Pair Programming or Bare(ly) Programming

image

“Sorry we don’t have enough resources, we only have four pairs” – As an engineering leader no other statement has made me cringe more.  After all four pairs is a healthy sized team of eight developers. 

Throughout my career I have run across CTOs, VPs, directors, development managers, teams, and individual developers who swear by pair programming with near religious devotion.   Personally I’ve maintained a healthy dose of skepticism when it comes to pairing as an overarching development philosophy.  

As an engineering leader my job is to build products that delight customers in the most efficient way possible.   Anecdotally, pairing consistently costs more and hence seems irresponsible to use exclusively as a development technique.    But admittedly anecdotal evidence is insufficient so I decided to dig through the research and see if I could find more empirical evidence to support my claim.

Background

Pair programming is an agile software development methodology where two programmers work on the same task using one computer and keyboard.   One programmer is called the driver and operates the keyboard and does the primary coding work.   The other developer, often called the navigator, is responsible for observing the driver and providing guidance in order to speed up problem solving, improve design, and minimize defects.

The potential negative impact of pair programming is immediately clear to most people.   By applying two resources to a task you are effectively doubling the cost.  So unless there’s an equal or greater improvement in other project variables, pair programming would be nearly impossible to justify. Exploring the problem through a project management lens, where we have three variables, cost (including resources), time, and quality/scope, If we double our cost we’d expect to see an equivalent decrease in time to deliver or increase in quality or scope (or some factor of each).

image

In mathematical terms let’s assume the value of any given project X is equal to a weighted linear combination of cost, time and quality/scope. 

image

When pairing our cost is automatically going to double since we’ve applied two resources for a task that in theory can be completed by one.

image

In order for our project value to remain equal or be better we need our other variables to proportionally change in the right direction.   For example if our project now takes 50% less time we could argue we net out even.  Or if our scope or quality double, we would similarly be in a good position.

image

However, In my experience I’ve not seen pair programming live up to these expectations.  Instead I’ve seen tasks or user stories take the same amount of time and produce similar results at nearly double the cost.  But you shouldn’t take my word for it.  Let’s review the literature and see what the experts have to say.

Research

There are actually a fair number of research papers that attempt to prove or disprove the efficacy of pair programming.  That said, in my survey of the literature I found most of the research to be ill designed for comparison to real world corporate product development organizations.  Specific issues include:

  • Developer Skills:  Most of the studies rely on university students that shouldn’t be compared to seasoned professional developers.
  • Non Production Environments:  The majority of the software used for evaluation is very far removed from real product development environments.
  • Organization Realities: Finally there is little or no accounting for organizational churn that happens in a real for-profit company
  • In spite of these issues it’s worth exploring these various research studies and the insights they provide on the impacts of pair programming.

    Many of the research papers evaluate the impact of pair programming on effort, which in at least one paper is defined as two times the duration or time required to complete a given task [1].  Specifically, effort increases ranging from 15% all the way to 100% have been observed [2].  In one of the more well conducted studies an effort increase of 84% was seen [1].   Since we know effort is just twice the duration of a single developer we can actually do some math to figure out how much faster pairs complete a task versus a single developer.

    image

    Or by using our earlier project management equation, with a little rounding we can assume our pairing time weight would be roughly 9/10 the weight required for a single developer.

    image

    This is nowhere near the factor of 1/2 or less we said we needed to make pair programming cost efficient.  Well if the research doesn’t support a sufficient decrease in time to completion perhaps there’s research indicating that a given project’s scope or quality will increase enough to offset the difference.  

    Unfortunately, once again the results are at best inconclusive, but in many cases support an actual decrease in scope and minimal or near zero increase in quality.  For example in [2] a reported 29% decrease in productivity was measured for pair programming team when measured as a function of completed use cases.

    Regarding quality, even in one of the more optimistic papers we only saw a 10% – 20% increase in quality (measured as test cases passed) [3].   According to [2], we only saw an 8% improvement in quality when measuring actual defects.   While these improvements are non trivial, when combined with the time and scope metrics it remains insufficient to offset the associated costs.

    Cherry Picking

    “But aren’t you just cherry picking the worst examples to justify your case” you might ask? Not really because even in the most optimistic research studies initial results were usually much worse and only improved over time.  For example in [3] initial increases in effort dropped from 60% to15% over time. Most of the research attributes these gains in effort to “pair jelling”.  In other words, as the pairs get to known each other they become more efficient.

    The problem with these studies is that they assume that once a pair jells the gain will hold.  However in any real for-profit organization there is potential for high variability in projects and staff which means pair jelling is unlikely to be a one off cost.  It is more likely a continuing cost to the business over time.

    Several studies also point out that the value of pair programming decreases with simpler tasks [4].   Therefor one must consider the ratio of simple to complex tasks in any given development cycle in order to understand the long term impacts of pair programming.  When I evaluated my own teams, I found multiple iterations where 75% of work items where smaller changes that could easily be tackled by a single developer in the same timeframe.  

    Finally, one paper [5] attempted to justify pair programming by evaluating Net Present Value (NPV).   In this paper an argument is made that even if it costs more to pair program, faster time to market warrants the cost.  I take issue with this calculation since it does not factor in the opportunity cost of having those extra resources not work on a different higher priority project.  

    For example if we take the reported 84% increase in effort and assume we finish our project in 9/10 the time of a single developer, we must ask ourselves what happens when a key customer asks for a critical bug fix?   I can tell that customer to wait until I finish my current project or I can split my pair and work on both at the same time at the small cost of a 1/10 increase in duration.  By splitting my pair I’ve delighted my key customer as quickly as possible at a trivial cost. Clearly you need to factor in the opportunity cost of not delighting that customer when evaluating the value of pair programming.

