Category: Neural Networks

Posted on

Elon Musk Indicted by SEC, Can AI Help?

The big news from the tech and finance world on September 27, 2018, is that Elon Musk has been sued by the US Securities and Exchange Commission (SEC) for his tweets about taking Tesla private at $420 per share. 

The SEC is seeking to have Musk banned from serving as an officer or director of any public company. Their reasoning? Musk was lying about having funding secured. This implies that he was trying to manipulate Tesla’s share price in an upward direction. Well, it worked, for about a day, that is. On the day of the tweet, Tesla’s share price rose to a high of $379.87 US per share from its previous price of around $350 per share, before falling back to $352 the next day (August 8, 2018). If the markets had actually believed Musk’s Tweet, Tesla’s share price likely would have climbed closer and closer to the mythical $420 price as the take-private day neared.

Tesla’s share price peaking after Musk’s announcement.

Instead, Tesla’s share price dropped like a rock because every savvy investor realized that Musk’s statement was either pure fanciful bluster, a joke about weed, or both. Of course, today has been much worse for Tesla’s share price than any of Musk’s recent ill-advised tweets. In after-hours trading, Tesla’s share price is down as much as 13%. That’s a lot and it is falling dangerously close to their 52 week low. This is all especially troubling considering that Tesla is expected to announce their best quarter ever, in terms of cash flow, in a few days.

  • Loading stock data...

So, what is the SEC doing, was it possible to predict this, and could AI make this type of situation any better? The answer to the first question is unclear, however, the answer to the second two questions is likely, yes.

AI is already being used in the legal profession to help identify responsive documents that must be turned over to the opposing party during a lawsuit. MIT Professor Emeritus Frank Levy leads a research that helps law firms apply machine learning software to the practice of law. 

If AI can predict what documents will be useful in a lawsuit, then whenever the CEO of a publicly traded company does something suspicious, it should be possible to use these same programs to parse historical cases and see what precedent there is for a lawsuit to be filed. At the very least, it could provide some insight into the likelihood of an indictment and, in the future, could even suggest potential courses of action for a company to take if it found itself in this type of situation.

Would the AI be able to help predict whether or not Elon will be convicted? Possibly. While I am not aware of any AIs currently being used to predict the outcome of legal matters, in my September 24, 2018 column, I covered the AI that perfectly predicted the outcome of last year’s Superbowl. While legal cases may be more complicated than a football score, there is likely several orders of magnitude more data about the outcome of various lawsuits than there is about football players, simply because there are WAY more lawsuits than there are football teams.

From a financial perspective, we could use this type of AI to predict potential lawsuits and their results and train the AI to make trades based on these predictions. If these types of AI were already in use, we could expect much smoother and more predictable share prices as the effect/implications of a particular news story would become apparent almost immediately after the information surfaces.

For now, I’ve programmed a simple AI for Elon Musk to help him decide if he should tweet something or not. You can try it, too, if you’d like. It’s posted below:


Posted on

The Best-Worst-Kept Secret in Machine Learning

Neural networks are pretty simple to build. Last time, we picked apart some of the fundamentals and how they really just boil down to linear algebra formulas. There is, however, one single algorithm that is incredibly useful for machine learning and I hadn’t heard of it until today. It’s called Logistic Regression. It’s a five-step process that enables nearly all modern Deep Learning software. Here’s how it goes:

  1. Take your data
  2. Pick a random model
  3. Calculate the error
  4. Minimize the error, and obtain a better model
  5. Become sentient, destroy all humans, dominate universe!

… Okay, I took a little creative licence on that last one. But seriously, it’s that simple. The only complicated part is calculating the ‘error function’ and generalizing it for large and varied datasets.

The error function itself is a big long, somewhat scary formula that looks like this:

Error Function = -1/m \sum_{i=1}^{m} (1-y)ln(1-\alpha(Wx^{i}+b)) + y_i ln(\alpha(Wx^{i}+b))

What the error function is doing though, is really quite simple. The error function is looking at a set of points (usually, pixels of an image). We can represent an image like this:

Dots on a graph (oddly reminiscent of a smartphone or computer display, don’t you think?) 

