How to Train Neural Network?

Minaxi Surya

5 months ago

Neural Network
Neural Network
As we know one of The most important parts of deep learning is training the neural networks.
So, let's learn how it actually works.
In this article we will try to learn how a neural network gets to train. We will also learn about the feed-forward method and backpropagation method in Deep Learning.
how neural network works
how neural network works

Why training is needed?

Training in deep learning
Training in deep learning
Training in deep learning is the process that helps machines to learn about the function/equation. We have to find the optimal values of the weights of a neural network to get the desired output.
To train a neural network, we use the iterative method using gradient descent. Initially we start with random initialization of the weights. After random initialization of the weights, we make predictions on the data with the help of forward-propagation method, then we compute the corresponding cost function C, or loss and update each weight w by an amount proportional to dC/dw, i.e., the derivative of the cost functions w.r.t. the weight. The proportionality constant is known as the learning rate.
Now we might be thinking what is learning rate?
The learning rate is a type of hyper-parameter that helps us to controls the weights of our neural network with respect to the loss gradient. It gives us an idea of how quickly the neural network updates the concepts it has learned.
learning Rate
learning Rate
A learning rate ought not to be excessively low as it will set aside more effort to chat the system and it ought to likewise not be even as well high as the system may never get chat. In this way, it is consistently attractive to have an ideal benefit of learning rate with the goal that the system combines to something valuable.
We can ascertain the inclinations productively utilizing the back-engendering calculation. The key perception of in reverse proliferation or in the reverse prop is that as a result of the chain rule of
separation, the angle at every neuron in the neural system can be determined to utilize the inclination at the neurons, it has active edges to. Subsequently, we ascertain the inclinations in reverse, i.e., first compute the angles of the yield layer, at that point the top-most shrouded layer, trailed by the first concealed layer, etc, finishing at the information layer.
The back-proliferation calculation is executed for the most part utilizing the possibility of a computational diagram, where every neuron is ventured into numerous hubs in the computational diagram and performs a basic numerical activity like expansion, augmentation. The computational diagram doesn't have any loads on the edges; all loads are allowed to the hubs, so the loads become their own hubs. The retrogressive spread calculation is then a sudden spike in demand for the computational diagram. When the computation is complete, just the angles of the weight hubs are required for updates. The rest of the angles can be disposed of.
A portion of the improvement method are:

Gradient Descent Optimization Technique

Gradient Descent Optimization Technique
Gradient Descent Optimization Technique
One of the most regularly utilized streamlining procedures that alters the loads as indicated by the blunder/misfortune they caused which is known as "angle plummet."
Inclination is nothing yet, it is slant, and incline, on a x-y chart, speaks to how two factors are identified with one another: the ascent over the run, the change in
separation over the adjustment in time, and so forth. For this situation, the incline is the proportion between the system's mistake and a solitary weight; i.e., how does the blunder change as the weight is shifted.
To place it in a straight forward way, here we predominantly need to discover which loads which produces the least blunder. We need to discover the loads that effectively speaks to the signals contained in the information, and makes an interpretation of them to a right arrangement.
As a neural system learns, it gradually modifies numerous loads with the goal that they can plan signs to importance effectively. The proportion between arrange Error and every one of those loads is a subsidiary, dE/dw that figures the degree to which a slight change in weight causes a slight change in the mistake.
Each weight is simply one factor in a profound neural system that includes many changes; the sign of the weight goes through actuations capacities and afterward wholes more than a few layers, so we utilize the chain rule of analytics to work back through the system initiations and yields. This leads us to the weight being referred to, and its relationship to by and large blunder.
Given two factors, mistake and weight, are intervened by a third factor, actuation, through which the weight is passed. We can compute how an adjustment in weight influences a change in blunder by first ascertaining how an adjustment in enactment influences a change in Blunder, and how an adjustment in weight influences an adjustment in enactment.
The essential thought in profound learning is simply that modifying a model's loads because of the blunder it produces, until you can't diminish the mistake any more. The profound net trains gradually if the inclination esteem is little and quick if the worth is high. Any mistakes in preparing prompts wrong yields. The way toward preparing the nets from the yield back to the information is gotten back to proliferation or back prop. We realize that forward spread beginnings with the info and works forward.
Back prop does the converse/inverse computing the slope from right to left.
Each time we compute a slope, we go through all the past angles to that point.
Let us start at a hub in the yield layer. The edge utilizes the angle at that hub. As we return into the shrouded layers, it gets progressively mind-boggling. The result of two numbers somewhere in the range of 0 and 1 gives you a more modest number. The slope esteem continues getting littler and thus back prop sets aside a ton of effort to prepare and creates terrible precision.
Difficulties in Deep Learning Calculations There are sure challenges for both shallow neural systems and profound neural systems, as overfitting and calculation time.
DNNs are simple influenced by overfitting on the grounds that the utilization of included layers of deliberation which permit them to show uncommon conditions in the preparation information.
Regularization techniques, for example, drop out, early halting, information growth, and move learning are utilized during preparing to battle the issue of overfitting.
Drop out regularization arbitrarily discards units from the shrouded layers during preparing which helps in keeping away from uncommon conditions. DNNs mull over a few preparing boundaries, for example, the size, i.e., the number of layers and the number of units per layer, the learning rate, and introductory loads. Finding ideal
boundaries aren't generally pragmatic because of the significant expense in time and
computational assets. A few hacks, for example, clumping can accelerate calculation. The huge preparation intensity of GPUs has altogether helped the preparing process, as the network and vector calculations required are top-notch on the GPUs.