    To Pair or Not to Pair

    So should you pair or not pair?  There are a lot of reasons a team might use pair programming.  In some cases the cost / benefit tradeoff may be worthwhile.  Pairing can be very effective at educating new team members, improving the skills of junior team members, cross training, and reducing the cost of complex tasks.  If you take anything away from this post let it be:

  • Challenge the Efficacy of Pair Programming: If your team or engineering manager wants to exclusively use pair programming, don’t blindly accept it.  Collect the data to validate if it is really cost effective
  • Pair when it makes Sense:  Use pairing selectively when it makes sense including educating new team members, improving the skills of junior team members, cross training, and reducing the cost of complex tasks.
  • Factor in Opportunity Costs: Make sure you consider the opportunity costs of projects not being worked on when pairing.
  • In short don’t allow yourself to be swayed by a dogmatic insistence that pair programming is better.  As a leader your job is to challenge your team to delight customers in the most cost effective way possible.   Pairing should only be used if it definitively contributes to that cause.

    References

    [1] Arisholm, Erik, et al. “Evaluating pair programming with respect to system complexity and programmer expertise.” IEEE Transactions on Software Engineering 33.2 (2007). – Summary available at https://pdfs.semanticscholar.org/9787/c9663cad3a1c21550f2e5e365e70fd01d3aa.pdf

    [2] Vanhanen, Jari, and Casper Lassenius. “Effects of pair programming at the development team level: an experiment.” Empirical Software Engineering, 2005. 2005 International Symposium on. IEEE, 2005. https://pdfs.semanticscholar.org/40dd/fa666bf367cfffaae421dbd3c6170a3e3dc3.pdf

    [3] Cockburn, Alistair, and Laurie Williams. “The costs and benefits of pair programming.” Extreme programming examined (2000): 223-247. http://www.cs.pomona.edu/~markk/cs121.f07/supp/williams_prpgm.pdf

    [4] Lui, Kim, and Keith Chan. “When does a pair outperform two individuals?.” Extreme programming and agile processes in software engineering (2003): 1011-1011. ftp://nozdr.ru/biblio/kolxo3/Cs/CsLn/E/Extreme%20Programming%20and%20Agile%20Processes%20in%20Software%20Engineering,%204%20conf.,%20XP%202003(LNCS2675,%20Springer,%202003)(ISBN%203540402152)(479s)_CsLn_.pdf#page=240

    [5] Padberg, Frank, and Matthias M. Muller. “Analyzing the cost and benefit of pair programming.” Software Metrics Symposium, 2003. Proceedings. Ninth International. IEEE, 2003. http://wwwipd.ira.uka.de/Tichy/uploads/publikationen/32/metrics03.pdf

    End-to-End Speech Recognition: Part 1 – Neural Networks for Executives (I Mean Dummies)

    When I originally contemplated the subject of my next blog post, I thought it might be interesting to provide a thorough explanation of the latest and greatest speech recognition algorithms, often referred to as End-to-End Speech Recognition, Deep Speech, or Connectionist Temporal Classification (CTC).   However, as I began to research the topic I quickly discovered that my basic knowledge of neural networks was woefully lacking.  Several weeks of reading and a few hundred lines of code later, I realized before I could teach a fellow plebe like myself about end-to-end speech recognition,  I probably needed to introduce the fundamentals first.

    With that in mind, what was intended to be a single entry will likely turn into multiple blog posts covering an overview of end-to-end speech recognition and some fundamentals of deep learning that make it possible.  In this first post I’d like to provide a brief introduction to end-to-end speech recognition and then give a more detailed tutorial about one of the basic components of deep learning, a multilayer perceptron, also known as a feed forward neural network.  I’ll then walk you through how I brought all this information together while building a very basic end-to-end speech recognition system.

    End-to-End Speech Recognition

    So what is end-to-end speech recognition anyway?  At it’s most basic level an end-to-end speech recognition solution aims to train a machine to convert speech to text by directly piping raw audio input with associated labeled text through a deep learning algorithm.   The resulting model is then able to recognize speech with no further algorithmic components.

    image

    And why is this any better than traditional speech recognition systems?  Traditional speech recognition systems use a much more complicated architecture that includes feature generation, acoustic modeling, language modeling, and a variety of other algorithmic techniques in order to be accurate and effective.   This in turn makes the training, testing, and code complexity far more difficult than would be with an end-to-end system.

    image

    In other words an end-to-end solution greatly reduces the complexity in building a speech recognition system.   And if that alone doesn’t convince you of the value an end-to-end recognizer brings to the table, several research teams, most notably the folks at Baidu, have shown that they can achieve superior accuracy results over traditional speech recognition systems.

    To validate the possibilities of an end-to-end speech recognition system I decided to build my own.  However, I quickly found that building such a system required advanced knowledge of deep learning techniques.   This is because the current end-to-end systems generally rely on more complex neural network algorithms like Recurrent Neural Networks (RNNs) and something called the connectionist temporal loss function that are difficult to understand if you don’t have a solid understanding of basic neural networks.   So I opted to take a simpler approach and see if I could build a very simple end-to-end recognizer using basic deep learning techniques.   Specifically a feed forward neural network or multi layer perceptron.

    Neural Network Fundamentals

    Before I dive into the details, let me provide a quick tutorial on the feed forward neural network.  The underlying element of a neural network is called a perceptron or an artificial neuron.  Much like a biological neuron, a perceptron takes a series of inputs, performs a function on those inputs, and produces and output that can be passed to other neurons.

    image

    The simplest function is just a sum of weighted inputs.  However this function is a linear relationship and the world is rarely linear so we apply something called an activation function to help impart nonlinearity.   There are actually numerous activation functions used in neural networks, some linear and some not, but the Sigmoid and TanH functions are two you will commonly see in the relevant literature.

    image

    Now that we know what a neuron is, a neural network is really just a collection of multiple interconnected neurons.   Neurons are grouped and connected in “layers”.   The simplest neural network is a single layer network that connects one or more inputs to one or more outputs.   There is no calculation on the input layer, only the output layer.

    image

    Neural networks can grow in complexity by adding additional layers which are commonly referred to as “hidden layers”.  In theory a network can contain an infinite number of layers with an infinite number of neurons although this is neither practical or necessary.

    image

    The only remaining question then is how do we know what weights will give us the outputs we are looking for.  A simple feed forward neural network uses a technique called forward and back propagation to train the network and find the optimal weights.   There are dozens of books and blog posts devoted to the subject of how the forward and back propagation algorithms work, but for the sake of this blog post I’ll provide an introductory explanation along with pointers to additional information.

