Artur Skrzeta

AI Procurement Chat Bot

Intro

Utilizing Deep Learning for procurement bot that is able to answer most frequently asked questions regarding corporate internal web-shop.

Artificial Neural Network:

• Consisted of artificial neurons which has their iunputs and outputs.
• Inputs can be values of some features, e.g. image brightness or number of pixels (in image recognition app).
• All conections between input values and neurons are weighted somehow - value of given feature is multiplied by a weight
• We sum all the results of multiplication from each feature_value * weight and the sum goes through the so-called activation function:
  x1 = input0 * weight
x2 = input1 * weight
     adding bias prevents from getting nulls
  sum = x1 + x2 + bias
activ_func(sum)
     we can simplify this with below equation:
  output = F(∑(xw)+bias)
     where F is an activation function.
• Activation function does the following transformation:
  - if sum > 0 then output = 1
  - if sum < 0 then output = 0
• The calculated output is a neuron's prediction itself.


source: udemy

Deep Learning:

• The adjective "deep" refers to use of multiple neuron layers, which raw data input comes through, between input and output layer.
• First outer layer is an input layer, last outer layer is an output layer, layers between are so-called hidden layers.
• Hidden layers are those ones that accept input data, perform computations and produce output.
• When 3 or more hidden layers we can call neural network as deep network.
• Neural network progressively transforms data input starting from low-level (simplified) to high-level (detailed) features.
• That way in image recognition application, first hidden layer recognizes edges when last hidden layer recognizes a human's face.

Cost Function:

• We can evaluate performance of neurons using a cost fucnction that measures how far we are with prediction from an expected value.
• Cost Function measures the error - difference between output (predicted value) and reality (expected value).
• Natural sequence of measureing error is attempting to correct the prediction (getting closer to expected values) in order to decrease an error.
• The process of producing predictions and correcting predictions, getting lower and lower error, is what in fact we call learning.
• Getting lower error requeires us to minimize a cost function. We can achieve this by using the algorithm: gradient descent.

Gradient Descent:

• Algorithm used for finding the local minimum of a function.
• Finding the minimum of the cost function is the main task in the deep learning process.
  
  source: udemy
• Using gradient descent we can determine the best possible neural network's parameters, like weights in neuron's inputs, for minimizing the cost.
• Adjusting the optimal parameters across an entire network requires us to apply so-called backpropagation.
• When data processed once, backpropagation goes back through the neuron layers and recalculates error contribution of each neuron.

Here is what makes the bot working:

• Script converts human's language into numbers.
• Each occurance of a word is being tracked down into a list which finally serves to train the ML model.
• Script takes advantage of tensorflow library that is commonly used for deep learning projects.