Basic Concepts of Deep Learning
Differentiate Artificial Intelligence(AI) vs. Machine Learning(ML) vs. Deep Learning(DL).
AI: It enables a machine to mimic human behavior.
ML: It is a technique to achieve ‘AI’ through algorithms trained with data. In ML, we’ve to train machine by ourselves. We extract features and feed them to machine then apply algorithms to it to train it. It is a subset of AI.
DL: It’s a type of ML inspired by human brain. In DL, the structure is called artificial neural network. In DL, machine learns itself using artificial neural network that mimics biological neural network. It’s a subset of ML.
Why activation function is used in a neural network?
In hidden layer(s), after summing function, activation function is applied which tells either neurons are activated or not. It also tells that either neurons should be fired or not.
_____________________________
Why cost function?
· It tells how good your model performance is. It’s also referred as loss/error.
· Its equation is C=1/2(Y-Y^)².
· Here, ‘Y’ shows original output and ‘Y^’ shows predicted output.
_____________________________
Why Gradient Descent?
It’s the optimization algorithm to minimize the cost function and to maximize the performance of model. It’s used to find the local and global minima of function.
_____________________________
Differentiate FeedForward Neural Network and Recurrent Neural Network.
FeedForward Neural Network:
· Signals travel in one direction from input to output.
· No feedback(loop).
· Considers only the current input.
· Can’t memorize the previous input like CNN.
Recurrent Neural Network:
· Signals travel in both direction making it a looped direction.
· Considers the current input along-with received inputs for generating the output layer.
· Has ability to memorize the last due to its internal memory.
· Example is : Sentiment analysis, text mining etc.
_____________________________
Differentiate Softmax function and ReLU function.
Softmax function:
· Generates output between ‘0’ and ‘1’.
· Divides each output such that total sum of the outputs is equal to ‘1’.
· Used in output layers.
ReLU function:
· Stands for ‘Rectified Linear Unit’ and it’s mostly used activation function.
· It gives output ‘X’ when X>=0.
· It’s used in hidden layers.
_____________________________
LSTM neural network
· It stands for ‘Long Short Term Memory’ neural network.
· It’s kind of neural network that’s capable of learning long-term dependencies.
· It remembers information for long periods of time and it’s its default behavior.
· It has three steps for that.
· First step: Decide what to forget, and what to remember.
· Second step: updates cell state values.
· Third step: Decide what part of current state makes it to output.
Differentiate Epoch, Batch, iteration.
Epoch: represents 1 or more iterations over the entire dataset.
Batch: we can’t pass entire dataset into neural network at once. So, we divide the whole dataset into batches.
Iteration: We can understand it with this example. If we’ve 10,000 images as data and a batch size of 200 size, then an epoch should run 10,000/200=50 iterations.
_____________________________
Define Boltzman machine.
· It’s an algorithm used for reduction, classification, regression, collaborative filtering, feature learning and topic modeling.
· It contains a set of visible and invisible units.
· It makes stochastic decision which means it determines that whether neuron is on or off.
· It contains only two layers (input layer and hidden layer).
_____________________________
Supervised learning vs. Unsupervised learning
Supervised learning:
· Machine is trained with the data that’s well labeled.
· It takes place in the presence of a supervisor or a teacher.
· We need to supervise a machine while training.
· Humans are involved in this type of learning.
· Its output is mostly known.
Unsupervised learning:
· No need to supervise a machine.
· Machine is trained with a data that’s not labeled.
· No need to supervise machine.
· No human involvement.
· Output is partially known and most unknown.
Deep Learning Frameworks
· TensorFlow
· DL4J
· PYTORCH
· Keras
· Caffe
· Microsoft Cognitive Toolkit
_____________________________
Limitations of Deep Learning
Deep learning is frequently used now-a-days but it has some limitations including:
Data:
DL is more efficient to deal with a large amount of unstructured data and neural network needs a mess of volume of data to train.
Computational Volume:
Training a neural network requires ‘Graphical Processing Units (GPUs)’ which have thousands of cores as compared to ‘CPUs’. And ‘GPUs’ are more expensive.
Training Time:
Deep neural network takes hours, months and days to train. As amount of data increases, the training time increases as the number of layers in network also increases.
_____________________________
So, students this is a brief article on the basic concepts of deep learning. If you like my article, don’t forget to hit ‘clap’.
Thank You.