    The main idea requires randomly initializing our weights and pushing the inputs “forward” through the network so we can make an output prediction.   We then use a cost or loss function to calculate how far our prediction was from the expected result.

    Our ultimate goals is to reduce our error or cost to the lowest point possible (sometimes referred to as the global minimum).  To do this we use an algorithm called gradient decent.   The goal of the gradient descent algorithm is to find the partial derivative of the cost function with respect to each weight.

    image

    In other words we’re looking for the direction (+/-) and slope of our cost function to tell us how large to adjust our weights and in which direction in order to get to zero cost (or close to it).  If the gradient is 0 we have reached our minima.   While I won’t go into the details thanks to the concept of the chain rule in calculus we can actually start at the output layer , perform the gradient descent algorithm, and “back” propagate it to the next layer and all the way back to our inputs.  Along the way we are calculating how much we need to adjust our weights to get closer to that zero cost.

    image

    When training a neural network we continue to forward and back propagate until we we have minimized the error.  While I have grossly oversimplified the explanation for forward and back propagation, this is fundamentally how neural networks work.  I have provided links to more detailed descriptions at the end of this post.

    Putting it All Together

    Now that we have some basic knowledge of end-to-end speech recognition systems and neural networks, we’re ready to make a simple end-to-end speech recognizer.  To build this recognizer I used python and the numpy library to help with the matrix math.

    However, before we start we need a simple speech data set.  Preferably one consisting of utterances with only single words.  This would eliminate the need to deal with time alignment (i.e. which text goes with which audio segment in time).  Luckily I found a great freely available dataset consisting of people speaking single digits 0 – 9 with fifty utterances per digit per person.   This data set met the criteria of being a single word while also being sufficiently large enough to train a neural network.

    With labeled audio data in hand the next step required is reading in the audio data and the associated labels  For this I used the python librosa library.  Librosa provides easy to use out-of-the-box functions for computing the Short Time Fourier Transform (STFT) which is necessary to get the frequency spectrum of our audio signal (e.g. our input signal).  Librosa additionally provides handy functions for computing other audio features like Mel Frequency Cepstral Coefficients (MFCC) which can also be a useful audio input feature (note my code provides an alternative implementation that uses MFCC’s instead of the raw spectrum)

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    
    for files in file_list:
        relative_path = 'recordings/' + files[0]
        file_name = os.path.join(os.path.dirname(__file__), relative_path)
        y, sr = load(file_name, sr=None)
        filesize = sys.getsizeof(y)
     
        if output_type == 'spectrum':
            spectrum = stft(y, nfft, hop_length=int(filesize / 2))
            mag, phase = magphase(spectrum)
            mag_input.append(mag)
     
        mfcc = feature.mfcc(y, sr, n_mfcc=nmfcc, hop_length=int(filesize / 2))
        mfcc = mfcc[1:nmfcc]
        mfcc_input.append(mfcc)
     
        digit.append(files[0][0])

    Beyond the audio, we also need to store the associated digit spoken in each audio file.   When training a multiclass classifier ( in our case our classes are 0 – 9) it’s common to use something called “one hot” vectors to represent the output.   This is just a vector where all the classes are represented by 0 except for the one element representing the actual output class.   So in our case we have a 10 element vector and if the audio file is someone saying “one’ the vector would look like [0 1 0 0 0 0 0 0 0 0 ].

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    
    class digits:
        zero    = [1, 0, 0, 0, 0, 0, 0, 0, 0, 0]
        one     = [0, 1, 0, 0, 0, 0, 0, 0, 0, 0]
        two     = [0, 0, 1, 0, 0, 0, 0, 0, 0, 0]
        three   = [0, 0, 0, 1, 0, 0, 0, 0, 0, 0]
        four    = [0, 0, 0, 0, 1, 0, 0, 0, 0, 0]
        five    = [0, 0, 0, 0, 0, 1, 0, 0, 0, 0]
        six     = [0, 0, 0, 0, 0, 0, 1, 0, 0, 0]
        seven   = [0, 0, 0, 0, 0, 0, 0, 1, 0, 0]
        eight   = [0, 0, 0, 0, 0, 0, 0, 0, 1, 0]
        nine    = [0, 0, 0, 0, 0, 0, 0, 0, 0, 1]

    With our inputs and outputs squared away it’s time to define our network. The variables that make up your network are also known as hyper-parameters. For my end-to-end recognizer I selected the following hyper-parameters: (*Note that selecting hyper-parameters is half art and half science and your choices will be critical to the success of your network.  I have provided additional resources below)

    • Number of layers:3 (input, output and one hidden layer)
    • Nodes in hidden layer: 2048 (1x our frequency bins)
    • Activation functions: TanH (HIdden), Sigmoid (Output)
    • Weight initialization algorithm: Xavier or Glorot
    • Learning rate = 0.001
    • Rate decay = 0.0001
    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    
    input_layer = layers.Layer(inputs=training_inputs.shape[0], neurons=training_inputs.shape[1] + 1)
     
    if mode == 'E2E':
        hidden_layer = layers.Layer(inputs=training_inputs.shape[1] + 1, neurons=2048,
                                    activation=activationfunctions.Tanh_Activation,
                                    activation_derivative=activationfunctions.Tanh_Activation_Deriv)
        hidden_layer.Initialize_Synaptic_Weights()
     
        output_layer = layers.Layer(inputs=2048, neurons=training_outputs.shape[1],
                                    activation=activationfunctions.Sigmoid_Activation,
                                    activation_derivative=activationfunctions.Sigmoid_Activation_Derivative)
        output_layer.Initialize_Synaptic_Weights()if mode == 'E2E':

    nnet = NeuralNetwork(layer1=input_layer, layer2=hidden_layer, layer3=output_layer, learning_rate=0.001,
    learning_rate_decay=0.0001, momentum=0.5)

    So now that we have our inputs and outputs, and we’ve defined our network, all we need to do is train using our forward and back propagation functions. Per my earlier description the forward propagation algorithm is quite simple and is really just summing the weighted inputs and applying the activation functions. Using matrix math this can be written in three or four simple lines of code.

