Combining Deep Neural Networks With Reinforcement Learning for Improved Performance

The performance of reinforcement learning can be improved by incorporating supervised learning techniques. Let us take a look at a concrete example.

You all might be familiar with the Roomba robot created by iRobot. The Roomba robot is perhaps the most popular robot vacuum sold in the United States. 


The Roomba is completely autonomous, moving around the room with ease, cleaning up dust, pet hair, and dirt along the way. In order to do its job, the Roomba contains a number of sensors that enable it to perceive the current state of the environment (i.e. your house). 

Let us suppose that the Roomba is governed by a reinforcement learning policy. This learning policy could be improved if we have accurate readings of the current state of the environment. And one way to improve these readings is to incorporate computer vision.

Since reinforcement learning depends heavily on accurate readings of the current state of the environment, we could use deep neural networks (a supervised learning technique) to pre-train the robot so that it can perform common computer vision tasks such as recognizing objects, localizing objects, and classifying objects before we even start running the reinforcement learning algorithm. These “readings” would improve the state portion of the reinforcement learning loop.

Deep neural networks have already displayed remarkable accuracy for computer vision problems. We can use these techniques to enable the robot to get a more accurate reading of the current state of the environment so that it can then take the most appropriate actions towards maximizing cumulative reward.

How Does Quickprop Address the Vanishing Gradient Problem in Cascade Correlation?

In order to answer this question, let us break it down piece-by-piece.

What is Quickprop?

Normal backpropagation when training an artificial feedforward neural network can take a long time to converge on the optimal set of weights. Updating the weights in the first layers of a neural network requires errors that are propagated backwards from the output layer. This procedure might work well for small networks, but when there are a lot of hidden layers, standard backpropagation can really slow down the training process.

Scott E. Fahlman, a Professor Emeritus at Carnegie Mellon University in the School of Computer Science, decided in the late 1980s to speed up backpropagation by inventing a new algorithm called Quickprop which was inspired by Newton’s method, a popular root-finding algorithm.

The idea behind QuickProp is to take the biggest steps possible without overstepping the solution (Kirk, 2015). Instead of the error information propagating backwards from the output layer (as is the case with standard backpropagation), QuickProp needs just the current and previous values of the weights, as well as the error derivative that was calculated during the most recent epoch. This twist on backpropagation means that the weight update process is now entirely local with respect to each node.

In Quickprop, we assume that the shape of the cost function for a given weight when near its optimal value is a paraboloid. With this assumption we can then follow a method similar to Newton’s method for converging on the optimal solution.

A paraboloid

The downside of Quickprop is that it can get chaotic if the surface of the cost function has a lot of local minima (Kirk, 2015).

Title Image Source: Wikimedia Commons

What is the Vanishing Gradient Problem?

The vanishing gradient problem is as follows: As a neural network contains more and more layers, the gradients of the cost function with respect to the weights at each of the nodes gets smaller and smaller, approaching zero. Teeny tiny gradients make the network increasingly more difficult to train.

The vanishing gradient problem is caused by the nature of backpropagation as well as the activation function (e.g. logistic sigmoid function). Backpropagation computes gradients using the chain rule. And since the sigmoid function generates gradients in the range (0, 1), at each step of training, you are multiplying a chain of small numbers together in order to calculate the gradients.

Consistently multiplying smaller and smaller gradient values produces smaller and smaller weight update values since the magnitude of the weight updates (i.e. step size) is proportional to the learning rate and the gradients. 

Thus, because the gradients become vanishingly small, the algorithm takes longer and longer to converge on an optimal set of weights. And since the weights will only make teeny tiny changes during each step of the training phase, the weights do not update effectively, and the neural network network losses accuracy. 

What is Cascade Correlation?

Cascade correlation is a method that addresses the question of how many hidden layers and how many hidden nodes we need. Instead of starting with a fixed network with a set number of hidden layers and nodes per hidden layer, the cascade correlation method starts with a minimal network and then trains and adds one hidden layer at a time. 

How Does Quickprop Address the Vanishing Gradient Problem in Cascade Correlation?

Some scientists have combined the Cascade correlation learning architecture with the Quickprop weight update rule (Abrahart, 2004). When combined in this fashion, Cascade correlation prevents there from being a vanishing gradient problem due to the fact the gradient is never actually propagated from output all the way back to input through the various layers. 

As noted by Marquez et al. (2018), “such training of deep networks in a cascade, directly circumvents the well-known vanishing gradient problem by ensuring that the output is always adjacent to the layer being trained.” This is because of the way the weights are locked as new layers are inserted, and the inputs are still tied directly to the outputs.

So to answer the question posed in the title, it’s really the cascade correlation architecture that has the greatest effect on reducing the vanishing gradient problem even though there is some benefit from the Quickprop algorithm.


Abrahart, Robert J., Pauline E. Kneale, and Linda M. See. Neural networks for hydrological modelling. Leiden New York: A.A. Balkema Publishers, 2004. Print.

Marquez, Enrique S., Jonathon S. Hare, Mahesan Niranjan: Deep Cascade Learning. IEEE Trans. Neural Netw. Learning Syst. 29(11): 5475-5485 (2018).

Kirk, Matthew, et al. Thoughtful Machine learning : A Test-driven Approach. Sebastopol, California: O’Reilly, 2015. Print.

Artificial Feedforward Neural Network With Backpropagation From Scratch

In this post, I will walk you through how to build an artificial feedforward neural network trained with backpropagation, step-by-step. We will not use any fancy machine learning libraries, only basic Python libraries like Pandas and Numpy.

Our end goal is to evaluate the performance of an artificial feedforward neural network trained with backpropagation and to compare the performance using no hidden layers, one hidden layer, and two hidden layers. Five different data sets from the UCI Machine Learning Repository are used to compare performance: Breast Cancer, Glass, Iris, Soybean (small), and Vote.

We will use our neural network to do the following:

  • Predict if someone has breast cancer
  • Identify glass type
  • Identify flower species
  • Determine soybean disease type
  • Classify a representative as either a Democrat or Republican based on their voting patterns

I hypothesize that the neural networks with no hidden layers will outperform the networks with two hidden layers. My hypothesis is based on the notion that the simplest solutions are often the best solutions (i.e. Occam’s Razor).

The classification accuracy of the algorithms on the data sets will be evaluated as follows, using five-fold stratified cross-validation:

  • Accuracy = (number of correct predictions)/(total number of predictions)

Title image source: Wikimedia commons

Table of Contents

What is an Artificial Feedforward Neural Network Trained with Backpropagation?



An artificial feed-forward neural network (also known as multilayer perceptron) trained with backpropagation is an old machine learning technique that was developed in order to have machines that can mimic the brain. Neural networks were the focus of a lot of machine learning research during the 1980s and early 1990s but declined in popularity during the late 1990s.

Since 2010, neural networks have experienced a resurgence in popularity due to improvements in computing speed and the availability of massive amounts of data with which to train large-scale neural networks. In the real world, neural networks have been used to recognize speech, caption images, and even help self-driving cars learn how to park autonomously.

The Brain as Inspiration for Artificial Neural Networks


In order to understand neural networks, it helps to first take a look at the basic architecture of the human brain. The brain has 1011 neurons (Alpaydin, 2014). Neurons are cells inside the brain that process information.

Each neuron contains a number of input wires called dendrites. Each neuron also has one output wire called an axon. The axon is used to send messages to other neurons. The axon of a sending neuron is connected to the dendrites of the receiving neuron via a synapse.

So, in short, a neuron receives inputs from dendrites, performs a computation, and sends the output to other neurons via the axon. This process is how information flows through the brain. The messages sent between neurons are in the form of electric pulses.

An artificial neural network, the one used in machine learning, is a simplified model of the actual human neural network explained above. It is typically composed of zero or more layers.


Each layer of the neural network is made up of nodes (analogous to neurons in the brain). Nodes of one layer are connected to nodes in another layer by connection weights, which are typically just floating-point numbers (e.g. 0.23342341). These numbers represent the strength of the connection between two nodes.

The job of a node in a hidden layer is to:

  1. Receive input values from each node in a preceding layer
  2. Compute a weighted sum of those input values
  3. Send that weighted sum through some activation function (e.g. logistic sigmoid function or hyperbolic tangent function)
  4. Send the result of the computation in #3 to each node in the next layer of the neural network.

Thus, the output from the nodes in a given layer becomes the input for all nodes in the next layer.

The output layer of a network does steps 1-3 above. However, the result of the computation from step #3 is a class prediction instead of an input to another layer (since the output layer is the final layer).