Dropout is a well-known regularization technique for neural networks. Deep neural networks are particularly prone to overfitting.
Let us now see what dropout is and how it works.
In the words of Geoffrey Hinton, one of the pioneers of Deep Learning, ‘If you have a deep neural net and it's not overfitting, you should probably be using a bigger one and using dropout’.
Dropout is a technique where during each iteration of gradient descent, we drop a set of randomly selected nodes. This means that we ignore some nodes randomly as if they do not exist.
Each neuron is kept with a probability of q and dropped randomly with probability 1-q. The value q may be different for each layer in the neural network. A value of 0.5 for the hidden layers, and 0 for input layer works well on a wide range of tasks.
During evaluation and prediction, no dropout is used. The output of each neuron is multiplied by q so that the input to the next layer has the same expected value.
The idea behind Dropout is as follows − In a neural network without dropout regularization, neurons develop co-dependency amongst each other which leads to overfitting.

Implementation trick

Dropout is implemented in libraries such as TensorFlow and Pytorch by keeping the output of the randomly selected neurons as 0. That is, though the neuron exists, its output is overwritten as 0.
Early Stopping
Implementation trick
Implementation trick
Here we try to  train neural networks using an iterative algorithm called gradient descent.
The idea behind early stopping is intuitive; we stop training when the error starts to increase. Here, by error, we mean the error measured on validation data, which is the part of training data used for tuning hyper-parameters. In this case, the hyper-parameter is the stop criteria.

Data Augmentation

Data Augmentation
Data Augmentation
It is where we increment the quantum of information we have or enlarge it by utilizing existing information furthermore, applying a few changes on it. The specific changes utilized depend on the undertaking we mean to accomplish. In addition, the changes that help the neural net rely upon its engineering.
For example, in numerous PC vision undertakings, for example, object order, a powerful information growth strategy is including new information focuses that are trimmed or interpreted variants of unique information.
At the point when a PC acknowledges a picture as information, it takes in a variety of pixel esteems. Allow us to state that the entire picture is moved left by 15 pixels. We apply a wide range of shifts in various ways, bringing about an increased dataset ordinarily the size of the first dataset.

Transfer Learning

Transfer Learning
Transfer Learning
The way toward taking a pre-prepared model and "adjusting" the model with our own dataset is called move learning. There are a few different ways to do this. A couple of ways are depicted beneath −
·       We train the pre-prepared model on an enormous dataset. At that point, we expel the last layer of the arrangement and supplant it with another layer with irregular loads. We train the pre-prepared model on a huge dataset. At that point, we evacuate the last layer of the organize and supplant it with another layer with arbitrary loads.
·       We at that point freeze loads of the various layers and train the system ordinarily. Here freezing the layers isn't changing the loads during inclination plunge or enhancement.
We at that point freeze loads of the various layers and train the system typically. Here freezing the layers isn't changing the loads during slope plummet or advancement.
The idea of driving this is the pre-prepared model will go about as a component extractor, and as it were the last layer will be prepared on the current errand.
I hope you enjoyed reading this article and finally, you came to know about How to Train Neural Network?
For more such blogs/courses on data science, machine learning, artificial intelligence and emerging new technologies do visit us at InsideAIML.
Thanks for reading…
Happy Learning…

Submit Review