The job of the logistic regression algorithm is to find a line that divides the red and blue pixels as best as it can. It shifts, translates and iterates, moving the line until it reaches the maximum possible percentage of blue pixels on one side of the line and the maximum possible percentage of red pixels on the other side of the line. As it iterates, it looks something like this:

The Logistic Regression algorithm starts by bisecting the graph at a random location and then it moves the line until it has maximized the number of blue pixels on one side and the number of red pixels on the other side. 

We call this ‘minimizing the error function‘ because what the algorithm is doing is finding the smallest number of blue pixels on the red side and the smallest possible number of red pixels on the blue side. These pixel mismatches are like errors if we’re trying to separate the two.

Here we can see the Error Plot as the algorithm iterates through the various stages and moves the dividing line. We can see that the percent error decreases with the number of epochs (iterations). It will not be able to get to zero in this case because there is no straight line that perfectly divides this particular plot, but it can surely reduce the errors to a minimum.

There, now you know about Logistic Regression, one of the foundational algorithms of machine-learning and deep neural networks. Of course, things start getting much more interesting when we’re no-longer using straight lines to divide the graph and we’re working with full-blown images instead of a few dots on a plot.

Let me know if you’ve learned something by reading this article. Soon we’ll start using these foundational principles and apply them to more complex tasks. Perhaps we’ll even be able to predict the next major market bubble before it bursts. But for now, that’s all!

Posted on

What AI Does Well

AI has become extremely adept at giving suggestions, sorting through huge volumes of data and providing summaries. Right now, I can log onto Google Photos and type any word I want and Google’s image classification algorithm will find me photos that contain whatever I search for. For example, I’m considering selling my 2013 VW Tiguan in order to help pay for another corporate vehicle (that happens to be a Tesla). Anyways, I typed Tiguan into the search bar on Google Photos to find images of the car that I could post online. Sure enough, every photo that I’ve ever taken of my car popped right up, and some photos showed up that had other people’s Tiguans in the background. I have around ten-thousand photos in my library, so finding those few is quite an impressive feat and would have been much more difficult had I tried to do it manually.

Some of the images Google’s AI found for me when I searched the word Tiguan

Most of the improvements in AI over the last 5-15 years have come from developments in a type of machine learning software called deep neural networks. They’re called neural networks because they form analogous structures to the human brain.

Basically, they’re a huge array of neurons (input neurons, output neurons and hidden neurons) connected by lines that represent weights. The connections between the neurons form matrices that modify the subsequent layers of the neural network. It all looks something like this:

Simplified neural network with only one hidden layer – Courtesy Udacity

Typically, deep neural networks have multiple hidden layers (it’s why they’re called ‘deep’ neural networks). What happens in the hidden layers is obstructed from view and it isn’t always obvious what each of the hidden layers is doing. Generally, the hidden layers are performing a simple matrix operation on the input values, the result, weighted by the lines (scalars) connecting the layers, is eventually passed to the output layer. The goal of an image classifier, for example, is to take an input, let’s say an image of a cat, and then produce an output, the word cat. Pretty simple, right?

Well, it kind of is. As long as you know what the input is and what the output should be it is relatively straightforward to ‘train’ a neural network to understand what weights to assign in order to transform a picture of a cat into the word cat. The problem arises when the network encounters something that it didn’t train for. Of course, if all the network has ever seen are picture of cats, if we feed it an image of something else, say, a mouse, the network might be able to tell you it’s not a cat, if it was trained with enough data, but more likely it will just think it’s a weird looking cat. If the network gets constantly rewarded by identifying everything as a cat, it’s probably going to think something is a cat when it sees it.

A neural network acts like a linear function that divides a boundary, in this case, cat vs not cat. Having a neural network with multiple layers allows the lines that can be drawn to be ‘curvier’ and include more cats and fewer dogs.

This is why having a large enough training and testing datasets is critical for neural networks. Neural networks need to train on large quantities of data. Google has billions (perhaps trillions) of photos stored in their servers, so they’ve been able to train their neural networks to be incredibly efficient at determining what is in an image.