Here is a diagram of the process I explained above:

Here is a diagram showing a single layer neural network:

b stands for the bias term. This is a constant. It is like the b in the equation for a line, y = mx + b. It enables the model to have flexibility because, without that bias term, you cannot as easily adapt the weighted sum of inputs (i.e. mx) to fit the data (i.e. in the example of a simple line, the line cannot move up and down the y-axis without that b term).

w in the diagram above stands for the weights, and x stands for the input values.

Here is a similar diagram, but now it is a two-layer neural network instead of single layer.

And here is one last way to look at the same thing I explained above:


Note that the yellow circles on the left represent the input values. w represents the weights. The sigma inside the box means that we calculated the weighted sum of the input values. We run that through the activation function f(S)…e.g. sigmoid function. And then out of that, pops the output, which is passed on to the nodes in the following layer.

Neural networks that contain many layers, for example more than 100, are called deep neural networks. Deep neural networks are the cornerstone of the rapidly growing field known as deep learning.

Training Phase

The objective during the training phase of a neural network is to determine all the connection weights. At the start of training, the weights of the network are initialized to small random values close to 0. After this step, training proceeds to the two main phases of the algorithm: forward propagation and backpropagation.

Forward Propagation

During the forward propagation phase of a neural network, we process one instance (i.e. one set of inputs) at a time. Hidden layers extract important features contained in the input data by computing a weighted sum of the inputs and running the result through the logistic sigmoid activation function. This output relays to nodes in the next hidden layer where the data is transformed yet again. This process continues until the data reaches the output layer.

The output of the output layer is a predicted class value, which in this project is a one-hot encoded class prediction vector. The index of the vector corresponds to each class. For example, if a 1 is in the 0 index of the vector (and a 0 is in all other indices of the vector), the class prediction is class 0. Because we are dealing with 0s and 1s, the output vector can also be considered the probability that an instance is in a given class.


After the input signal produced by a training instance propagates through the network one layer at a time to the output layer, the backpropagation phase commences. An error value is calculated at the output layer. This error corresponds to the difference between the class predicted by the network and the actual (i.e. true) class of the training instance.

The prediction error is then propagated backward from the output layer to the input layer. Blame for the error is assigned to each node in each layer, and then the weights of each node of the neural network are updated accordingly (with the goal to make more accurate class predictions for the next instance that flows through the neural network) using stochastic gradient descent for the weight optimization procedure.

Note that weights of the neural network are adjusted on a training instance by training instance basis. This online learning method is the preferred one for classification problems of large size (Ĭordanov & Jain, 2013).

The forward propagation and backpropagation phases continue for a certain number of epochs. A single epoch finishes when each training instance has been processed exactly once.

Testing Phase

Once the neural network has been trained, it can be used to make predictions on new, unseen test instances. Test instances flow through the network one-by-one, and the resulting output (which is a vector of class probabilities) determines the classification. 

Helpful Video

Below is a helpful video by Andrew Ng, a professor at Stanford University, that explains neural networks and is helpful for getting your head around the math. The video gets pretty complicated in some spots (esp. where he starts writing all sorts of mathematical notation and derivatives). My advice is to lookup anything that he explains that isn’t clear. Take it slow as you are learning about neural networks. There is no rush. This stuff isn’t easy to understand on your first encounter with it. Over time, the fog will begin to lift, and you will be able to understand how it all works.

Return to Table of Contents

Artificial Feedforward Neural Network Trained with Backpropagation Algorithm Design

The Logistic Regression algorithm was implemented from scratch. The Breast Cancer, Glass, Iris, Soybean (small), and Vote data sets were preprocessed to meet the input requirements of the algorithms. I used five-fold stratified cross-validation to evaluate the performance of the models.

Required Data Set Format for Feedforward Neural Network Trained with Backpropagation

Columns (0 through N)

  • 0: Instance ID
  • 1: Attribute 1
  • 2: Attribute 2
  • 3: Attribute 3
  • N: Actual Class

The program then adds two additional columns for the testing set.

  • N + 1: Predicted Class
  • N + 2: Prediction Correct? (1 if yes, 0 if no)

Breast Cancer Data Set

This breast cancer data set contains 699 instances, 10 attributes, and a class – malignant or benign (Wolberg, 1992).

Modification of Attribute Values

  • The actual class value was changed to “Benign” or “Malignant.”
  • Attribute values were normalized to be in the range 0 to 1.
  • Class values were vectorized using one-hot encoding.

Missing Data

There were 16 missing attribute values, each denoted with a “?”. I chose a random number between 1 and 10 (inclusive) to fill in the data.

Glass Data Set

This glass data set contains 214 instances, 10 attributes, and 7 classes (German, 1987). The purpose of the data set is to identify the type of glass.

Modification of Attribute Values

  • Attribute values were normalized to be in the range 0 to 1.
  • Class values were vectorized using one-hot encoding.

Missing Data

There are no missing values in this data set.

Iris Data Set

This data set contains 3 classes of 50 instances each (150 instances in total), where each class refers to a different type of iris plant (Fisher, 1988).

Modification of Attribute Values

  • Attribute values were normalized to be in the range 0 to 1.
  • Class values were vectorized using one-hot encoding.

Missing Data

There were no missing attribute values.

Soybean Data Set (small)

This soybean (small) data set contains 47 instances, 35 attributes, and 4 classes (Michalski, 1980). The purpose of the data set is to determine the disease type.

Modification of Attribute Values

  • Attribute values were normalized to be in the range 0 to 1.
  • Class values were vectorized using one-hot encoding.
  • Attribute values that were all the same value were removed.

Missing Data

There are no missing values in this data set.

Vote Data Set

This data set includes votes for each of the U.S. House of Representatives Congressmen (435 instances) on the 16 key votes identified by the Congressional Quarterly Almanac (Schlimmer, 1987). The purpose of the data set is to identify the representative as either a Democrat or Republican.

  • 267 Democrats
  • 168 Republicans

Modification of Attribute Values

  • I did the following modifications:
    • Changed all “y” to 1 and all “n” to 0.
  • Class values were vectorized using one-hot encoding.

Missing Data

Missing values were denoted as “?”. To fill in those missing values, I chose random number, either 0 (“No”) or 1 (“Yes”).

Stochastic Gradient Descent

I used stochastic gradient descent for optimizing the weights.

In normal gradient descent, we need to calculate the partial derivative of the cost function with respect to each weight. For each partial derivative, we have to tally up the terms for each training instance to compute the partial derivative of the cost function with respect to that weight. What this means is that, if we have a lot of attributes and a large dataset, gradient descent is slow. For this reason, stochastic gradient descent was chosen since weights are updated after each training instance (as opposed to after all training instances).

Here is a good video that explains stochastic gradient descent.

Logistic (Sigmoid) Activation Function

The logistic (sigmoid) activation function was used for the nodes in the neural network.

Description of Any Tuning Process Applied

Learning Rate

Some tuning was performed in this project. The learning rate was set to 0.1, which was different than the 0.01 value that is often used for multi-layer feedforward neural networks (Montavon, 2012). Lower values resulted in much longer training times and did not result in large improvements in classification accuracy.


The number of epochs chosen for the main runs of the algorithm on the data sets was chosen to be 1000. Other values were tested, but the number of epochs did not have a large impact on classification accuracy.

Number of Nodes per Hidden Layer

In order to tune the number of nodes per hidden layer, I used a constant learning rate and constant number of epochs. I then calculated the classification accuracy for each data set for a set number of nodes per hidden layer. I performed this process using networks with one hidden layer and networks with two hidden layers. The results of this tuning process are below.


Note that the mean classification accuracy across all data sets when one hidden layer was used for the neural network reached a peak at eight nodes per hidden layer. This value of eight nodes per hidden layer was used for the actual runs on the data sets.

For two hidden layers, the peak mean classification accuracy was attained at five nodes per hidden layer. Thus, when the algorithm was run on the data sets for two hidden layers, I used five nodes per hidden layer for each data set to compare the classification accuracy across the data sets.

