properties which distinguish humans from automatons:
intentionality, forethought, self-reactiveness, and self-reflectiveness. These four properties are
sometime referred in educational literature as self-regulation and metacognitive processes.
According to Vermunt and Verloop (1999), metacognitive regulation activities are those thinking
activities learners use to decide on their goals, to exert control over their processing and affective
activities and to steer the course and outcomes of their activities. For purposes of illustration,
consider the following analogy of a software engineer and a learnable software. The software
engineer represents the consciousness who sets the goals, plans for experiences, monitors and
evaluates the progress of learning. A learnable software represents a neural pattern written in the
brain. This learnable software is an elastic program which can automatically detects its mistakes
and rewrites itself but under supervision of the engineer. The engineer manages, directs, and
gives instructions to the learnable software but does not engage in writing the software by hand.
The engineer is not a programmer and he does not even aware of how the software is written.
Once the software is written ‘correctly’, the consciousness releases its control over the written
software and the software is working deliberately. Only when something goes unexpected, the
consciousness comes back to exercise control and regulates the process of rewriting the software
again. Bandura (1999, 2006) criticizes connectionist’s view of human learning as it concentrates
merely on neural patterns to interpret learning and argues that this view strips humans of agentic
capabilities and a self-identity. In contrary, Bandura (2006) conceives consciousness as an
emergent property of brain activities which is not reducible solely to the property of neurons
activity. In other words, the consciousness is higher-level force which is a result of lower-level
neural activities but its properties are not limited to them. As clarified in this study, ANN design
shows the need for consciousness force to manage and regulate ANN learning but this force does
not occur as an emergent property of neural activity as Bandura proposes. Rather, it is a
completely distinct entity which uses, guides and manages the neural activity and does not result

March 2017 Vol. 14 No.3.- 21 -
from it. Siemens (2005) defines learning as a process of connecting specialized nodes but, as far
as we know, connectivism does not refer to learning rate and other hyper-parameters in its
assumptions. Connectivism is also criticized for its ambiguity in that it does not show how pattern
recognition is done (Aldahdouh et al. 2015). It is not clear in connectivism what the
characteristics of pattern recognition are.
Up to now, the paper presents learning in one neuron level. Even though the rules of learning in
one neuron level are applicable for a network level, the complexity and time spent in training a
network increase dramatically. Suppose that you have a deep fully-connected network as shown
in Fig. 17:
Fig. 17 Fully connected neural network with 44 neurons
In this network, you need to repeatedly calculate biases and weights according to the equations:
b
c
¶
¶
-=- holdnew bb ,
w
c
¶
¶
-=- holdnew ww (8)
Old values of weight (wold) and bias (bold) were given from the previous step or were set randomly
at the first step. You provide an arbitrary value for learning rate (η) and tune it by try-and-error
strategy. The remaining part is to calculate partial derivatives (∂c/∂b) and (∂c/∂w). Our network
has 44 neurons but a typical ANN may have millions of neurons. In such networks, finding
gradient becomes a tedious task. Part of the difficulty in finding partial derivatives for each
neuron returns to the fact that a tiny shift in single neuron weight or bias will propagate to all
neurons in the next layer. And the next layer will propagate changes to the next layer, and so on.
This tiny shift significantly changes the cost of the whole network. ANN researchers have found a
way to trace these changes called back-propagation. Back-propagation outperformed all other
methods used previously to compute gradient (Goodfellow et al. 2016; Nielsen 2015). The core
idea of back-propagation depends on calculating partial derivatives using multivariable chain rule
in mathematics. Presenting the mathematical model of back-propagation does not serve the
purpose of this study. Intuitively speaking, back-propagation starts by computing the error in the
output layer and then uses this error to compute the error in the preceding layers, one after
another. That is, it goes through the network backward. At the end, it uses the error matrix to