In general, problems where there is a large enough training dataset and both the input and the answer are known for the training set are fairly tractable for AI programs today. One task that is generally more difficult for today’s AI software is explaining how and why it got the answer it did. Luckily, researchers and businesses are hard at work solving this problem. Hopefully soon, Google Photos will be able to not only show us images of all the cats in our photo library, but also be able to tell us why they’re so cute and yet so cold all at the same time.

‘Blackbox’ AI happens when a system can provide the correct answer but gives no indication of how it arrived at the solution.
Posted on

Robots Still Can’t Win at Golf

Tiger Woods won his 80th PGA tour title this Sunday, September 23. I was planning to delve deeper into my MIT course on AI and study the details of natural language processing, specifically syntactic parsing and the value of training data. Instead, I found myself glued to a browser window for four and a half hours this afternoon, watching my favourite golfer relive his glory days, winning by 2 shots and capturing the Tour Championship. It was totally worth every minute.

You see, even though humans are being vastly outpaced by AI and machines at every turn, humans are still better at many nuanced tasks. Sure, you can program a robot to swing a golf club and hit repeatable shots, but even the best golf robots still can’t beat the best humans over 18 holes with all the nuanced shots required for a round. Still, they can make a hole-in-one from time to time:

Despite humanity’s increasing incompetence compared to machines, it is still incredibly fun to watch a talented person, who has worked their entire life to perfect their craft, get out there and show the world what they’ve got. Doubly so if that person has recently recovered from spinal fusion surgery and hasn’t won for over five years on tour. Yeah, it’s just putting a little white ball in a hole, but the crowds and excitement that Tiger Woods is able to generate while he plays are unparalleled in golf, and possibly even in sports.

Tiger didn’t win the FedEx Cup, the PGA Tour’s season-long points-based title, but he came really close. If he had, he would’ve made $10,000,000 on the spot. Not too shabby. Regardless, with the highest viewership numbers in the history of the tournament and crowds so large that commentators said they’d never seen anything like it, Tiger Woods undoubtedly made the tour, its sponsors and network partners well over $10 Million this weekend. The amount of value that he generates for the tour and for golf is almost incalculable.

Of course, if you’re not a golf fan, you probably think that it’s boring to watch. That can be said about just about any sport or event that one doesn’t understand. Something is boring to us because we don’t understand the context, the history, and the implications of a certain event happening. Once we understand the subject and can opine and converse with other people about the topic then it becomes much more real and tangible.

I think the same principle applies to artificial intelligence as well as finance. Few understand the topic. It takes time to learn and understand the nuances that make the topics interesting and valuable. Once one does build the knowledge and expertise to apply skills in these areas, the results can be extraordinary. 

So I’m going to pose a question for my future self and any would-be AI experts. In 2 years, will we be able to build software that can perfectly predict the outcome of major events in sports, specifically golf tournaments, with better results than the best human statisticians and algorithms?

Before you say pfft and walk away thinking I’m a complete idiot for saying that, already this year, an AI has perfectly predicted the outcome of the Superbowl. Let that sink in.

It’s going to happen. My hope is that I’m the one building that software.

Posted on

Google is Lightyears Ahead in AI

Google’s Deepmind is incredible. In 2016, AlphaGo (one of the Google Deepmind projects) bested the world champion Go player, Lee Sedol in the 2500-year-old Chinese game that many AI experts thought would not be cracked for at least another decade.

AlphaGo won that match against Sedol 4-1, smashing the belief of experts and Go fanatics around the world who knew that a computer couldn’t yet beat a human. There’s a great Netflix documentary on the feat that chronicles the AlphaGo team’s quest to defeat the grandmasters of the ancient game. It reminded me of when I was younger and I watched IBM’s Deep Blue defeat the Grandmaster, Garry Kasparov at Chess, except this was much scarier.