Return to Table of Contents

Artificial Feedforward Neural Network Trained with Backpropagation Algorithm in Python, Coded From Scratch

Here are the preprocessed data sets:

Here is the full code for the neural network. This is all you need to run the program:

import pandas as pd # Import Pandas library 
import numpy as np # Import Numpy library
from random import shuffle # Import shuffle() method from the random module
from random import seed # Import seed() method from the random module
from random import random # Import random() method from the random module
from collections import Counter # Used for counting
from math import exp # Import exp() function from the math module

# File name:
# Author: Addison Sears-Collins
# Date created: 7/30/2019
# Python version: 3.7
# Description: An artificial feedforward neural network trained 
#   with backpropagation (also called multilayer perceptron)

# Required Data Set Format
# Columns (0 through N)
# 0: Attribute 0
# 1: Attribute 1 
# 2: Attribute 2
# 3: Attribute 3 
# ...
# N: Actual Class

# 2 additional columns are added for the test set.
# N + 1: Predicted Class
# N + 2: Prediction Correct?

ALGORITHM_NAME = "Feedforward Neural Network With Backpropagation"
SEPARATOR = ","  # Separator for the data set (e.g. "\t" for tab data)

def normalize(dataset):
    Normalize the attribute values so that they are between 0 and 1, inclusive
    :param pandas_dataframe dataset: The original dataset as a Pandas dataframe
    :return: normalized_dataset
    :rtype: Pandas dataframe
    # Generate a list of the column names 
    column_names = list(dataset) 

    # For every column except the actual class column
    for col in range(0, len(column_names) - 1):  
        temp = dataset[column_names[col]] # Go column by column
        minimum = temp.min() # Get the minimum of the column
        maximum = temp.max() # Get the maximum of the column

        # Normalized all values in the column so that they
        # are between 0 and 1.
        # x_norm = (x_i - min(x))/(max(x) - min(x))
        dataset[column_names[col]] = dataset[column_names[col]] - minimum
        dataset[column_names[col]] = dataset[column_names[col]] / (
            maximum - minimum)

    normalized_dataset = dataset

    return normalized_dataset

def get_five_stratified_folds(dataset):
    Implementation of five-fold stratified cross-validation. Divide the data
    set into five random folds. Make sure that the proportion of each class 
    in each fold is roughly equal to its proportion in the entire data set.
    :param pandas_dataframe dataset: The original dataset as a Pandas dataframe
    :return: five_folds
    :rtype: list of folds where each fold is a list of instances(i.e. examples)
    # Create five empty folds
    five_folds = list()
    fold0 = list()
    fold1 = list()
    fold2 = list()
    fold3 = list()
    fold4 = list()

    # Get the number of columns in the data set
    class_column = len(dataset[0]) - 1

    # Shuffle the data randomly

    # Generate a list of the unique class values and their counts
    classes = list()  # Create an empty list named 'classes'

    # For each instance in the dataset, append the value of the class
    # to the end of the classes list
    for instance in dataset:

    # Create a list of the unique classes
    unique_classes = list(Counter(classes).keys())

    # For each unique class in the unique class list
    for uniqueclass in unique_classes:

        # Initialize the counter to 0
        counter = 0
        # Go through each instance of the data set and find instances that
        # are part of this unique class. Distribute them among one
        # of five folds
        for instance in dataset:

            # If we have a match
            if uniqueclass == instance[class_column]:

                # Allocate instance to fold0
                if counter == 0:

                    # Append this instance to the fold

                    # Increase the counter by 1
                    counter += 1

                # Allocate instance to fold1
                elif counter == 1:

                    # Append this instance to the fold

                    # Increase the counter by 1
                    counter += 1

                # Allocate instance to fold2
                elif counter == 2:

                    # Append this instance to the fold

                    # Increase the counter by 1
                    counter += 1

                # Allocate instance to fold3
                elif counter == 3:

                    # Append this instance to the fold

                    # Increase the counter by 1
                    counter += 1

                # Allocate instance to fold4

                    # Append this instance to the fold

                    # Reset the counter to 0
                    counter = 0

    # Shuffle the folds

    # Add the five stratified folds to the list

    return five_folds

def initialize_neural_net(
    no_inputs, no_hidden_layers, no_nodes_per_hidden_layer, no_outputs):
    Generates a new neural network that is ready to be trained.
    Network (list of layers): 0+ hidden layers, and output layer
    Input Layer (list of attribute values): A row from the training set 
    Hidden Layer (list of dictionaries): A set of nodes (i.e. neurons)
    Output Layer (list of dictionaries): A set of nodes, one node per class
    Node (dictionary): Contains a set of weights, one weight for each input 
      to the layer containing that node + an additional weight for the bias.
      Each node is represented as a dictionary that stores key-value pairs
      Each key corresponds to a property of that node (e.g. weights).
      Weights will be initialized to small random values between 0 and 1.
    :param int no_inputs: Numper of inputs (i.e. attributes)
    :param int no_hidden_layers: Numper of hidden layers (0 or more)
    :param int no_nodes_per_hidden_layer: Numper of nodes per hidden layer
    :param int no_outputs: Numper of outputs (one output node per class)
    :return: network
    :rtype:list (i.e. list of layers: hidden layers, output layer)

    # Create an empty list
    network = list()

    # Create the the hidden layers
    hidden_layer = list()
    hl_counter = 0

    # Create the output layer
    output_layer = list()

    # If this neural network contains hidden layers
    if no_hidden_layers > 0:

        # Build one hidden layer at a time
        for layer in range(no_hidden_layers):

            # Reset to an empty hidden layer
            hidden_layer = list()

            # If this is the first hidden layer
            if hl_counter == 0:

                # Build one node at a time
                for node in range(no_nodes_per_hidden_layer):

                    initial_weights = list()
                    # Each node in the hidden layer has no_inputs + 1 weights, 
                    # initialized to a random number in the range [0.0, 1.0)
                    for i in range(no_inputs + 1):

                    # Add the node to the first hidden layer

                # Finished building the first hidden layer
                hl_counter += 1

                # Add this first hidden layer to the front of the neural 
                # network

            # If this is not the first hidden layer

                # Build one node at a time
                for node in range(no_nodes_per_hidden_layer):

                    initial_weights = list()
                    # Each node in the hidden layer has 
                    # no_nodes_per_hidden_layer + 1 weights, initialized to 
                    # a random number in the range [0.0, 1.0)
                    for i in range(no_nodes_per_hidden_layer + 1):


                # Add this newly built hidden layer to the neural network

        # Build the output layer
        for outputnode in range(no_outputs):

            initial_weights = list()
            # Each node in the output layer has no_nodes_per_hidden_layer 
            # + 1 weights, initialized to a random number in 
            # the range [0.0, 1.0)
            for i in range(no_nodes_per_hidden_layer + 1):

            # Add this output node to the output layer

        # Add the output layer to the neural network
    # A neural network has no hidden layers

        # Build the output layer
        for outputnode in range(no_outputs):
            initial_weights = list()
            # Each node in the hidden layer has no_inputs + 1 weights, 
            # initialized to a random number in the range [0.0, 1.0)
            for i in range(no_inputs + 1):

            # Add this output node to the output layer


    # Finished building the initial neural network
    return network

def weighted_sum_of_inputs(weights, inputs):
    Calculates the weighted sum of inputs plus the bias
    :param list weights: A list of weights. Each node has a list of weights.
    :param list inputs: A list of input values. These can be a single row
        of attribute values or the output from nodes from the previous layer
    :return: weighted_sum
    :rtype: float
    # We assume that the last weight is the bias value
    # The bias value is a special weight that does not multiply with an input
    # value (or we could assume its corresponding input value is always 1)
    # The bias is similar to the intercept constant b in y = mx + b. It enables
    # a (e.g. sigmoid) curve to be shifted to create a better fit
    # to the data. Without the bias term b, the line always goes through the 
    # origin (0,0) and cannot adapt as well to the data.
    # In y = mx + b, we assume b * x_0 where x_0 = 1

    # Initiate the weighted sum with the bias term. Assume the last weight is
    # the bias term
    weighted_sum = weights[-1]

    for index in range(len(weights) - 1):
        weighted_sum += weights[index] * inputs[index]

    return weighted_sum

def sigmoid(weighted_sum_of_inputs_plus_bias):
    Run the weighted sum of the inputs + bias through
    the sigmoid activation function.
    :param float weighted_sum_of_inputs_plus_bias: Node summation term
    :return: sigmoid(weighted_sum_of_inputs_plus_bias)
    return 1.0 / (1.0 + exp(-weighted_sum_of_inputs_plus_bias))

def forward_propagate(network, instance):
    Instances move forward through the neural network from one layer
    to the next layer. At each layer, the outputs are calculated for each 
    node. These outputs are the inputs for the nodes in the next layer.
    The last set of outputs is the output for the nodes in the output 
    :param list network: List of layers: 0+ hidden layers, 1 output layer
    :param list instance (a single training/test instance from the data set)
    :return: outputs
    :rtype: list
    inputs = instance

    # For each layer in the neural network
    for layer in network:

        # These will store the outputs for this layer
        new_inputs = list()

        # For each node in this layer
        for node in layer:

            # Calculate the weighted sum + bias term
            weighted_sum = weighted_sum_of_inputs(node['weights'], inputs)

            # Run the weighted sum through the activation function
            # and store the result in this node's dictionary.
            # Now the node's dictionary has two keys, weights and output.
            node['output'] = sigmoid(weighted_sum)