March 2017 Vol. 14 No.3.- 22 -
calculate the gradient that is required to compute the next values of bias bnew and weight wnew (see
Goodfellow et al., 2016 and Nielsen, 2015 for more details).
Combined with back-propagation, gradient descent algorithm moves gradually to reduce the gap
between correct output and network output. In each step of this movement, it computes the cost
function and its gradient. The result is that each neuron in the network moves one step toward
learning its right weights and bias. Eventually, the gradient descent reaches a point in which it
can't reduce the cost anymore. At that point, the network reaches the maximum approximation to
the correct output. This process is done automatically by gradient descent algorithm. In a network
of millions of neurons, what are those right weights and bias that a single neuron learns? What is
the meaning of those connections and biases? Why does each neuron connect to other neurons in
that way? Until now no one has a theory. As Nielsen (2015) put it,
In the early days of AI research people hoped that the effort to build an AI would also
help us understand the principles behind intelligence and, maybe, the functioning of the
human brain. But perhaps the outcome will be that we end up understanding neither the
brain nor how artificial intelligence works (ch. 1)!
Learning in network level clearly supports the core assumption of connectivism: A single
connection between two nodes does not have meaning in its own. A meaning is distributed across
group of connections or patterns (Aldahdouh et al. 2015; Downes 2006). Looking at the network
from higher level mitigates its complexity (Downes 2016). But that does not give us the answer
and the exact meaning of the entities in the lower level. The inner entities are well organized and
they certainly serve the purpose of the whole network but by looking at single entity or small
number of entities, one may fall in the illusion of finding conflicting and contradictory ideas with
the whole network.
Discussion
It has been argued that machine learning has nothing to do with human learning. Some
researchers in education field do not even see the difference between recent technologies and
traditional tools like books (Verhagen 2006). Connectivists, on the other hand, argue that learning
resides in non-human appliances (Siemens 2005). They insist on the relevance of machine
learning and, frequently, share ANN findings to support their assumptions. This paper tried to
bridge the gap and examine the relationship between connectivism’s assumptions and AI
findings. Table 1 below summarizes, in points, the relationship between ANN concepts and
connectivism’s assumptions.
Table 1
Relationship between ANN concepts and connectivism’s assumptions
ANN concept Yes No
Not
clear
Clarification
Perceptron neuron X The connection is not necessarily sharp; it can be
graded.
Neuron bias X A bias of neuron represents learner autonomy.
Sigmoid neuron X The graded feature of sigmoid is congruent with
connectivism but the flat part of the curve is not.
Neuron as logic gate X A group of logic gates interprets the emergent
property.

March 2017 Vol. 14 No.3.- 23 -
ANN concept Yes No
Not
clear
Clarification
Network Layers X Arranging network in layers contradicts with chaos
theory.
Feedforward net X Information in connectivism flows in both directions.
Recurrent net X Self-join connections are allowed in connectivism
Fully connected net X Connectivism seeks to increase connectivity as much
as possible.
Convolutional net X Connectivism does not see the value of limiting
connectivity.
Supervised learning X Supervised learning techniques are congruent with
constructivism and former learning theories.
Unsupervised learning X Unknown and dynamic knowledge can only be
learned by pattern recognition.
Learning rate X Connectivism has been criticized for its ambiguity in
interpreting the process of pattern recognition.
Learning in network
level
X A single connection between two nodes does not
have meaning in its own as connectivism suggests.
In this paper, we argue that the definition of learning as pattern recognition can only describe a
learning process under certain conditions. As shown in the paper, ANN researchers use
constructivist’s principles for teaching a network. Although not mentioned previously, we found
that AI researchers use almost all previous learning theories. For example, one technique in AI
mimics behaviorism and it is called reinforcement learning. In this technique, the agent learns
from a series of rewards and punishments (Russell and Norvig 2010). ANN researchers also use a
technique called Long-Short Term Memory which mimics cognitivism (Olah 2015). But all these
methods should be seen from the lens of known and static knowledge only. That is, if the
knowledge is known and static, ANN researchers may use labeled training data and make use of
the gap between the network output and the correct output. In a situation where knowledge is not
known or is not static, all these techniques will fail. The question of how to teach a network in
these conditions remains not fully answered in AI field. As Russell and Norvig (2010) say, "how
do we know that the hypothesis h is close to the target function f if we don't know what f is?"
(p. 713). In such cases, the only hope is to use unsupervised learning model in which a learner
should extract the pattern from given examples without explicit feedback. The repetition and
relative similarity between objects in given examples may help a learner to cluster and combine
different ideas together to come up with new object. And that is where connectivist's theory lies.
Even though connectivism’s assumptions are not congruent with some points in ANN, they are
certainly valuable in interpreting machine learning algorithm in general. Assuming knowledge as
a network; viewing learner as a node in network; and the connectivity between learners serve us
to understand the relationship between neurons and the complexity of artificial learning network.
Without such assumptions, one may find it extremely difficult to approach machine learning
algorithm from an educational perspective.
Studying ANN reveals that the flow of information should be free but educators should not seek
to increase the connectivity as much as they can. A fully-connected learning environment