    1
    2
    3
    4
    5
    
    def Feed_Forward(self, inputs):
        self.l1_inputs[:,0:self.layer1.neurons-1] = inputs
        self.l2_hidden = self.layer2.activation(dot(self.l1_inputs, self.layer2.synaptic_weights))
        self.l3_output = self.layer3.activation(dot(self.l2_hidden, self.layer3.synaptic_weights))
        return  self.l3_output

    The forward propagation algorithm gives us our predicted output.  Using that predicted output we can perform our back propagation.  Much like my earlier explanation we need to perform a series of steps for each layer.   Specifically we need to calculate the error, calculate the gradient, and adjust our weights based on the previous two calculations.

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    20
    21
    22
    
    def Back_Propogate(self, outputs):
     
        output_deltas = numpy.zeros((self.layer1.inputs, self.layer3.neurons))
        l3_output_error = -(outputs - self.l3_output)
        if self.layer3.activation_derivative == activationfunctions.Sigmoid_Activation_Derivative:
            output_deltas = self.layer3.activation_derivative(self.l3_output) * l3_output_error
        elif self.layer3.activation_derivative == activationfunctions.softmax_derivative:
            output_deltas = l3_output_error
        elif self.layer3.activation_derivative == activationfunctions.Oland_Et_Al_Derivative:
            output_deltas = self.layer3.activation_derivative(self.l3_output) - outputs
     
        hidden_deltas = numpy.zeros((self.layer1.inputs, self.layer2.neurons))
        l2_hidden_error = output_deltas.dot(self.layer3.synaptic_weights.T)
        hidden_deltas = self.layer2.activation_derivative(self.l2_hidden) * l2_hidden_error
     
        adjustment1 = self.l2_hidden.T.dot(output_deltas)
        self.layer3.synaptic_weights = self.layer3.synaptic_weights - (adjustment1 * self.learning_rate) #+ self.l3_output_adjustment * self.momentum
        self.l3_output_adjustment = adjustment1
     
        adjustment2 = self.l1_inputs.T.dot(hidden_deltas)
        self.layer2.synaptic_weights = self.layer2.synaptic_weights - (adjustment2 * self.learning_rate) #+ self.l2_hidden_adjustment * self.momentum
        self.l2_hidden_adjustment = adjustment2

    To bring it all together we just need to iterate over our forward and back propagation algorithms until we have stopped learning or have reduced our cost or error to it’s lowest possible point.

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    
    def Train(self, inputs, outputs, iterations):
        for iteration in range(iterations):
            error = 0.0
     
            # random.shuffle(patterns)
            # turn off random
            randomize = numpy.arange(len(inputs))
            numpy.random.shuffle(randomize)
            inputs = inputs[randomize]
            outputs = outputs[randomize]
     
            self.Feed_Forward(inputs)
            error = self.Back_Propogate(outputs)
            error = numpy.average(error)
            if iteration % 10 == 0:
                print('error %-.5f' % error)
            # learning rate decay
            self.learning_rate = self.learning_rate * (
            self.learning_rate / (self.learning_rate + (self.learning_rate * self.learning_rate_decay)))

    That’s it!  While there is a lot more glue code and learning that went into this implementation what I have presented here represents the fundamental building blocks of a basic end-to-end speech recognition system.  I have made the full project available on GitHub and you can evaluate the code yourself in order to fully comprehend all the details.  I’ve also provided a bevy of resources below that helped get me to this point and can do the same for you.

    Final Thoughts

    You might be asking why a senior leader in my position would spend the time required to go through this exercise.  There are some general principles I like to follow and I think anybody managing a research oriented (or really any engineering) team should consider as well.  Specifically:

    • ABL – Always Be Learning:  If you want to innovate you need to be up to speed on the latest technology trends.
    • Earn your team’s respect:  The best way to earn the respect of your technical team is to get into the trenches.  Show them that you understand their job and all the pain that comes with it.  In other words write code (any code), test it, check it in, and push it to production.
    • Lead by example: If you want your team to “innovate for the masses”, it’s best demonstrate the behaviors you are looking for.

    Hopefully this post has given you a basic understanding of end-to-end speech recognition systems and neural networks  If you’re really brave perhaps you’ve learned how to build your own simple end-to-end recognizer.  But if you take nothing else away from this article I hope it’s that you’ll invest your time improving your own technical skills and getting in the trenches to earn your team’s respect.

    In an upcoming post I’ll dig deeper into end-to-end speech recognition algorithms and how they work.  Specifically we’ll cover recurrent neural networks and the connectionist temporal classification algorithms that truly allow these systems to be superior over traditional speech recognition systems.  In the mean time I hope you get a chance to “wreck a nice beach”!

    References
    1. “How to build a simple neural network in 9 lines of Python code” – Milo Spencer-Harper
    2. “How to build a multi-layered neural network in Python” – Milo Spencer-Harper
    3. “Understanding and coding Neural Networks from Scratch in Python and R” – Sunil Ray
    4. “How to Compute the Derivative of  Sigmoid Function (fully worked example)” – Jeremy (no last name)
    5. “Practical recommendation for Gradient-Based Training of Deep Architectures” – Yoshua Bengio
    6. “How to train your Deep Neural Network” – Rishabh Shukla
    7. “Understanding the difficulty of training deep feedforward neural networks” – Xavier Glorot and Yoshua Bengio
    8. “Deep Learning Basics: Neural Networks, Backpropegation, and stochastic Gradient Descent” –  Alex Minnaar
    9. “Speech Recognition: You down with CTC” – Karl N.
    10. “Deep Speech: Scaling up end-to-end speech recognition” – Andrew Y. Ng et al.
    11.  “Connectionist Temporal Classification: Labeling Unsegmented Sequence Data with Recurrent Neural Networks” – Alex Graves et al.

    Creating a Management Philosophy

    Just In case you couldn’t infer this from my previous posts some folks consider me opinionated and occasionally dogmatic.  What else would you expect from a born and raised New Yorker, who grew up in a household where arguing your point was considered a cultural birthright!

    Unfortunately, while having strong opinions and ideas can be a positive, I’ve found throughout my career that those ideas don’t always resonate with coworkers.  Even when those ideas are sound.  When I started to lead and manage larger teams this increasingly became a handicap and I soon realized I needed a better way to get my thoughts across.