            # Used for debugging

            # Add the output of the node to the new_inputs list

        # Update the inputs list
        inputs = new_inputs

    # We have reached the output layer
    outputs = inputs

    return outputs

def sigmoid_derivative(output):
    The derivative of the sigmoid activation function with respect 
    to the weighted summation term of the node.
    Formally (after a lot of calculus), this derivative is:
        derivative = sigmoid(weighted_sum_of_inputs_plus_bias) * 
        (1 - sigmoid(weighted_sum_of_inputs_plus_bias))
                   = node_ouput * (1 - node_output)
    This method is used during the backpropagation phase. 
    :param list output: Output of a node (generated during the forward
        propagation phase)
    :return: sigmoid_der
    :rtype: float
    return output * (1.0 - output)

def back_propagate(network, actual):
    In backpropagation, the error is computed between the predicted output by 
    the network and the actual output as determined by the data set. This error 
    propagates backwards from the output layer to the first hidden layer. The 
    weights in each layer are updated along the way in response to the error. 
    The goal is to reduce the prediction error for the next training instance 
    that forward propagates through the network.
    :param network list: The neural network
    :param actual list: A list of the actual output from the data set
    # Iterate in reverse order (i.e. starts from the output layer)
    for i in reversed(range(len(network))):

        # Work one layer at a time
        layer = network[i]

        # Keep track of the errors for the nodes in this layer
        errors = list()

        # If this is a hidden layer
        if i != len(network) - 1:

            # For each node_j in this hidden layer
            for j in range(len(layer)):

                # Reset the error value
                error = 0.0

                # Calculate the weighted error. 
                # The error values come from the error (i.e. delta) calculated
                # at each node in the layer just to the "right" of this layer. 
                # This error is weighted by the weight connections between the 
                # node in this hidden layer and the nodes in the layer just 
                # to the "right" of this layer.
                for node in network[i + 1]:
                    error += (node['weights'][j] * node['delta'])

                # Add the weighted error for node_j to the
                # errors list
        # If this is the output layer

            # For each node in the output layer
            for j in range(len(layer)):
                # Store this node (i.e. dictionary)
                node = layer[j]

                # Actual - Predicted = Error
                errors.append(actual[j] - node['output'])

        # Calculate the delta for each node_j in this layer
        for j in range(len(layer)):
            node = layer[j]

            # Add an item to the node's dictionary with the 
            # key as delta.
            node['delta'] = errors[j] * sigmoid_derivative(node['output'])

def update_weights(network, instance, learning_rate):
    After the deltas (errors) have been calculated for each node in 
    each layer of the neural network, the weights can be updated.
    new_weight = old_weight + learning_rate * delta * input_value
    :param list network: List of layers: 0+ hidden layers, 1 output layer
    :param list instance: A single training/test instance from the data set
    :param float learning_rate: Controls step size in the stochastic gradient
        descent procedure.
    # For each layer in the network
    for layer_index in range(len(network)):

        # Extract all the attribute values, excluding the class value
        inputs = instance[:-1]

        # If this is not the first hidden layer
        if layer_index != 0:

            # Go through each node in the previous layer and add extract the
            # output from that node. The output from the previous layer
            # is the input to this layer.
            inputs = [node['output'] for node in network[layer_index - 1]]

        # For each node in this layer
        for node in network[layer_index]:

            # Go through each input value
            for j in range(len(inputs)):
                # Update the weights
                node['weights'][j] += learning_rate * node['delta'] * inputs[j]
            # Updating the bias weight 
            node['weights'][-1] += learning_rate * node['delta']

def train_neural_net(
    network, training_set, learning_rate, no_epochs, no_outputs):
    Train a neural network that has already been initialized.
    Training is done using stochastic gradient descent where the weights
    are updated one training instance at a time rather than after the
    entire training set (as is the case with gradient descent).
    :param list network: The neural network, which is a list of layers
    :param list training_set: A list of training instances (i.e. examples)
    :param float learning_rate: Controls step size of gradient descent
    :param int no_epochs: How many passes we will make through training set
    :param int no_outputs: The number of output nodes (equal to # of classes)
    # Go through the entire training set a fixed number of times (i.e. epochs)
    for epoch in range(no_epochs):
        # Update the weights one instance at a time
        for instance in training_set:

            # Forward propagate the training instance through the network
            # and produce the output, which is a list.
            outputs = forward_propagate(network, instance)

            # Vectorize the output using one hot encoding. 
            # Create a list called actual_output that is the same length 
            # as the number of outputs. Put a 1 in the place of the actual 
            # class.
            actual_output = [0 for i in range(no_outputs)]
            actual_output[int(instance[-1])] = 1
            back_propagate(network, actual_output)
            update_weights(network, instance, learning_rate)

def predict_class(network, instance):
    Make a class prediction given a trained neural network and
    an instance from the test data set.
    :param list network: The neural network, which is a list of layers
    :param list instance: A single training/test instance from the data set
    :return class_prediction
    :rtype int
    outputs = forward_propagate(network, instance)

    # Return the index that has the highest probability. This index
    # is the class value. Assume class values begin at 0 and go
    # upwards by 1 (i.e. 0, 1, 2, ...)
    class_prediction = outputs.index(max(outputs))
    return class_prediction

def calculate_accuracy(actual, predicted):
    Calculates the accuracy percentages
    :param list actual: Actual class values
    :param list predicted: predicted class values
    :return: classification_accuracy
    :rtype: float (as a percentage)
    number_of_correct_predictions = 0
    for index in range(len(actual)):
        if actual[index] == predicted[index]:
            number_of_correct_predictions += 1
    classification_accuracy = (
        number_of_correct_predictions / float(len(actual))) * 100.0
    return classification_accuracy

def get_test_set_predictions(
    training_set, test_set, learning_rate, no_epochs, 
    no_hidden_layers, no_nodes_per_hidden_layer):
    This method is the workhorse. 
    A new neutal network is initialized.
    The network is trained on the training set.
    The trained neural network is used to generate predictions on the
    test data set.
    :param list training_set
    :param list test_set
    :param float learning_rate
    :param int no_epochs
    :param int no_hidden_layers
    :param int no_nodes_per_hidden_layer
    :return network, class_predictions
    :rtype list, list
    # Get the number of attribute values
    no_inputs = len(training_set[0]) - 1

    # Calculate the number of unique classes
    no_outputs = len(set([instance[-1] for instance in training_set]))
    # Build a new neural network
    network = initialize_neural_net(
        no_inputs, no_hidden_layers, no_nodes_per_hidden_layer, no_outputs)

        network, training_set, learning_rate, no_epochs, no_outputs)
    # Store the class predictions for each test instance
    class_predictions = list()
    for instance in test_set:
        cl_prediction = predict_class(network, instance)

    # Return the learned model as well as the class predictions
    return network, class_predictions


def main():
    The main method of the program
    LEARNING_RATE = 0.1 # Used for stochastic gradient descent procedure
    NO_EPOCHS = 1000 # Epoch is one complete pass through training data
    NO_HIDDEN_LAYERS = 1 # Number of hidden layers
    NO_NODES_PER_HIDDEN_LAYER = 8 # Number of nodes per hidden layer

    # Welcome message
    print("Welcome to the " +  ALGORITHM_NAME + " Program!")

    # Directory where data set is located
    #data_path = input("Enter the path to your input file: ") 
    data_path = "vote.txt"

    # Read the full text file and store records in a Pandas dataframe
    pd_data_set = pd.read_csv(data_path, sep=SEPARATOR)

    # Show functioning of the program
    #trace_runs_file = input("Enter the name of your trace runs file: ") 
    trace_runs_file = "vote_nn_trace_runs.txt"

    ## Open a new file to save trace runs
    outfile_tr = open(trace_runs_file,"w") 

    # Testing statistics
    #test_stats_file = input("Enter the name of your test statistics file: ") 
    test_stats_file = "vote_nn_test_stats.txt"

    ## Open a test_stats_file 
    outfile_ts = open(test_stats_file,"w")

    # Generate a list of the column names 
    column_names = list(pd_data_set) 

    # The input layer in the neural network 
    # will have one node for each attribute value
    no_of_inputs = len(column_names) - 1

    # Make a list of the unique classes
    list_of_unique_classes = pd.unique(pd_data_set["Actual Class"])

    # The output layer in the neural network 
    # will have one node for each class value
    no_of_outputs = len(list_of_unique_classes)