March 2017 Vol. 14 No.3.- 24 -
increases the complexity and makes the learning process harder. The learner should be free to
share knowledge and to act within a reasonable number of valuable connections. In addition,
studying ANN shows that learning is a slow and smooth process. This is in comparison to
knowledge which is changing rapidly. Connectivism presents two solutions for this paradox:
(1) to allow students to copy-and-paste and (2) to store their information outside themselves.
Educators are also invited to find other innovative solutions.
This paper presents a very limited part of ANN literature. There are still some other topics that
are valuable and deserve careful study from an educational perspective. Such topics include the
number of nodes in each layer; different types of neurons; different cost functions; weight
initialization; overfitting and regularization; Long Short Term Memory technique; bidirectional
recurrent network; and Deep Belief Network.
References
Aldahdouh, Alaa A. and António J. Osório. 2016. “Planning to Design MOOC? Think First!” The
Online Journal of Distance Education and e-Learning 4(2):47–57.
Aldahdouh, Alaa A., António J. Osório, and Susana Caires. 2015. “Understanding Knowledge
Network, Learning and Connectivism.” International Journal of Instructional Technology
and Distance Learning 12(10):3–21. Retrieved
(http://www.itdl.org/Journal/Oct_15/Oct15.pdf#page=7).
Bandura, Albert. 1999. “Social Cognitive Theory: An Agentic Perspective.” Asian Journal of
Social Psychology 2:21–41.
Bandura, Albert. 2006. “Toward a Psychology of Human Agency.” Perspective on Psychological
Science 1(2):164–80. Retrieved (http://www.jstor.org/stable/40212163).
Bell, Frances. 2010. “Network Theories for Technology-Enabled Learning and Social Change:
Connectivism and Actor Network Theory.” Pp. 526–33 in 7th International Conference on
Networked Learning. Retrieved (http://usir.salford.ac.uk/9270/1/Bell.pdf).
Bell, Frances. 2011. “Connectivism: Its Place in Theory-Informed Research and Innovation in
Technology-Enabled Learning.” The International Review of Research in Open and
Distance Learning 12(3):98–118. Retrieved
(http://www.irrodl.org/index.php/irrodl/article/view/902/1664).
Brabazon, Tara. 2016. The University of Google: Education in the (Post) Information Age.
Routledge. Retrieved
(https://books.google.fi/books?hl=en&lr=&id=QdeXCwAAQBAJ&oi=fnd&pg=PT7&ots=J
bNthv0zbm&sig=QTtI95zOhlT9ibCjonqJeuO3he8&redir_esc=y#v=onepage&q&f=false).
Britz, Denny. 2015. “Recurrent Neural Networks Tutorial, Part 1 – Introduction to RNNs.”
Retrieved August 22, 2016 (http://www.wildml.com/2015/09/recurrent-neural-networks-
tutorial-part-1-introduction-to-rnns/).
Clarà, Marc and Elena Barberà. 2014. “Three Problems with the Connectivist Conception of
Learning.” Journal of Computer Assisted Learning (OCTOBER 2013).
Cormier, Dave. 2016. “Content Is a Print Concept.” Retrieved June 19, 2016
(http://davecormier.com/edblog/2016/06/11/content-is-a-print-concept/).
Dirckinck-Holmfield, Lone, Chris Jones, and Berner Lindström. 2009. Analysing Networked
Learning Practices in Higher Education and Continuing Professional Development. Sense
Publishers.
Downes, Stephen. 2005. “An Introduction to Connective Knowledge.” Pp. 77–102 in, edited by
K. Habitzel, T. D. Märk, and S. Prock. innsbruck university press. Retrieved