    Enter Pete Carroll.   That’s right, the same Pete Carrol who led the Seattle Seahawks to two super bowls and USC to two college national championships.  Not too long ago a good friend and neighbor happened to tell me about a management training class he attended based on material from Pete Carroll and his “Win Forever” philosophy.  It sounded very compelling and I immediately purchased Pete’s book based on the same concepts.

    It was immediately clear from reading this book that Pete Carroll had faced similar challenges earlier in his career.

    “But while I had a sense inside me of what we needed, I hadn’t articulated it very well.  I didn’t have the details worked out in my own mind so that I could lay them out clearly and convincingly to anybody else”

    In short, this book preaches a simple strategy for dealing with an inability to convey your ideas, which is to write them down, iterate on them, and formulate them into a single cohesive vision.   By doing so you change the conversation from “hey, here is my opinion” to “hey, I have a strategy for winning and here it is”.   Or in Pete’s words;

    “by December I finally had a clear, organized template of my core values , my philosophy, and – most importantly – my overarching vision for what I wanted to stand for as a person, a coach, and a competitor”

    Armed with a singular vision and philosophy you have a solid foundation to convey your thoughts.  And suddenly you have transformed disparate ideas into a recipe for success.  The implications of documenting your philosophy are huge and by doing so you will:

    • Set clear expectations for your employees
    • Set expectations for executives, higher level managers and peers for how you operate and how it will benefit them
    • Have a recipe for success that you can continually improve and iterate on

    With this information in mind I decided to take years of ideas I had accumulated and started to jot them down.  I refined them and wove them into overarching vision.  And when I thought about what I was ultimately trying to achieve it became clear that I was always trying to deliver truly innovative software to as many people as possible.   And so my Innovate for the Masses™  philosophy was born.   I present it to you below unedited.  It is a continual work in progress but something that has served me well so far.


    Philosophy: Innovate for the Masses™

    Innovate for the Masses™
      • Create unique and defensible value: Products should deliver something truly special that cannot be found in other solutions and simultaneously provide a defensible moat.
      • Build best in class solutions: Products should be fully functional and should not cut corners. We should do everything required to delight customers, nothing more and nothing less.
      • Support all customers: Products should be accessible by all existing and future customers. One off solutions are never okay.
      • Enterprise class reliability and scalability: Product should be robust with 4 9’s reliability and the ability to scale to all customer demands.
    How we do it
    • Ruthless prioritization: We question the necessity and value of every feature or piece of code. We only work on things that deliver essential value to the customer.
    • Avoid premature optimization: we only build exactly what the customer needs. No more and no less.
    • Read between the lines: We listen to our customers but don’t just cater to their demands. We find the commonality amongst all our customer’s requirements and build a truly unique and defensible product that surprises, delights, and addresses their needs.
    • Communicate like crazy: We are one team with one vision and one goal. Everybody must constantly be talking to innovate and build cohesive products
    • Right spot right time: We believe every team member plays an important role in the team’s success whether that is in a leading role or a supporting role.
    • Work harder than anyone else: We will win by out working all of our competitors.
    • Don’t chase the competition: We don’t chase every move our competitors make. We pay attention but follow our vision and goals and methodically work towards delivering on them without being distracted.
    Expectations for our people
    • Be insanely passionate: Our employees exude passion. We are a passion first organization.
    • Get a lot done / execute like crazy: Our employees are insanely productive. They get more done than anyone else.
    • Care: Our employees give a shit. They care about the product, team, company, and customer like something they hold dear.
    • Have a sense of humor: Our employees laugh. At themselves and each other. We believe you should leave work every day having smiled so much it hurts.
    • Don’t whine or complain: Our employees don’t whine or complain they express their opinions and try to instigate change in the direction they want to see. If a decision doesn’t go their way they disagree and commit.
    • Don’t play politics: Our employees don’t play politics. They lay it all out on the table and do their job to the best of their abilities … that is what they get rewarded for.
    • Dare to disagree: Our employees disagree loudly and proudly. Good disagreement is central to progress. Different opinions are valued and we seek out constructive conflict

    Having a vision and philosophy is not all rainbows and unicorns.   Creating a philosophy and broadcasting it to your coworkers is the equivalent of driving a giant metal stake deep into the ground.   You may find throughout the course of your career that sometimes people don’t agree with your strategy and when they don’t you only have three options, change your strategy, change their minds, or move on.  Or again to quote Pete:

    “Coach Seifert was specifically adamant that I not change who I was or my mentality. He said clearly “Pete, you’ve got to do it the way you know how.” After my experience in New York, I wondered if I shouldn’t try to be more political, but the advice I got from the two mentors was uncompromising – and some of the best I ever received.”

     

    In closing if you are anything like me or Pete Carroll I strongly encourage you to write down your great ideas and formulate them into a cohesive philosophy.   It will be will worth your while.

    **For reference Pete Carroll explicitly calls out John Wooden for influencing his strategies and techniques.  I highly encourage people to also read John Wooden’s book “Wooden on Leadership”

    Twitter Feed

    Who would think a playboy image would bring back fond memories of my ............... adaptive signal processing class! @ragomusic twitter.com/ozm/status/119…

    About 3 weeks ago from Jason Flaks's Twitter via Twitter for iPhone

    Managing Research Projects in an Agile Development Environment

    Anyone who has worked in an agile organization has found that certain projects don’t quiet fit the agile mold.   Nowhere is this more apparent than with research oriented projects.   After all if there is complete uncertainty in the scope and outcome of a project, as would be the case in a research project, how do you create user stories and estimate story points?   And if you can’t create stories and estimate the associated costs how can you hold your team accountable, communicate status to the rest of the organization, and make cost / benefit tradeoffs?  Simple!  You can’t.

    I’ve personally dealt with this issue after hiring several researchers to work on an agile software product team.  Initially, I struggled to interleave our research projects with our other production work so I started looking for a solution.  The answer to my problem came after reviewing the agile literature and the scientific method and concluding that research projects really just represent an extreme of what the agile process is ultimately trying to solve. Below I will walk you through how I arrived at this solution and details on how you can apply similar tactics in your own research organization.

    AGILE PROCESS

    Early in my career at Microsoft someone handed me a copy of Steven McConnell’s book Code Complete.