    # Replace all the class values with numbers, starting from 0
    # in the Pandas dataframe.
    for cl in range(0, len(list_of_unique_classes)):
        pd_data_set["Actual Class"].replace(
            list_of_unique_classes[cl], cl ,inplace=True)

    # Normalize the attribute values so that they are all between 0 
    # and 1, inclusive
    normalized_pd_data_set = normalize(pd_data_set)

    # Convert normalized Pandas dataframe into a list
    dataset_as_list = normalized_pd_data_set.values.tolist()

    # Set the seed for random number generator

    # Get a list of 5 stratified folds because we are doing
    # five-fold stratified cross-validation
    fv_folds = get_five_stratified_folds(dataset_as_list)
    # Keep track of the scores for each of the five experiments
    scores = list()
    experiment_counter = 0
    for fold in fv_folds:
        print("Running Experiment " + str(experiment_counter) + " ...")
        outfile_tr.write("Running Experiment " + str(
            experiment_counter) + " ...\n")

        # Get all the folds and store them in the training set
        training_set = list(fv_folds)

        # Four folds make up the training set

        # Combined all the folds so that all we have is a list
        # of training instances
        training_set = sum(training_set, [])
        # Initialize a test set
        test_set = list()
        # For each instance in this test fold
        for instance in fold:
            # Create a copy and store it
            copy_of_instance = list(instance)
        # Get the trained neural network and the predicted values
        # for each test instance
        neural_net, predicted_values = get_test_set_predictions(
            training_set, test_set,LEARNING_RATE,NO_EPOCHS,
        actual_values = [instance[-1] for instance in fold]
        accuracy = calculate_accuracy(actual_values, predicted_values)

        # Print the learned model
        print("Experiment " + str(
            experiment_counter) + " Trained Neural Network")
        for layer in neural_net:
        outfile_tr.write("Experiment " + str(
            experiment_counter) + " Trained Neural Network")
        for layer in neural_net:

        # Print the classifications on the test instances
        print("Experiment " + str(
            experiment_counter) + " Classifications on Test Instances")
        outfile_tr.write("Experiment " + str(
            experiment_counter) + " Classifications on Test Instances")
        test_df = pd.DataFrame(test_set, columns=column_names)

        # Add 2 additional columns to the testing dataframe
        test_df = test_df.reindex(
        ), 'Predicted Class', 'Prediction Correct?'])

        # Add the predicted values to the "Predicted Class" column
        # Indicate if the prediction was correct or not.
        for pre_val_index in range(len(predicted_values)):
            test_df.loc[pre_val_index, "Predicted Class"] = predicted_values[
            if test_df.loc[pre_val_index, "Actual Class"] == test_df.loc[
                pre_val_index, "Predicted Class"]:
                test_df.loc[pre_val_index, "Prediction Correct?"] = "Yes"
                test_df.loc[pre_val_index, "Prediction Correct?"] = "No"

        # Replace all the class values with the name of the class
        for cl in range(0, len(list_of_unique_classes)):
            test_df["Actual Class"].replace(
                cl, list_of_unique_classes[cl] ,inplace=True)
            test_df["Predicted Class"].replace(
                cl, list_of_unique_classes[cl] ,inplace=True)

        # Print out the test data frame

        # Go to the next experiment
        experiment_counter += 1
    print("Experiments Completed.\n")
    outfile_tr.write("Experiments Completed.\n\n")

    # Print the test stats   
    print(ALGORITHM_NAME + " Summary Statistics")
    print("Data Set : " + data_path)
    print("Learning Rate: " + str(LEARNING_RATE))
    print("Number of Epochs: " + str(NO_EPOCHS))
    print("Number of Hidden Layers: " + str(NO_HIDDEN_LAYERS))
    print("Number of Nodes Per Hidden Layer: " + str(
    print("Accuracy Statistics for All 5 Experiments: %s" % scores)
    print("Classification Accuracy: %.3f%%" % (

    outfile_ts.write(ALGORITHM_NAME + " Summary Statistics\n")
    outfile_ts.write("Data Set : " + data_path +"\n\n")
    outfile_ts.write("Learning Rate: " + str(LEARNING_RATE) + "\n")
    outfile_ts.write("Number of Epochs: " + str(NO_EPOCHS) + "\n")
    outfile_ts.write("Number of Hidden Layers: " + str(
        NO_HIDDEN_LAYERS) + "\n")
    outfile_ts.write("Number of Nodes Per Hidden Layer: " + str(
        NO_NODES_PER_HIDDEN_LAYER) + "\n")
        "Accuracy Statistics for All 5 Experiments: %s" % str(scores))
    outfile_ts.write("Classification Accuracy: %.3f%%" % (

    ## Close the files


Return to Table of Contents

Artificial Feedforward Neural Network Trained with Backpropagation Output

Here are the trace runs:

Here are the results:


Here are the test statistics for each data set:


Breast Cancer Data Set

The breast cancer data set results were in line with what I expected. The simpler model, the one with no hidden layers, ended up generating the highest classification accuracy. Classification accuracy was just short of 97%. In other words, the neural network that had no hidden layers successfully classified a patient as either malignant or benign with an almost 97% accuracy.

These results also suggest that the amount of training data has a direct impact on performance. Higher amounts of data (699 instances in this case) can lead to better learning and better classification accuracy on new, unseen instances.

Glass Data Set

The performance of the neural network on the glass data set was the worst out of all of the data sets. The ability of the network to correctly identify the type of glass given the attribute values never exceeded 70%.

I hypothesize that the poor performance on the glass data set is due to the high numbers of classes combined with a relatively smaller data set.

Iris Data Set

Classification accuracy was superb on the iris dataset, attaining a classification accuracy around 97%. The results of the iris dataset were surprising given that the more complicated neural network with two hidden layers and five nodes per hidden layer outperformed the simpler neural network that had no hidden layers. In this case, it appears that the iris dataset benefited from the increasing layers of abstraction provided by a higher number of layers.

Soybean Data Set (small)

Performance on the soybean data set was stellar and was the highest of all of the data sets but also had the largest standard deviation for the classification accuracy. Note that classification accuracy reached a peak of 100% using one hidden layer and eight nodes for the hidden layer. However, when I added an additional hidden layer, classification accuracy dropped to under 70%.

The reason for the high standard deviation of the classification accuracy is unclear, but I hypothesize it has to do with the relatively small number of training instances. Future work would need to be performed with the soybean large dataset available from the UCI Machine Learning Repository to see if these results remain consistent.

The results of the soybean runs suggest that large numbers of relevant attributes can help a machine learning algorithm create more accurate classifications.

Vote Data Set

The vote data set did not yield the stellar performance of the soybean data set, but classification accuracy was still solid at ~96% using one hidden layer and eight nodes per hidden layer. These results are in line with what I expected because voting behavior should provide a powerful predictor of whether a candidate is a Democrat or Republican. I would have been surprised had I observed classification accuracies that were lower since members of Congress tend to vote along party lines on most issues.

Summary and Conclusions

My hypothesis was incorrect. In some cases, simple neural networks with no hidden layers outperformed more complex neural networks with 1+ hidden layers. However, in other cases, more complex neural networks with multiple hidden layers outperformed the network with no hidden layers. The reason why some data is more amenable to networks with hidden layers instead of without hidden layers is unclear.

Other conclusions include the following:

  • Higher amounts of data can lead to better learning and better classification accuracy on new, unseen instances.
  • Large numbers of relevant attributes can help a neural network create more accurate classifications.
  • Neural networks are powerful and can achieve excellent results on both binary and multi-class classification problems.

Return to Table of Contents


Alpaydin, E. (2014). Introduction to Machine Learning. Cambridge, Massachusetts: The MIT Press.

Fisher, R. (1988, July 01). Iris Data Set. Retrieved from Machine Learning Repository:

German, B. (1987, September 1). Glass Identification Data Set. Retrieved from UCI Machine Learning Repository:

Ĭordanov, I., & Jain, L. C. (2013). Innovations in Intelligent Machines –3 : Contemporary Achievements in Intelligent Systems. Berlin: Springer.

Michalski, R. (1980). Learning by being told and learning from examples: an experimental comparison of the two methodes of knowledge acquisition in the context of developing an expert system for soybean disease diagnosis. International Journal of Policy Analysis and Information Systems, 4(2), 125-161.

Montavon, G. O. (2012). Neural Networks : Tricks of the Trade. New York: Springer.