Deep Blue was able to defeat the best Chess players in the world because chess is a game with a relatively small number of possible moves each turn and a computer can essentially ‘brute force’ calculate all the potential moves that a player could make, and easily counter them. Go is a very different game. With an average of 200 potential moves every turn, the number of possible configurations of a Go board quickly exceeds the number of atoms in the entire universe. There are approximately 10^{78} atoms in the universe. A game of Go can last up to 2.08*10^{170} moves, and at each point has at most 361 legal moves. From the lower bound, the number of possible go games is at least {10^{10}}^{48} , which is much, much larger than the number of atoms in the observable universe. This means that brute-forcing the outcome of a Go match is essentially impossible, even with the most powerful supercomputers in the world at your disposal.

In order to surmount this problem, AlphaGo combines three cutting-edge techniques to predict outcomes and play the game better than any human. Alpha Go combines two neural networks – one conducting deep reinforcement learning and the other using a value network to predict the probability of a favourable outcome, with a Monte Carlo search algorithm that works similarly to how Deep Blue worked. The deep reinforcement learning piece works by having the system play itself over and over again improving the neural network and optimizing the system to win the game by any margin necessary. The value network was then trained on 30 million game positions that the system experienced while playing itself to predict the probability of a positive outcome. Finally, the tree search procedure uses an evaluation function to give a ‘tree’ of possible game moves and select the best one based on the likely outcomes.

The architecture of the system is impressive advanced and its ability to beat humans is impressive, however, the most amazing moment in the Netflix documentary comes when the world champion, Sedol, realizes that Alpha Go is making moves that are so incredible, so creative, that he has never seen a human player even think of playing those moves.

Since beating the 9-dan professional two years ago, Google has since iterated its AlphaGo platform with 2 new generations. The latest generation AlphaGo Zero, beat the iteration of AlphaGo that defeated humanity’s best Go player by a margin of 100 – 0. That’s right, the newest version of AlphaGo destroyed the version of AlphaGo that beat the best human player 100 times over. The craziest thing is, the new version of AlphaGo was not given any directions except for the basic rules of the game. It simply played itself over and over again, millions of times, until it knew how to play better than any human or machine ever created.

Courtesy Google’s DeepMind Blog

This awesome video by Two Minute Papers talks about how Google’s Deep mind has iterated over the past few years and how AlphaGo is now exponentially better than the smartest human players and is trained and runs on hardware equivalent to an average laptop.

Courtesy Google’s DeepMind Blog

It is scary and incredible how fast DeepMind is advancing in multiple fields. Right now it is still considered a narrow AI (i.e. it doesn’t perform or learn well in areas outside of its primary domain), however, the same algorithms are being applied to best the greatest humans in every area from medicine to video gaming. In my opinion, it is only a matter of a few years before this system will be able to beat humans at nearly everything that we think we do well right now. We can only hope that Google built in the appropriate safeguards to protect us from its own software.

If you want to learn more about how our future robot overlords think, there’s no better way to get started than by racing an autonomous robocar around a track. Come check out @DIYRobocarsCanada on Instagram and join us on our Meetup group to get involved today!

Posted on

Natural Language Processing – I’m Lookin’ at you, Siri

According to some experts, in 2018, natural language and audio processing (NLP for short) are the two areas that AI excels most at. It seems fairly obvious that machine audio processing would be good when you look at the number of US homes with a smart speaker device like Amazon’s Echo, Apple Home Pod, or Google Home (nearly 50%). That’s a lot of people with a device whose sole purpose is to collect data on what you’re saying and translate it into usable commands for you, and into monetizable data for the company that builds them. 

“Hey, Siri!”

— me, just now

In fact, if you read that out loud right now, there is a good chance that your phone just pinged at you wondering what you want from it. With so many devices listening constantly to our every conversation and vocal machination, is it any wonder that they’re starting to understand us better than we understand ourselves?

In my home, I have Apple Home, and by proxy, Siri, set up to control everything from my locks to my lights, smoke alarm, security system and thermostat. If Siri ever became sentient and went Skynet on my ass, she could do A LOT of damage. Although, based on the latest tests, I’m much less worried about Siri taking over than I am about Google. Or more specifically Google’s Duplex AI. It can literally trick humans working at shops into believing that it is a person. If that’s not scarily impressive, I don’t know what is.