    At the time my greatest take away from that book was the concept of the “Cone of Uncertainty”.   The “Cone of Uncertainty” states that the uncertainty of a given project decreases as time progresses and more details are flushed out.

    image

    Historically the “Cone of Uncertainty” was dealt with by creating detailed upfront plans and using waterfall project management approaches.  The trouble with those methodologies is that they’re extremely resistant to scope change.   Largely because scope change reintroduces uncertainty.

    The agile manifesto attempts to eliminate the “cone of uncertainty” problem by following the principle of “Responding to change over following a plan”.   Most agile methodologies use some form of iterative development to reduce uncertainty, with the idea being that if you’re working on smaller well defined chunks of a larger project uncertainty is removed and the project can slowly adapt to changing requirements.  Mike Cohn wrote in an article titled “The Certainty of Uncertainty”.

    “The best way to deal with uncertainty is to iterate. To reduce uncertainty about what the product should be, work in short iterations and show (or, ideally give) working software to users every few weeks. Uncertainty about how to develop the product is similarly reduced by iterating. For example, missing tasks can be added to plans, inadequate designs can be corrected sooner rather than later, bad estimates can be amended, and so on.”

    If I take the above information together I can conclude two things.  First, the agile method attempts to reduce or eliminate uncertainty by making every project a function of smaller work items iterated over time.  Or framed in mathematical notation:

    clip_image002[8]

    Where: T = Max Iterations, M = Backlog, N = User Stories belong to M

    Secondly, if a research project is really just a project with maximum uncertainty then the same framework should apply.   Only there would be an unbounded number of work items over an unbounded amount of time.   Or framed in mathematical notation:

    clip_image002[14]

    According to this logic a research project should actually work within an agile framework.   We just need to figure out how to construct M (i.e. backlog) and how to bound M and T (i.e. number of iterations).

    SCIENTIFIC METHOD

    So what are reasonable user stories for a research project and why are they potentially infinite?  It occurred to me that research in general follows the scientific method and that the scientific method may be a good framework for story generation.

    image

    In essence the scientific method can be boiled down to three phases: a research phase, an iterative hypothesis testing phase, and a communicate or productize phase.  The unbounded component of research is that many hypotheses end in failure leading to another hypothesis that must be tested and this can potentially go on ad nauseam.  This provided me a compelling framework for how to break research into user stories.

    image

    The first story in any research project correlates to the first phase in the scientific method.  This story should be a time bounded spike that frames the initial question, covers any background research, and has an acceptance criteria of generating the required stories for the next phase of the project, hypothesis testing.

    The next set of stories are all part of the hypothesis testing phase.  These stories include any development work required to test the hypothesis, any data collection required, running the tests, and analyzing the results.   If the hypothesis proves false the team should circle back to the background research phase and continue on with the process.

    The final phase in this framework is only relevant when a hypothesis is proven to be true.   This phase contains multiple stories including any communication or publishing of results, IP protection, and a handoff to whomever might be building the final product (which might be the same team).   The final handoff story should also be a spike and the acceptance criteria should include the user stories required for the production deployment.

    BOUNDING AN UNBOUNDED PROJECT

    Now how do you go about making sure research stories don’t go on forever?  How do you bound T and M?  And how do you communicate the cost / value trade offs with management?

    I have found that the previously described framework only works if you apply the following guidelines in conjunction.  Specifically

    1. For any research project to be considered we must have enough information for the project to pass the “sniff test” (i.e. Is it possible in a reasonable amount of time and does it make business sense).
    2. The initial estimate for research projects are based on the expected number of hypothesis iterations and the cost must be inline with the expected project value (i.e. if the research is perceived to have large value it may be worth iterating for a long time).
    3. If the number of hypothesis iterations exceeds the original cost the cost/benefit analysis must be revisited and the project should be canceled if the cost has exceeded expected value.
    CONCLUSION

    What I have presented here is a process by which you can take an unbounded research project and place a structure around it that will work in companies using an agile development methodology.  Besides allowing research projects to function in an agile organization this framework also provides a method for bounding research problems and communicating the cost / benefit trade offs to management and other relevant parties.   For those who have faced similar issues integrating research oriented projects into an agile culture I hope this methodology provides some ideas on how you can better integrate research into your processes.

    Microsoft’s 5.1% Word Error Rate (WER) Announcement is Complete and Utter Bullshit

    eoln69_hi

    I apologize! That title was actually generated by Microsoft’s speech recognition system incorrectly transcribing “Microsoft’s 5.1% Word Error Rate (WER) Announcement is Completely Misleading”.   Okay, that was snarky, but I promise Microsoft compelled me to write that.  You see in the course of editing my previous post Microsoft had to go and put out a press release announcing “Microsoft Researchers Achieve new Conversational Speech Recognition Milestone”.  Their announcement flies in the face of my previous post and therefore I had no choice but to attempt an epic takedown.

    Before I try to dismantle Microsoft’s irrational clam I would like to state that the the researchers at Microsoft (some of whom I have crossed paths with while working on the Xbox Kinect and HoloLens) have done some solid research with potential implications on how we build production speech recognition systems.   I have no issues with the technical nature of the research paper underpinning the press release, but I do take issue with the marketing and PR spin applied on top of it.  So without further ado “LET’S GET READY TO RUMBLE”.

    There are two primary issues with the announcement made by Microsoft:

    1. Does Microsoft’s testing provide conclusive evidence that the 5.1% WER results will generalize
    2. Are the tactics used viable from a cost/compute/timeliness perspective in a production system

    Let’s tackle each of these issues independently.

    Will the Results Generalize

    In my previous post I discussed why large data sets were critical for training truly accurate conversational speech recognition systems.   While I do take issue with the data size used to train the Microsoft speech recognition system, the larger issue is with the test set used to validate the word error rate.

    In Andrew Ng’s seminal talk on the “nuts and bolts of machine learning”, he goes into great detail on the different data sets required for training, testing and validating machine learning algorithms.  I encourage anybody interested in the optimal process for training and testing machine learning / AI like algorithms to watch this seriously awesome video.   In terms of Microsoft’s research I want to focus on the relatively small size of their test corpus, it’s overlap with the training data, and the fact that the chosen corpus appears cherry-picked.