Schlimmer, J. (1987, 04 27). Congressional Voting Records Data Set. Retrieved from Machine Learning Repository:

Wolberg, W. (1992, 07 15). Breast Cancer Wisconsin (Original) Data Set. Retrieved from Machine Learning Repository:

Return to Table of Contents

Artificial General Intelligence and Deep Learning’s Long-Run Viability

Learning and true intelligence are more than classification and regression (which is what deep learning is really good at). Deep learning is deficient in perhaps one of the most important types of intelligence, emotional intelligence. Deep learning cannot negotiate trade agreements between countries, resolve conflict between a couple going through a divorce, or craft legislation to reduce student debt.

The long-run viability of deep learning as well as progress towards machines that are more intelligent will depend both on improvements in computing power as well as our ability to understand and subsequently quantify general intelligence. 

Concerning the quantification of general intelligence, this really is the inherent goal of not just artificial intelligence, but of computer science as a whole. We want a machine that can automate tasks that humans do and to be as independent as possible while doing it. We want to teach a computer to think the way a human thinks, to understand the way a human understands, and to infer the way a human infers. At present, the state-of-the-art is a long way from achieving this goal. 

Deep learning is only as good as the data fed to it: garbage in, garbage out. Deep learning has a lot of limitations and is a long way from the kind of artificial general intelligence envisioned in those science fiction movies.

At this point, the state of the art in deep learning is identifying images, detecting fraud, processing audio, and performing advertising. The only thing that we really know for sure about deep learning at this point is that it performs some tasks at a very high level, and we aren’t 100% sure why it does. We understand the probability and the algorithm behind the representations that we are making, but we don’t understand what the network is actually learning like we would understand a classification tree or a run of Naive Bayes.  

Understanding what abstractions that the network is learning is essential to move forward in the field. If at some point the theory is proven to be invalid, then all the emerging techniques will be rendered useless.

In any case, even if some theoretical basis is proven, it may or may not show a strong relationship that warrants continued development in these fields.

Deep Learning Needs Generalists Not Just Specialists

The Renaissance Man, Leonardo da Vinci

My prediction is that the real breakthroughs on the route to artificial general intelligence will come from the T-shaped generalists, those professionals who have the breadth and depth to see the unseen interconnectedness. 

In order for deep learning and AI to be viable over the long run, researchers in this field will need to be able to think outside the box and be well-versed in the fundamentals of multiple disciplines, not just their own silo. Mathematicians need to learn a little bit about computer science, neuroscience, and psychology. Computer scientists need to learn a little bit about neuroscience and psychology. Neuroscientists need to cross over and learn a little bit about computer science as well as mathematics. 

While learning about other fields is no guarantee that a researcher will make a breakthrough, staying in a narrow area and not venturing out is a guarantee of not making a breakthrough. As they say, to the hammer, everything looks like a nail.

As a final note, I’m keeping my eye out on quantum computing, and, more importantly, quantum machine learning and quantum neural networks. Quantum machine learning entails executing machine learning algorithms on a quantum computer. It will be interesting to see how getting away from the 1s and 0s of a standard computer will impact the field, and if it will usher in a new era of artificial intelligence. Only time will tell.

The Limitations of Deep Learning

Despite the explosion in the popularity of deep learning, we are still a long way from the kind of artificial general intelligence that is the inherent long term goal of computer science. Let us take a look at what the limitations are of deep learning. I’m going to pull from a paper written by Professor Gary Marcus of New York University about this topic. The paper is entitled: Deep Learning: A Critical Appraisal.

Deep Learning Methods Require Lots and Lots of Data


Humans do not need hundreds, thousands, or even millions of examples of horseshoe crabs in order to learn what a horseshoe crab is. In fact, a 3 year old child could look at one example of a horseshoe crab, be told it is a horseshoe crab, and immediately identify other horseshoe crabs, even if those new horseshoe crabs do not look exactly like that first horseshoe crab example. 

Deep learning models on the other hand, need truck loads of examples of horseshoe crabs in order to distinguish horseshoe crabs from say spiders, or other similar-looking creatures.

Deep Learning Is Not Deep


The word “deep” in deep learning just refers to the structure of the mathematical model built during the training phase of the algorithm. The learning is not really deep, in the true sense of the word. A deep learning algorithm does not understand the why behind its output. 

Deep Learning Does Not Quickly Adapt When Things Change


Deep learning works well when test data looks like training data. However, it cannot easily cope with novel situations, such as when the domain changes from the one that it was trained on.

For example, if a deep neural network learns that school buses are always yellow, but, all of a sudden, school buses become blue, deep learning models will need to be retrained. A five year old would have no problem recognizing the vehicle as a blue school bus.

Deep Learning Does Not Do Well With Inference


Deep learning algorithms cannot easily tell the difference between phrases like: “John promised Mary to leave” and “John promised to leave Mary.”

Deep Learning Is Not Transparent


Part of the confusion and lack of clarity of deep learning is that it is difficult to understand for the general public. For example, there are many formulas and recalculations that are completed in order to create the neural network that forms the learning model for our deep learning program. 

These neural networks could also be biased depending on how the programmer sets up the formulas…for example, weighting some features more than others instead of considering the collaborative effect of all the features.

It also does not help that the layers of the neural network are called “hidden layers” which makes deep learning and its methodologies sound even more mysterious.  

Also, the mathematics….ahh the mathematics. Try explaining the whole sigmoid and gradient descent thing to a child. Better yet, imagine a doctor explaining to a patient suffering from cancer that the cancer diagnosis was determined based on the output of an artificial feedforward neural network trained with backpropagation. How would a patient respond to this? How would a health insurance company respond?

Neural networks are relatively black boxes that look like magic to the untrained eye, containing hundreds if not millions of parameters. In fields like medicine and finance, humans want to know exactly (and simply) why a particular decision was made. You cannot just say, “because my deep neural network said so.” 

Deep Learning Cannot Take Full Advantage of the Five Basic Senses

The Five Senses: Look, Sound, Smell, Taste, Feel

Humans have five basic senses: vision, hearing, smell, taste, touch. When I learn what a rooster is, I can see it, hear it, smell it, taste it, and touch it. Deep learning at this stage focuses mainly on the vision part and lacks the other four senses. Those other senses can be important.

For example, imagine a driverless car. A human could hear a train coming long before it sees the train. The driver would then stop. A driverless car on the other hand would need to see the train first before making the decision to stop.

Humans use all five senses to learn, and these five senses play an important role in getting a bigger picture of recognizing objects and understanding what makes a rooster a rooster, or a train, a train.

Deep Learning Is Not yet Able to Know Something With 100% Certainty


I could look at a trash can inside my kitchen, know it is a trash can, and tell you with 100% certainty that it is a trash can. A deep neural network on the other hand, no matter how many parameters or data it has been trained on, works in the world of probabilities and numbers. It will never have 100% confidence that the trash can is a trash can. 

A trained deep neural network, for example, might be 96.3% sure it is a trash can but not 100% sure. There is always that miniscule probability it could be something else, like say a tree stump or a recycle bin. 

I’m in front of my laptop now. It is 100% my laptop. A deep learning algorithm, on the other hand, might say it is 99.3% sure that the object currently in front of me is my laptop. It would need a human to validate that the object is, in fact, a laptop.

This limitation this has to do with the way a neural network algorithm “learns.” 

The reality is that a program does not look at a dog and intuitively know that it is a dog like a human does. Instead, deep learning is more like a statistical analysis of patterns that are observed from the sample data points. So a deep learning program might only identify that an image is a dog based on the shapes in the image and the fact that statistical patterns tell you that these shapes will mean that the image is likely a dog.  

But the fact is that the deep learning program cannot be 100% sure that an image is a dog; for example if there is a weird picture of a fox or bear that looks a lot like a dog. Due to this uncertainty (the fact that the deep learning algorithm cannot be 100% sure), it is difficult for humans to trust deep learning especially when applied to critical, possibly life-threatening applications.  

For example, if a deep learning algorithm cannot always correctly identify that an object is an obstacle when processing the images from a driverless car, then it would be concerning for people to trust this program to drive for us.

Furthermore, there may be an inherent bias in how the deep learning algorithm identifies patterns. For example, if a programmer decided that a road is only definable by separating lines for lanes, then the algorithm may be biased to using lines for identifying roads. This would mean that the algorithm misses identifying roads that don’t have clear lane markings or even dirt roads. Such a bias means that a deep learning program could miss important identifying features/patterns. 

Is this a road or a walking path to a beach? Humans would easily be able to recognize this as a road due to the tire tracks. Computers in driverless cars would need to have seen something similar to this in order to recognize it as a road.

In short, it is difficult for a deep learning algorithm to account for every possible example and every variation, which means that it would not be 100% correct.     

