In trying to do more screencasts as well as blog posts I created a screencast that goes over this Wintellect post on ML.NET I did recently.
Hope you enjoy it! And please submit any feedback if you have any.
During the same week PyCon was going on, Microsoft had their annual Build conference. If you're not familiar with this conference this is where Microsoft announces a lot of new things for developers. The main focus of this year's Build was about artificial intelligence.
Also in the same way as PyCon, Microsoft records all of the sessions at Build so we're free to watch them later. With that, I present to you what I think are the top five session at Build that go over artificial intelligence. Why top five instead of top ten like in the PyCon session post? Well, there just wasn't enough to do a top ten. :)
This session goes a lot into the Cognitive Services. Not only how much it can improve your applications by incorporating them, but also how easy they are to implement.
Continuing with more information about Cognitive Services, this session goes into some of the other capabilities that were announced at Build. A nice demo they showed went over what they did with the released JFK files and how they used Cognitive Search to analyze all of that data.
This is an interesting talk that goes into the basics of data science and machine learning, but also goes into how to integrate the mindset of DevOps into data science. Doing so can help with things like testing and version control as a process to doing data science which can help reproducability and putting models into production.
This video is full of great demos that show off the power of several Cognitive Services. Although, my favorite demo is the one where they show the power of having a bot on your site.
There are quite a few deep learning videos out there, but I firmly believe that this is one of the best. The way that Seth and Chris describe the deep learning process of the neural network algorithms makes it very understandable to what's going on behind the scenes when you train a model.
So a lot of interesting things happened at build in terms of AI, mostly involving their Cognitive Services APIs. These APIs, I think, are going to really help make your apps stand out from the rest of the crowd and Microsoft looks to continue to add to them with APIs from their Cognitive Services Labs.
I'm definitely looking forward to what else they come up with.
I've done book reviews in the past, but as I mentioned in my post on deliberate practice, I intended to get into more research papers. I don't have any academic experience in reading research papers, I can't quite be able to reach the very math centric or abstract papers (at least not yet), but there are quite a few that, with some patience, anyone can read and learn from.
With that, I'll do my best to summarize the more interesting research papers I read here.
When first getting started with reading research papers is that you don't read the whole thing in one pass through. Read a bit of the Abstract and Introduction sections to make sure it's a paper that looks interesting.
For more on how to read a research paper, Siraj Raval has an interesting video:
For this first Research Minute, I chose the paper A Few Useful Things to Know about Machine Learning by Pedro Domingos from the University of Washington. This paper has some very practical thoughts on creating your machine learning models. Below are a few of the ones I found the most helpful.
This is the main reason you create a model, but often times you won't have any true test data from the wild to test on. Because of this it is recommended to split your data even before you do any pre-processing on it. scikit-learn has a very handy test_train_split method just for this purpose.
What we're trying to avoid is overfitting your model (which we explore a bit later in this post). An overfit model won't give you accurate results if you deploy it and it predicts with real data that it hasn't seen before when the model was trained.
The author puts this very nicely that doing machine learning is a lot of work. It's not magic that we just give it data and you're done. There's a lot more to it than that. A good analogy from the paper is that creating machine learning models is like farming. Nature does a good bit of the work, but a lot of up-front work has to be done for nature to do its thing. It's the same way with creating machine learning models - a lot of pre-processing and business understanding of the data is needed before starting the modeling process.
It's often said that most of the work in data science and machine learning is cleaning and getting your data in a state that you can feed it into a model and also to evaluate how your model's performance does. Fitting data into a model is easy with the current libraries.
Overfitting a model is where the model will score high on training data, but when it sees real data that it hasn't yet seen, it will score low. Ways in which overfitting happens is with bias and variance.
Bias is where the algorithm learns the same wrong thing, which results in the model missing relevant relationships. Recall my previous post about how bias in machine learning is a big topic that we still need to work on. Variance is where the model is sensitive to irrelevant data which can cause it to model random noise in the training data.
A popular way to battle against overfitting is to use cross validation) on your model. Cross validation will take your data and create subsets of it. In each subset it will take a random portion of the data as test data. It does this with each iteration, or fold, and it compares each score of the sampled test data of all iterations. This is a good way to test your model on different data to make sure you get consistent scores. If you don't, then your model may be sensitive to part of the data.
The image below from the article helps explain bias vs. variance through the game of darts.
Sometimes the best way to get a more accurate model is to do feature engineering on your data. This combines or manipulates columns of your data to create columns of new or a different type of data which may be better for the algorithm to detect a pattern. The author states in this paper that feature engineering is the best thing you can do to increase the accuracy of your model. Though, it is a tough thing to learn and there's a book strictly about feature engineering coming out soon.
Feature engineering is tricky. In the book mentioned above, doing feature engineering on your data is more like an art than it is a science. You have to know a lot about your data and have the creativity to think that unrelated parts of data can come together, or that one part of the data can be represented in a different way.
The author points out here in the paper that there can be a tradeoff of three elements when training a model - the amount of data, computational power, and the time it takes to train athe model. The author then goes into how the most important of those three is the time. Since time is a resource we can never get back or improve upon, unless we can invent the time machine, we need to choose algorithms that generalize well to our data as well as being able to be trained quickly. The best way to get both is to use more data with the simplest algorithm possible.
Wouldn't more complex algorithms generalize better than simpler algorithms? Not necessarily. Take a look at the below image from the paper. Multiple algorithms, some more complex than the others, still can generalize the same as the simpler algorithms. However, these simpler algorithms can be much faster to train and predict on new data.
The paper has several more ideas, thirteen in total, but these are the ones I thought gave the most insights from the projects I've done. I'm sure I'll come back to this paper several times as I do more projects and gain more machine learning experience.
You hear a lot about machine learning and how it's transforming industries, but there may be something about these algorithms you may not have heard - its biases.
A bias is favoring one group over the others and these machine learning algorithms are supposed to prevent decisions based on any type of bias a person would have. However, the machine learning models are only as good as the data that you give it. The historical data that we have has a built in bias that we need to address.
I came across an interesting talk from the 2017 NIPS conference that goes over this very well.
There's also an interesting podcast episode of the TED Radio Hour that goes over algorithmic bias.
Researchers know this is a problem in algorithms and they are actively doing research on how to better beat these biases from machine learning models. Until then, though, we would need to be diligent when testing our models to make sure these types of biases aren't integrated in our data.
Jonathan is a normal software developer residing in Columbia, SC. During the day he creates software using the Microsoft technology stack: by night he's doing martial arts or off learning and experiencing new things. He can also be found taking long walks around the parking lot or just sitting around reading.
Jon is a random software developer in North Carolina developing in C# for web and machine learning with ML.NET. I also post videos on YouTube.