    Corpus Size

    The test set Microsoft selected for calculating the reported  5.1% WER is the 2000 NIST CTS SWITCHBOARD corpus.  While I was unable to find the specific number of hours of conversation in this test corpus I was able to confirm that the 1998 and 2001 NIST CTS data sets contained 3 and 5 hours of conversation respectively.  We can therefore assume the number of hours of conversation in the 2000 set is similar in duration.   When considering the overall size of the conversational speech domain explained in my previous post  a test set of this size is hardly sufficient for making any broad claims about meeting or beating human transcription accuracy.

    Training Data Overlap

    As you dig into the details of the NIST corpus a dirty little secret is quickly revealed.  Let me start by quoting directly from the source:

    “Of the forty speakers in these conversations thirty-six appear in conversations of the published Switchboard Corpus.”

    Let me translate that for you.   Thirty-six of the speakers in the test corpus are the same speakers used in Microsoft’s training corpus.   I’ll also remind you that the Switchboard corpus only has 543 speakers to begin with.  This raises a foundational questions about whether the test data is really distinct relative to the training set.   You see almost all modern speech recognition systems use something called i-vectors to help achieve speaker independence (sometimes called speaker adaptation).  Since the same speakers, on the same devices, in the same environments exist in both the training and test corpus there will invariably be a correlation between the i-vectors generated by the two data sets.

    Per the diagram below, a truly honest measure of WER would require the the test data be truly distinct from the training set .  In other words it should pull from a data set that includes different speakers, different content, and different acoustic environments.   What is clear from the Microsoft paper is that this didn’t happen which calls into question whether the published results will truly generalize.  It also greatly diminishes the the validity of any claim about a new “milestone” being achieved in conversational speech recognition.

    Cherry-picking

    It’s worth noting that the full 2000 NIST CTS corpus actually contains a total of 40 conversations.   Twenty of those conversations are from the Switchboard corpus and twenty are from a different corpus called “Call Home”.   This begs the question of why Microsoft only validated against the Switchboard portion of the corpus.   While I can’t say for sure what their intent was, my best guess is because if they had used the Call Home data the results would not have led to the desired goal of meeting or beating “human accuracy”.

    Taken altogether, the small corpus, with overlapping data, and a cherry picked data set you can’t help but ask did Microsoft really achieve a “new conversational speech recognition milestone”?

    Is it Production Ready

    EBTKS.  For those not familiar with texting slang, that stands for “Everything But the Kitchen Sink”,  and it’s really the best description of the system Microsoft used for this research.  This calls into question the production viability of their proposed solution.

    Ensemble Models

    At the acoustic model (AM) and language model (LM) layer Microsoft is using an ensemble model technique.   This technique requires training multiple models and processing each utterance through every model.   A separate algorithm is used to combine the outputs of the different models.   In essence this equates to trying to run multiple recognizers at once for every audio utterance.  It currently requires an enormous number of machines to transcribe phone calls in real-time at scale   Microsoft appears to be running 4 distinct AMs and multiple LMs which will have serious performance impacts.   This raises questions about the number of machines and associated costs required to run a system like the one used in Microsoft’s paper.

    Language MODEL RESCORING

    On top of the ensemble modeling Microsoft is also using language Model Rescoring.  In order to rescore you usually have an initial language model produce an N-BEST lattice which is basically the top N paths predicted by the language model.   This lattice needs to be stored or held in memory in order for the rescoring to take place.   In Microsoft’s case they are generating a 500-best lattice.   While not crazy holding a 500-best lattice in memory in a scaled production speech recognition system would not be ideal unless it provided significant accuracy gains.   According to the paper the gains from rescoring were minimal at best.

    In Conclusion

    So where does that leave us?  Microsoft has done some great research on advancing speech recognition algorithms.   Research that I greatly appreciate and hope to review further.   However for Microsoft to even imply that they achieved some epic milestone in matching human transcription accuracy is downright preposterous.

    In the words of renowned Johns Hopkins speech recognition researcher Daniel Povey:

    “… … this whole competition between IBM and Microsoft on Switchboard is just a pissing contest, in which they both try to add in more data and bigger system combinations to beat the other one’s number.  It doesn’t really indicate any special progress.”

    “Blinded by the Light, Revved up like a ???”

    Image result for i don't understand what you're saying

    I probably sang that Manfred Mann song a thousand times in my teen years and I was pretty sure the last word in that lyric was a feminine hygiene product until Google came along and taught me otherwise.   It turns out the lyrics to Blinded by the Light are very difficult to understand and so is conversational speech.

    For my first substantive blog post on this site I’d like to continue on a theme we have been covering over at Marchex around the complexity in building automatic speech recognition (ASR) systems that can accurately understand unbounded conversational speech.   In this post I intend to dive a little deeper into WHY conversational ASR systems are so difficult to build, possible solutions to improve them, and the bounty for those who finally succeed.

    There are really three primary issues that are limiting current systems from accurately recognizing conversational speech: Data, Data, and Data.    More specifically: Required Data Size,  Lack of Publically Available Data Sets, and Cost and Complexity with Acquiring the Required Data.

    Required Data Size

    There is no strict answer for how much data is needed to solve a given machine learning problem, but one oft-cited rule is the “rule of 10”.   The rule of 10 states that you need roughly 10 times as many examples as you have parameters.   While there are multiple parts of an ASR system including an acoustic model (AM) and a language model (LM), for now I am going to focus on the LM.   One parameter used in an LM is called an n-gram, specifically in most cases a trigram.   A trigram is basically the probabilities of any 3 words being seen next to each other.   So if we take the rule of 10 that would imply we need 10 times the number of 3 word combinations required for our task.