Significance of the Misunderstanding of Deep Learning

In my previous post on deep learning, I posted about some of the potential fallout that could occur due to the misunderstanding of what deep learning is really learning. I mentioned some possible significant consequences:

  • Wasted resources as venture capitalists throw money at anything that has to do with deep learning. 
  • Wasted resources as non-expert government agencies fund any research project that has the term “deep learning” in it.
  • A boat load of computer science graduates around the world that, all of a sudden, have found their “passion” in deep learning.
  • Disappointed companies as deep learning does not have the expected impact on their bottom line.
  • Another AI winter.

Let’s look at the last bullet point. Rodney Brooks, former MIT professor and co-founder of iRobot and Rethink Robotics, predicts that we will enter a new AI winter in 2020

AI winter is a period of reduced funding and interest in artificial intelligence research that comes at the tail end of an AI hype cycle. Each AI hype cycle begins with some major breakthrough. Then for the next 5-10 years after that breakthrough, all sorts of papers get written on AI, companies that are doing “X + [insert some new hot, AI technology]” get funded, computer science students around the world change their career paths, and the media goes into a feeding frenzy about how the new breakthrough will change the world.

Executives at big companies around the world then shout out quotations like this: “AI is more profound than … electricity or fire” –  Sundar Pichai (CEO of Google). 

Experts in AI then chime in, “This time is different!” 

When you hear comments like this, ask yourself “what are they selling?”.

Deep learning is a tool like any other tool…like a wrench for a car mechanic or a serrated knife for a master chef. Deep learning can currently solve specific problems really well but others not so well. 

Machine learning, the field that encompasses deep learning, is about automating the process of finding relationships based on empirical data. It is a powerful tool that has an enormous amount of potential, but it is not a panacea and is still a long way away from replacing the human brain. 

I do agree with Mr. Pichai that when we have true artificial general intelligence that such a breakthrough would be as profound as electricity or fire. We are not there yet. Much more work needs to be done (and that is a great thing for us scientists and engineers). The future is bright.

How Can We Help Others Gain a Better Understanding of What These Models Are Learning?

The example that is often marketed to explain deep learning is a neural network that first takes the inputs and learns lines, curves, and other shapes. Each successive layer abstracts and combines the data more and more until we see letters and fully formed images. This method of explaining deep learning seems like a good example of how we might gain an understanding of what exactly these algorithms actually learn.

I think that one reason why people might not trust deep learning is because they don’t understand how it works and even when they do, we cannot see the hidden layers in the neural network. When we look at the neural network for deep learning, we have multiple layers including hidden layers that are compose some of the neurons.

Neural Network

With deep learning, we allow the program to learn and distinguish the key features from our sample data set. The problem is that we may not be easily able to understand the features that are distinguished and how they might be related to each other as defined by the hidden layers in the algorithm’s learning model.

I think that one strategy for improving, or at least understanding what deep learning is doing is to unpack these abstracted layers in order to hand tune the results into something that is more relevant.

What are Deep Learning Methods Really Learning?

It is not exactly clear what deep learning methods are really learning. Sure, they are highly effective and are learning something, but I’m still trying to get my head around exactly what they are learning.

Consider your run-of-the-mill deep neural network. “Learning” is nothing more than an optimization procedure. We are trying to produce an optimized mathematical formula that takes in a set of training examples and then can, as accurately as possible, map the inputs (i.e. attributes, features, etc.) of those examples to the outputs (i.e. class, target variable, etc.). We then use this formula to classify a new set of examples.

Gradient Descent. Is this really all there is to learning?

At its core, deep learning is about input-process-output. It is not true learning in the sense of the word (the way we humans do). True learning entails understanding, and understanding is nonexistent during deep learning. 

You can memorize a book, chapter-by-chapter, word-for-word; but that doesn’t mean you are learning. You still would not understand the plot. Similarly, in deep learning there is no understanding. Deep learning “memorizes” a mapping between inputs and outputs without any real understanding of the why behind those relationships. And in my view, the why is a huge part of learning. True learning (in the human sense of the word) without understanding is not learning. Perhaps then we should call deep learning something different? Deep optimization perhaps??? Guess that didn’t sound as marketable and sexy as deep learning.

If you look out in nature — the human brain or the brain of any living organism — nothing out there learns in a way that even remotely resembles backpropagation. Neural networks are about classification error, but real learning — the way humans learn — is deeper than that (pun intended). 

A neural network, for example, has a completely different concept of what it is to be a dog. That concept could involve where certain groups of pixels are placed and may have nothing to do with the actual structure of the animal. Where a human would see legs, arms, torso, etc., a deep learning algorithm may abstract a completely different set of things. This has led a rise in adversarial attacks, where an attacker is able to determine what representation provides the highest probability for an image to be classified as anything and is able to insert noise that causes things to be misclassified.

Another point to consider is that neural networks generate something. It may be a relationship that we did not previously understand, but it may also just be nonsense that happens to work.  The abstractions may result in some representational form that is at its most basic just complete nonsense. If there is no real understanding of what the abstractions that the algorithm makes, then it is hard to confirm that it’s actually doing anything.

A Basic Neural Network

The significance of the lack of understanding of what deep learning is, is yet to be seen; but here are just a few of the consequences if the hype gets unchecked:

  • Wasted resources as venture capitalists throw money at anything that has to do with deep learning. 
  • Wasted resources as non-expert government agencies fund any research project that has the term “deep learning” in it.
  • A boat load of computer science graduates around the world that, all of a sudden, have found their “passion” in deep learning.
  • Disappointed companies as deep learning does not have the expected impact on their bottom line.
  • Another AI winter.

Remember, there is the marketing element in there too. Using anthropomorphic terms like machine “learning” and deep “learning” is a much better sell to a general audience than machine mathematical optimization or deep optimization. Researchers gotta sell their ideas too!

Bottom Line: Artificial intelligence is not yet intelligent, and deep learning is not yet deep (yay! we still have work to do!)…nor is it learning in the true sense of the word. Deep learning certainly will continue to have an enormous impact on the world, but there needs to be more awareness and discussion of not just the enormous potential of deep learning but also its limitations so non-technical stakeholders can make more informed decisions.

Why Deep Learning Has Received So Much Attention Lately

Deep learning has been receiving an enormous amount of interest over the last seven years in the academic and business communities. Let’s take a look at the definition of deep learning, and then we will take a look at how this field has become so popular so quickly.

What is Deep Learning

Deep learning is a machine learning technique in which we teach a computer how to make predictions. Predictions are made by mapping a set of inputs to a set of outputs. 

Input Data —–> Deep Learning Algorithm (i.e. Process) —–> Output Data

For example, let’s say our input data into a deep learning algorithm is a set of photos. We want to be able to automatically tag each photo as either being dogs or elephants.


Input Data (lots of images containing dogs and elephants) —–> Deep Learning Algorithm —–> Classification of Each Image (i.e. Dogs or Elephants)

The “learning” part of the term deep learning entails looking at a bunch (hundreds, thousands, even millions+) of photos of elephants and dogs to develop a mathematical model of what both animals look like. Once the deep learning algorithm has been trained to recognize dogs and elephants, it can then be used to classify new photos as either dogs or elephants.

Most deep learning algorithms use neural network architectures as the structure of the underlying mathematical model. For this reason, deep learning methods are commonly called deep neural networks. 

Neural networks consist of layers and interconnected nodes. The first layer is the input layer. This layer might consist of, for example, thousands of matrices of pixels that represent photos of dogs or elephants. Each layer after the input layer transforms the data slightly so that the data is more abstract and complete than the previous layer. 

The layer after the input layer (i.e. second layer), for example, might contain nodes that recognize simple shapes like circles and edges (that at this point look nothing like a dog or elephant). The third layer contains nodes that recognize more complex shapes that look like a dog’s body parts (e.g. nose, eye, ear, etc.). Then the final layer, the output layer, outputs the classification of a photo as being either a dog or elephant.

A basic multi-layer neural network architecture. The first layer on the left is the input layer. The two inner layers of nodes (neurons) are the hidden layers. The fourth layer on the right is the output layer that outputs the classification. In this case, the network expects four different classes in the data set (e.g. dogs, elephants, cows, horses).

Forbes Magazine has a good image showing the basic deep neural network structure I described above.

The “deep” part of deep learning refers to the number of hidden layers in the neural network. Standard neural networks have two or three (like in my example above) hidden layers, but deep neural networks can have 100+ layers. 

In short, a deep neural network is one that has several hidden layers, with the idea that these layers learn different levels of abstraction of the input attributes; thereby allowing the network to solve more complex problems, such as face recognition, object tracking and so on.

Origin of the Deep Learning Revolution: AlexNet

This post at shows the graphs of the percentage of selected arXiv publications with either “deep”, “adversarial” or “convolutional” in the title. Note how the graph was virtually all 0s prior to 2010. It then took off like a rocket in 2012. What happened in 2012?

