Learning Problem&

Cognitive Modelling
Prof. Magdy M. Aboul-Ela

Dr. Heba Zaki

[Link]@[Link]
 What about learning?
• In the previous Lec, the neural networks were simple enough that we could
set the connection strengths by hand to perform the desired computations.
• This is not a feasible general strategy, since models get much more
complex
• Moreover, we would like computational tools for trying to understand
learning and cognitive development

2
 Connection strengths
• In practice, connection strengths are almost never set by hand.

3
4
Learning by optimizing an objective function

5
 Logistic regression = single connectionist
neuron
• Logistic regression only accepts 0 or 1 discrete output, whereas this
formulation can accept 0-1 continuous output.
• Logistic regression uses a negative log-likelihood loss. This measures the
performance of a classification model where the output is a probability value
between 0 and 1.
• It is designed to fit a Maximum Likelihood Estimate (MLE). This
ensures the model parameters are chosen to maximize the probability of the
observed data occurring.
6
 A naive, informal algorithm: Parameter
wiggling
• Computing the error:
• x = [1, 1]
• ŷ = 0.26 (predicted output)
• y = 1.0 (target)
• E(w,b) = 0.54 (squared error)
• What if we try wiggling each parameter?
• If we increase w0 by 0.02, our error E(w,b) becomes 0.51
• If we decrease w0 by 0.02, our error E(w,b) becomes 0.55 7
 A smarter algorithm: stochastic gradient
descent
• Computing the gradient tells us which direction to go for steepest
descent:

• α : learning rate

8
9
Learning via stochastic gradient descent

• Computing the error:

• x = [1, 1]
• ŷ = 0.26 (predicted output)
• y = 1.0 (target)
• E(w,b) = 0.54 (squared error)
= 2(0.26-1)(0.26)(1-0.26)1 = - 0.285
10
(if we increase weight, error goes down)
 Visualizing an update to the weights

Notice that the new prediction

is a little bit better! (closer to
• Update with gradient descent the target y = 1.0)
11
Learning via stochastic gradient descent (Cont.)

• We repeatedly cycle through each pattern in the training set, making updates
after each pattern as in the previous slide.

12
Failing to learn XOR with a linear classifier
(“one or the other, but not both”)

13
Learning XOR with a multi-layer classifier

14
15
 Backpropagation algorithm for computing
gradient
• Updates for these weights the same as before:
• Multi-step strategy:

Step 1) Compute how error changes as a function of hidden unit activation (we
worked most of this step out already for single layer net)
Step 2) Compute how hidden unit activation changes as a function of weight
• As before, update with gradient descent:
• Conceptually, still no different than wiggling w3 and seeing how error
changes! 16
 Important tricks for training neural networks

• The learning rate is extremely important

• Your model may not learn if you set the rate too high or too low.
• Often, you want to start the rate high and decrease it over the course of
learning.
• Mini-batch learning
• Instead of updating the network after every single training example,
we group several examples into a mini-batch.
• This is more stable and usually much faster 17
 Important tricks for training neural networks
(cont.)
• The are much better activation functions than the “sigmoid”.

• Classification requires a different loss and activation function

• Softmax output layer (for c possible classes)
• Loss Function (negative log-likelihood loss)
18
Backpropagati
on allows us
to efficiently
optimize
a wide range
of “deep
neural
network”
models 19
Key principles of deep learning for data science
and machine learning
• Learning is central. Models should have a very large number of
parameters, and you should feed your model as much data as
possible.
• Learning from raw data. Learn from the raw data if possible. It
is better to learn a representation of your data than to hand-craft
a representation.
• Gradient-based learning. Gradient descent is very effective in
high-dimensional parameter spaces (e.g., millions of weights).
20
RNN [memory]
• Type of deep learning model designed to process sequential data, such as text,
speech, or time series, by incorporating loops that allow it to maintain a form of
memory, considering past inputs to influence current predictions and outputs. the
output from the previous step is fed as input to the current step.
• In traditional neural networks, all the inputs and outputs are independent of each
other, but in cases when it is required to predict the next word of a sentence, the
previous words are required and hence there is a need to remember the previous
words.
• The main and most important feature of RNN is its Hidden state, which
remembers some information about a sequence. 21
 Hopfield Networks
• Hopfield networks are a beautiful form of Recurrent Artificial Neural
Networks (RNNs)
• Hopfield Networks were the first instance of associative neural
networks: (ANN)
RNN architectures which are capable of producing an emergent
associative memory.
• Associative memory, or content-addressable memory, is a system in which a
memory recall is initiated by the associability of an input pattern to a
memorized one.
22
 What is Recurrent Neural
Network (RNN)?

• These parameters are updated

using Backpropagation.

• However, since RNN works on

sequential data here we use an
updated backpropagation which
is known as
• Backpropagation through time
(BPTT).
Slide
23
 Self-Organizing Networks

• Self-Organizing Neural Network (SONN) most famously refers to the

Self-Organizing Map (SOM), also known as the Kohonen Map

• The SOM is a powerful type of Artificial Neural Network (ANN) used

for unsupervised learning, primarily for dimensionality reduction and
data visualization. It is designed to produce a low-dimensional (typically
2D) representation of the input data space.
24
The Problem of Long-Term Dependencies
[learning]
• “LSTMs,” a very special kind of recurrent neural network which works, for
many tasks, much better than the standard version.
• Almost all exciting results based RNNs are achieved with them.
• Sometimes, we only need to look at recent information to perform the present
task. For example, consider a language model trying to predict the next word
based on the previous ones.
• If we are trying to predict the last word in “the clouds are in the sky” we don’t
need any further context – it’s pretty obvious the next word is going to be sky.
25
 The Problem of Long-Term Dependencies
• Consider trying to predict the last word in the text
“I grew up in France… I speak fluent French.”
• Recent information suggests that the next word is probably the name of a
language, but if we want to narrow down which language, we need the context
of France, from further back.
• It’s entirely possible for the gap between the relevant information and the point
where it is needed to become very large.
• Unfortunately, as that gap grows, RNNs become unable to learn to connect the
information.
• Sadly, in practice, RNNs don’t seem to be able to learn them, but LSTMs don’t
have this problem!
26
 LSTM Networks architecture
• LSTM architecture has three gates
using which regulate the flow of
information.
• These gates are used to protect and
control the cell state.
• Whether information is useful or not is
decided by various gates in LSTM.
• These gates can learn and unlearn
information based on its relevance.

27Slide
27
 Applications of LSTM Networks
1- Text generation or language modelling involves the calculation of words
whenever a sequence of words is supplied as input.
Language models can be used at the level of characters or n-gram level as
well as at the sentence or the level of a paragraph.
2- Image processing is the process of the analysis of a photograph and
converting the result into sentences. In order to do this, we will need to have a
set of data consisting of many photos with the appropriate descriptive
captions.
28
 Applications of LSTM Networks (cont.)
3- Speech and Handwriting Recognition
4- Music generation is identical to text generation, where LSTMs can predict
the musical notes, not text, by studying a mix of notes fed into the input.
5- Language Translation involves translating a sequence of one language to a
similar sequence in a different language. Like image processing, an image-
based dataset that includes words and translations is cleaned first before the
relevant portion to build the model. An encoder-decoder LSTM model can
convert the input sequences into their formatted vector (encoding) and then
convert the translated version. 29