March 2017 Vol. 14 No.3.- 25 -
(http://www.oapen.org/download?type=document&docid=449459#page=84).
Downes, Stephen. 2006. “Learning Networks and Connective Knowledge.” Pp. 1–26 in
Collective intelligence and elearning, edited by H. H. Yang and S. C.-Y. Yuen. New York:
Information Science Reference. Retrieved
(https://books.google.fi/books?hl=en&lr=&id=GGx0GlvbYa0C&oi=fnd&pg=PA1&dq=Lea
rning+Networks+and+Connective+Knowledge&ots=t48QLHJ56R&sig=z-
9LiYdqRJG8H4Syp2fkT9WW5Z4&redir_esc=y#v=onepage&q=Learning Networks and
Connective Knowledge&f=false).
Downes, Stephen. 2009. “New Tools for Personal Learning.” Presentation delivered at the
MEFANET 2009 Conference, Brno, Czech Republic. Retrieved May 11, 2016
(http://www.downes.ca/presentation/234).
Downes, Stephen. 2010a. “Dimensions of a Learning Network.” Retrieved January 10, 2017
(http://www.slideshare.net/Downes/dimensions-of-a-learning-network-5560762).
Downes, Stephen. 2010b. “What Is Democracy in Education.” Retrieved February 1, 2016
(http://halfanhour.blogspot.pt/2010/10/what-is-democracy-in-education.html).
Downes, Stephen. 2012. Connectivism and Connective Knowledge: Essays on Meaning and
Learning Networks. Retrieved (http://www.downes.ca/files/books/Connective_Knowledge-
19May2012.pdf).
Downes, Stephen. 2016. “NRC01PL 07 - The MOOC Ecosystem - YouTube.” Retrieved
September 2, 2016 (https://www.youtube.com/watch?v=kxstCBbIrUs&feature=youtu.be).
Goodfellow, Ian, Yoshua Bengio, and Aaron Courville. 2016. Deep Learning. Book in
preparation for MIT Press. Retrieved (http://www.deeplearningbook.org/).
Kop, Rita and Adrian Hill. 2008. “Connectivism: Learning Theory of the Future or Vestige of the
Past?” International Review of Research in Open and Distance Learning 9(3). Retrieved
(http://www.irrodl.org/index.php/irrodl/article/view/523).
Nielsen, Michael a. 2015. Neural Networks and Deep Learning. Determination Press. Retrieved
(http://neuralnetworksanddeeplearning.com/).
Olah, Christopher. 2014. “Conv Nets: A Modular Perspective.” Retrieved August 23, 2016
(http://colah.github.io/posts/2014-07-Conv-Nets-Modular/).
Olah, Christopher. 2015. “Understanding LSTM Networks.” Retrieved August 22, 2016
(http://colah.github.io/posts/2015-08-Understanding-LSTMs/).
Piaget, Jean. 2013. The Construction of Reality in the Child. Routledge. Retrieved
(https://books.google.fi/books?hl=en&lr=&id=PpfGxMDZP-
4C&oi=fnd&pg=PP2&dq=The+construction+of+reality+in+the+child&ots=LLsJNq8Qud&
sig=RtAqeUgF2hA30rsG9-GOW4e9kn4&redir_esc=y#v=onepage&q=The construction of
reality in the child&f=false).
Russell, Stuart and Peter Norvig. 2010. Artificial Intelligence: A Modern Approach. Third Edit.
New Jersey: Prentice Hall. Retrieved (http://aima.cs.berkeley.edu/).
Siemens, George. 2005. “Connectivism: A Learning Theory for the Digital Age.” International
journal of instructional technology and distance learning 2(1):3–10. Retrieved
(http://www.itdl.org/Journal/Jan_05/article01.htm).
Siemens, George. 2006. Knowing Knowledge. Lulu. com. Retrieved
(http://www.elearnspace.org/KnowingKnowledge_LowRes.pdf).
Siemens, George and Peter Tittenberger. 2009. Handbook of Emerging Technologies for
Learning. Manitoba, Canada: University of Manitoba. Retrieved April 11, 2014
(http://elearnspace.org/Articles/HETL.pdf).

March 2017 Vol. 14 No.3.- 26 -
Verhagen, Pløn. 2006. “Connectivism : A New Learning Theory?” University of Twente 2(1):3–
10. Retrieved (http://elearning.surf.nl/e-learning/english/3793).
Vermunt, Jan D. and Nico Verloop. 1999. “Congruence and Friction between Learning and
Teaching.” Learning and Instruction 9(3):257–80.
Vermunt, Jan D. and Yvonne Vermetten. 2004. “Patterns in Student Learing: Relationships
between Learning Strategies, Conceptions of Learning and Learning Orientations.”
Educational Psychology Review 16(4):359–84.
Wang, Zhijun, Li Chen, and Terry Anderson. 2014. “A Framework for Interaction and Cognitive
Engagement in Connectivist Learning Contexts.” International Review of Research in Open
and Distance Learning 15(2):121–41.
About the author
Alaa A. AlDahdouh PhD candidate, Institute of Education of
University of Minho, Portugal. He has B.S in computer engineer with
cumulative experience in software programming. He is interested in
studying human and machine learning as well. His work focuses on
investigating new educational theories such as connectivism and their
role to enhance educators' understanding of learning process.
E-mail: AlaaAldahdouh@gmail.com
Find him on Social Media: ResearchGate & Twitter
Return to Table of Contents

Does artificial neural network support connectivism’s assumptions?

More Related Content

What's hot (19)

Similar to Does artificial neural network support connectivism’s assumptions? (20)

Recently uploaded (20)

Does artificial neural network support connectivism’s assumptions?