In 2010 and 2011, Fei-Fei Li held the ImageNet competition, an annual machine learning contest. Contest participants were given millions of images to use to train their models. These images were pre-labeled with one of ~1,000 different categories (e.g. leopard, cherry, mushroom, etc.). The objective of the contest was to correctly classify examples that were not in the training set. 

During those first two years of the competition in 2010 and 2011, the winning teams had a classification accuracy of 72%. None of the winners of those competitions used deep learning methods. Then in 2012, a team from the University of Toronto led by Alex Krizhevsky won the competition with a classification accuracy of 84%. The second place contestant had a classification accuracy of 74%. The team from the University of Toronto used deep learning methods combined with the computational power of graphical processing units (GPUs) to completely blow the competition out of the water.

The results were remarkable and gave birth to the deep learning era that continues to this day.

Why Deep Learning Has Received So Much Attention Lately

With traditional machine learning approaches, you would have to design a feature extraction algorithm which generally involves a lot of heavy mathematics (complex design), may not be very efficient, and does not perform well (i.e. accuracy may not be suitable for real-world applications). After doing all of that, you would also have to design a whole classification model to classify your input given the extracted features (i.e. attributes).

That’s a lot of work!

Enter Deep Learning…

  • With deep neural networks, we can perform feature extraction and classification in one shot, which means we could only need to design one model.
  • The availability of large amounts of labeled data as well as GPUs, which can process data in parallel at high speeds, enables these models to be much faster than previous methods.
  • Using the back-propagation algorithm, a well-designed loss function, and millions of parameters, these deep networks are able to learn highly complex features (which had to traditionally be hand designed)…i.e. no more complex design!
  • Deep neural networks have become fairly easy to implement with high-level open source libraries such as Keras, Pytorch, and TensorFlow.

Deep Learning has made many new applications practically feasible. We wouldn’t have been able to make good language translators pre-deep learning, because we simply had no technique at the time that would perform well enough or at a high enough speed for a real-world application.  Deep learning techniques have been applied to not just image recognition, but automatic speech recognition, natural language processing, drug discovery, customer relationship management, robotics, self-driving cars, and more.

How to Choose an Optimal Learning Rate for Gradient Descent

One of the challenges of gradient descent is choosing the optimal value for the learning rate, eta (η). The learning rate is perhaps the most important hyperparameter (i.e. the parameters that need to be chosen by the programmer before executing a machine learning program) that needs to be tuned (Goodfellow 2016).

If you choose a learning rate that is too small, the gradient descent algorithm might take a really long time to find the minimum value of the error function. This defeats the purpose of gradient descent, which was to use a computationally efficient method for finding the optimal solution.

On the other hand, if you choose a learning rate that is too large, you might overshoot the minimum value of the error function, and may even never reach the optimal solution. 

Before we get into how to choose an optimal learning rate, it should be emphasized that there is no value of the learning rate that will guarantee convergence to the minimum value of the error function (assuming global) value of a function. 

Algorithms like logistic regression are based on gradient descent and are therefore what is known as “hill climber.” Therefore, for non-convex problems (where the graph of the error function contains local minima and a global minimum vs. convex problems where global minimum = local minimum), they will most likely get stuck in a local minimum. 


So with that background, we are ready to take a look at some options for selecting the optimal learning rate (eta) for gradient descent.

Choose a Fixed Learning Rate

The standard gradient descent procedure uses a fixed learning rate (e.g. 0.01) that is determined by trial and error. For example:

“Typical values for a neural network with standardized inputs (or inputs mapped to the (0,1) interval) are less than 1 and greater than 10-6 but these should not be taken as strict ranges and greatly depend on the parametrization of the model. A default value of 0.01 typically works for standard multi-layer neural networks but it would be foolish to rely exclusively on this default value. If there is only time to optimize one hyper-parameter and one uses stochastic gradient descent, then this is the hyper-parameter that is worth tuning.”

Bengio (2013)

Use Learning Rate Annealing

Learning rate annealing entails starting with a high learning rate and then gradually reducing the learning rate linearly during training. The learning rate can decrease to a value close to 0.

The idea behind this method is to quickly descend to a range of acceptable weights, and then do a deeper dive within this acceptable range.

Use Cyclical Learning Rates

Cyclical learning rates have been proposed: 

“Instead of monotonically decreasing the learning rate, this method lets the learning rate cyclically vary between reasonable boundary values.”

Leslie (2017)

Use an Adaptive Learning Rate

Another option is to use a learning rate that adapts based on the error output of the model. Here is what the experts say about adaptive learning rates.

Reed (1999) notes on page 72 of his book Neural Smithing: Supervised Learning in Feedforward Artificial Neural Networks that an adaptable learning rate is preferred over a fixed learning rate:

“The point is that, in general, it is not possible to calculate the best learning rate a priori. The same learning rate may not even be appropriate in all parts of the network. The general fact that the error is more sensitive to changes in some weights than in others makes it useful to assign different learning rates to each weight”

Reed (1999)

Melin et al. (2010) writes:

“First, the initial network output and error are calculated. At each epoch new weights and biases are calculated using the current learning rate. New outputs and errors are then calculated.

As with momentum, if the new error exceeds the old error by more than a predefined ratio (typically 1.04), the new weights and biases are discarded. In addition, the learning rate is decreased. Otherwise, the new weights, etc. are kept. If the new error is less than the old error, the learning rate is increased (typically by multiplying by 1.05).”

Melin et al. (2010)


Bengio, Y. Practical Recommendations for Gradient-based Training of Deep Architectures. In K.-R. Muller, G. Montavon, and G. B. Orr, editors, Neural Networks: Tricks of the Trade. Springer, 2013.

Goodfellow, Ian, Yoshua Bengio, and Aaron Courville. Deep Learning. Cambridge, Massachusetts: The MIT Press, 2016. Print.

Melin, Patricia, Janusz Kacprzyk, and Witold Pedrycz. Soft Computing for Recognition Based on Biometrics. Berlin Heidelberg: Springer, 2010. Print.

Reed, Russell D., and Robert J. Marks. Neural Smithing: Supervised Learning in Feedforward Artificial Neural Networks. Cambridge, Mass: MIT Press, 1999. Print.

Smith, Leslie N. Cyclical learning rates for training neural networks. In Applications of Computer Vision (WACV), 2017 IEEE Winter Conference on. IEEE, pages 464–472.

My Master’s Thesis

Straight from the vault, my Master’s thesis: An Analysis of the Robustness and Relevance of Meteorological Triggers for Catastrophe Bonds.


Each year weather-related catastrophes account for an estimated United States dollars (USO) $40 billion in damage across the world, although only a fraction of this risk of loss is insured. Losses from hurricanes in the United States have increased over the past several years to the extent that many insurance companies have become increasingly reluctant to insure in certain locations along the coast. Several insurance companies have become insolvent as a result of the active hurricane seasons of 2004 and 2005. In order to cope with this hurricane risk, some insurance and reinsurance firms have shifted part of their risk to the capital markets in the form of catastrophe bonds.

Two problems are observed with catastrophe bonds based on parametric triggers (e.g. Saffir-Simpson scale rating of a hurricane at landfall). First, the trigger mechanisms are measured imprecisely, with the degree of imprecision depending on the choice of trigger mechanism, the available sensor systems, and the methods by which meteorologists analyze the resulting observations. Second, the trigger mechanisms might not relate well to the economic harm caused by the weather phenomena, suggesting that they were not selected on the basis of adequate understanding of relevant meteorology and its relationship to storm damage. Both problems are documented, and perhaps ameliorated in part, by a thorough study of the relevant meteorology and meteorological practices.

Development of a set of robust and relevant triggers for catastrophe bonds for hurricanes is the objective of this study. The real-time and post-landfall accuracy of measured hurricane parameters such as minimum central pressure and maximum sustained surface wind speed were analyzed. Linear regression and neural networks were then employed in order to determine the predictability of storm damage from these measurable hurricane parameters or combination thereof. The damage data set consisted of normalized economic losses for hurricane landfalls along the United States Gulf and Atlantic coasts from 1900 to 2005. The results reveal that single hurricane parameters and combinations of hurricane parameters can be poor indicators of the amount of storm damage.

The results suggest that modeled-loss type catastrophe bonds may be a potentially superior alternative to parametric-type bonds, which are highly sensitive to the accuracy of the measurements of the underlying storm parameters and to the coastal bathymetry, topography, and economic exposure. A procedure for determining the robustness of a risk model for use in modeled-loss type catastrophe bonds is also presented.