    This is where the problem arises.  You see we humans write beautifully but we speak like idiots.   Grammar goes out the window when people talk, we stutter, words are often repeated over and over while people search for their next thought, and honestly some folks downright make up words that don’t even exist.   Taken together that means one can expect to see almost ANY combination of 3 words in the wild.   Everything from “a a a “ to “zebra zebra zebra” .   So if you don’t mind rewinding your brain to highschool math and combinatorics that means the number of 3 word combinations is:

    | Number of Words in US English | ^3

    or

    | ~500,000 | ^3 = 125 QUADRILLION (i.e. a really #$%&’ing big number)

    If we apply the rule of 10 we would need 1.25 QUINTILLION (i.e. an even bigger #$%&’ing number) utterances (basically a spoken sentence) containing examples of these trigrams.   Let me put this in perspective for you.    A single spoken utterance saved in a text file is roughly 50 bytes in size.   So in order to to store 1.25 QUINTILLION utterances I would need 50 * 1.25 QUINTILLION bytes of storage.  Or … 62,500 Petabytes!   For reference 20 years of internet archiving only consumed 23 petabytes as of 2015.  And if that doesn’t frame it for you think about it this way.   The average utterance duration is roughly 1.5 seconds. If I were to string 1.25 QUINTILLION recorded utterances together it would take approximately 60 millennia to play it back!

    So what’s the point?   The point is that the data size required to cover all possible examples of spoken US English is almost inconceivable.  Is the rule of 10 an exact science?  No.  Does it matter?  No, because even if this estimate is wrong by 1/2 or 3/4 it is still huge.   Ultimately the data size needed to properly train a conversational ASR system is gargantuan.

    Lack of Publically Available Data Sets

    Okay so we need a lot of data.   Can’t we just buy it?  No!  Most publically available data sets are shockingly small compared to the size of the domain I described above.  As my fellow Marchex coworkers reported in our recently published research paper the size of 2 of the most commonly used data sets, fisher English and switchboard, is prohibitively small.

    Switchboard Fisher English Marchex
    Hours

    309

    2,000

    5,000+

    Speakers

    543

    20,407

    605,000

    Utterances

    391,000

    1,600,000

    11,700,000

    Conversations

    2,400

    16,000+

    288,000

    Words

    3,000,000

    18,000,000

    79,500,000

    Dat Acquisition Cost and Complexity

    Alright if you can’t buy it why don’t companies just go and collect the data themselves?   Well it turns out collecting 62,500 petabytes or 60 millennia’s worth of people conversing is no simple task.   There are two primary problems, collecting that amount of audio data and labeling it.

    Audio Data

    Where could someone acquire that quantity of audio data?   Well, there are countless hours of TV and Radio interviews out there but the dialog is generally scripted and edited so not reflective of true conversational speech.  On top of that in most cases companies do not have the legal rights to the data and acquiring those rights would be prohibitively expensive.

    Amazon, Apple, Microsoft, Google, and other companies are all collecting mountains of data from various voice assistants (Alexa, Siri, etc.) and voicemail messages.   However all that speech data is mostly unidirectional and non-conversational (“Alexa tell me the weather” is not really conversational).

    That leaves one obvious channel for acquiring conversational speech and that is phone calls.    So why don’t companies just collect call recordings at scale? The answer is simple:  WIRETAPPING.

    In the US wiretapping is a federal and state statute aimed at ensuring your communications are private and there are criminal and civil penalties for those who violate the law.   What makes wiretapping laws particularly problematic is that the law varies by state specifically around who must consent to being recorded.

    So why does this matter?   Well because 12 states require bidirectional consent and phone networks are open (nobody can guarantee they control both sides of the call).  While any company can update their “terms of service” to notify you that you are being recorded, they would have no easy way to guarantee that the other party has consented.   Unless they start playing that pesky message “this call might be recorded or monitored” in front of every call, including your weekly call with your mother!  This puts scalable call recording for consumer oriented phone services mostly out of reach since the risk of violating a criminal law is too high (I think it is safe to say Mark Zuckerberg has no interest in going to jail).

    In fact just ask Google who has dealt with an ongoing wiretapping case because they were scanning the emails of GMAIL users to place targeted ads.   The argument is in fact incredibly similar in that Google was not just reading the email of GMAIL users but also any yahoo, Hotmail, etc. user who sent a GMAIL user an email.   In the 12 states requiring bidirectional consent the non-Gmail users never consented which has potentially caused Google to violate the law.

    Labeling

    Even if by some miracle we could collect that amount of audio data how would we label it?   In general ASR systems (and all other machine learning systems) require accurately labeled data (sometimes called “ground truth”).   In the speech recognition world that generally involves hand transcribing data.   And if it would take 60 millennia to read out that much speech imagine how long it would take to hand transcribe it.   Simply put, it is not feasible in our lifetimes at any reasonable cost.

    What’s the Solution

    It turns out almost all companies record phone calls.  Recordings from any one company would have highly biased content but in aggregate consumer to business recorded phone calls are an amazing source of conversational speech at scale.   Because you need a  wide cross section of content to ensure subject matter diversity, companies who provide platform call recording solutions and have legal access to the aggregate data are really the best sources of this content.

    But what about the labeling?  Well the only reasonable solution for labeling that content is using unsupervised or semi-supervised automated solutions for labeling the data.   This is where Marchex has invested and you can read more details about our semi-supervised approach in our research paper.   I hope to cover this topic in detail in a future post.

    Why Does Any of this Matter

    You might be asking if highly accurate conversational speech recognition is really necessary.   Or you might be thinking “My Alexa already works awesome”   But if you are a sci-fi nerd like me you’re anxiously awaiting the day that you can step foot on the holodeck and have a real conversation with an AI entity or crack open a beer with a fully conversant robot like Data from Star Trek.   For that to happen we need to truly understand conversational speech.  We need to understand it so machines can properly decipher what humans are saying and we need to understand it so machines can generate speech that mimics human dialog.

    Highly accurate conversational speech recognition is necessary for us to fulfill the promised vision of artificial intelligence.  Who knows maybe in a few years a holographic Manfred Mann and I will be doing a duet in my own personal holodeck.   Can you hear it?   “Blinded by the light, revved up like a deuce …”

    Hello World!

    I’m back!  After a long hiatus from blogging I’ve decided to give it a go again.  I’d like to use this blog as a forum for sharing my thoughts on advanced technologies like speech recognition, audio signal processing, natural language processing, and artificial intelligence.   Additionally I will be covering my theories on how to manage team’s building these types of technologies and how we bring these products to market.   I hope you find the content engaging and stay tuned for my first official post in the coming weeks!