Are we doomed? Probably not… Yet. Duplex is still very limited according to Google. All it can do is book appointments for you. But people got so scared that they’d be talking to a machine and not know it that the outcry forced Google to hamper its own system. Google has announced that Duplex will now announce itself as a computer at the beginning of every phone conversation so as not to creep people out. To me, this kind of defeats the purpose of the software to begin with. But hey, it’s probably bad form to test your AI on a bunch of unsuspecting hair salon and restaurant employees without them knowing about it.

This software is getting really good at understanding and communicating with humans via voice. This is mainly because it has a TON of really good data collected by millions of devices around the world. It’s also because engineers, mathematicians, and programmers have made some serious breakthroughs in Deep Learning in the past 10 years and Apple and Google have tens of billions of dollars invested into making it work.

This weekend I’ll be learning from MIT Professor Regina Barzilay about what it means for machines to understand something. We’ll also be covering which NLP problems have been solved, where progress is being made and tasks that are still very difficult for computers to solve. Hopefully, I’ll come back with a better understanding of how it all works (I think it has something to do with phonemes and triphones) and what we can do with it. Once I do, I’ll report back here and let you know how close we are to the robopocalypse.  

Posted on

Matrix Madness

Linear algebra is an important tool used in modern deep learning algorithms. Unfortunately, when I did my undergrad in Electrical and Computer Engineering, I had no idea that the ability to transform vectors and matrices would ever be practicably useful for anything (other than giving me migraines at 2AM the night before my midterm exams). So, once I had learned enough to pass the course, I immediately forgot everything.

It was only when I decided to pursue a deeper understanding of machine learning and AI, in order to apply it to my business and to my work in finance, that it dawned on me. I should have paid attention in Linear Algebra II when I was back at the UofA in Engineering school! Well, since I didn’t, and even if I did, all my books on the subject mysteriously burned up in the great notes fire of ’09, I guess it’s time to re-learn me some matrix math.

Lucky for me, Udacity has brought on some of the best professional educators in the world for their AI Nanodegree program including former Khan Academy animator and 3 Blue 1 Brown creator, Grant Sanderson.

So, now I’m super good at manipulating matrices thanks to the magic of a YouTube superstar and the LaTeX plugin for WordPress websites.

A = \begin{bmatrix} a_{11} & \cdots & a_{1j} & \cdots & a_{1n} \\ \vdots & \ddots & \vdots & \ddots & \vdots \\ a_{i1} & \cdots & a_{ij} & \cdots & a_{in} \\ \vdots & \ddots & \vdots & \ddots & \vdots \\ a_{m1} & \cdots & a_{mj} & \cdots & a_{mn} \end{bmatrix}

…I spent 30 minutes trying to get this matrix to display correctly.

If you want to learn why vectors are cool and how to use matrix multiplication to rule the world, watch the video series that I’ve linked below.

We haven’t arrived at the part about why Linear Algebra is so important for creating neural networks and deep learning algorithms, but we will. If you’re still with me, keep plugging along, learn how to understand all matrix transformations using only the unit vector and a 2×2 matrix. Eventually, we’ll discover how to program a computer to predict Apple’s  share price the day after they launch a ‘new’ iPhone

/insert corny iPhone XS Max joke here/.

  • Loading stock data...

That’s all for today. Now go out on your own and learn the basics of linear algebra! If you message me directly, I’ll even send you my notes. Tomorrow, we’ll talk about why this stuff is important for machine learning.

Linear algebra joke: One year for halloween I was ‘Snakes on a Plane’

If you’re still here and you’re wondering why yesterday we were talking about valuing a company and today we’re talking about undergraduate linear algebra, you’re probably not alone. So I’ll tell you why: It’s going to take a foundational understanding of programming, mathematics and finance to get where we need to go. To understand machine learning, we have to understand how the software is built and to build software that is capable of doing what a CFA can do, we’ll need to know what a CFA knows. I’m bringing you along on this journey as I learn the fundamentals on both ends, machine learning and finance. Let’s see how it goes!

Do you get it now?