Whenever there is a new announcement or breakthrough with AI, it always strikes me how out of reach the results would be to replicate for individuals and small organizations. Machine learning algorithms, and especially deep learning with neural networks, are often so computationally expensive that they are infeasible to run without immense computing power.
As an example, OpenAI Five (OpenAI’s Dota 2 playing bot) used 128,000 CPUs and 256 GPUs which trained continuously for several months:
In total, the current version of OpenAI Five has consumed 800 petaflop/s-days and experienced about 45,000 years of Dota self-play over 10 realtime months.
Running a collection of more than a hundred thousand CPUs and hundreds of GPUs for ten months would cost several million dollars without discounts. Needless to say, a hobbyist such as myself would never be able to replicate those results. Cutting edge AI research like this has an implicit disclaimer: “Don’t try this at home”.
Even on a smaller scale, it is not always possible to run machine learning algorithms without certain trade-offs. I can sort a list of a million numbers in less than a second, and even re-compile a fairly complex web application in a few seconds, but training a lyrics-generating neural network on less than three thousand songs takes several hours to complete.
Although a comparison between number sorting and machine learning seems a bit silly, I wonder if we will ever see a huge reduction in computational complexity, similar to going from an algorithm like bubble sort to quicksort.1
Perhaps it is not fair to expect to be able to replicate the results of a cutting edge research institution such as OpenAI. Dota 2 is a very complex game, and reinforcement learning is an area of research that is developing fast. But even OpenAI acknowledges that recent improvements to their OpenAI Five bot are primarily due to increases in available computing power:
OpenAI Five’s victories on Saturday, as compared to its losses at The International 2018, are due to a major change: 8x more training compute. In many previous phases of the project, we’d drive further progress by increasing our training scale.
It feels slightly unnerving to see that the potential AI technologies of the future are currently only within reach of a few companies with access to near-unlimited resources. On the other hand, the fact that we need to throw so many computers at mastering a game like Dota should be comforting for those with gloomy visions of the future :-)
Automatic, machine-generated music has been a small interest of mine for some time now. A few days ago, I tried out a deep learning approach for generating music… and failed miserably. Here’s the story about my efforts so far, and how computational complexity killed the post-rock.
The spark of an idea
When Photo Amaze was created in 2014, I thought it would be fun to have some kind of ambient music playing while navigating through the 3D maze. But I did not want to play pre-recorded music. I wanted it to be automatically generated on-the-fly, based on the contents of the pictures in the maze.
That was the spark. A picture is worth a thousand words, so why can’t it be worth a few seconds of music as well? For example, take a look at this picture:
motional impact, like the ambient sound of a running water stream or the whistle of the wind picking up speed over the mountain.
Since I can make a connection between photo and music, perhaps a machine could do this automatically as well. This is not a novel idea, but it is a nut that has yet to be cracked, and it was an intriguing idea to start exploring.
Hard-coded music mappings
The first experiment I did was to create a more-or-less fixed mapping between an image’s content and some kind of sound output. The high-level idea of the implementation was to simply map brighter colors to brighter sound notes. The steps to produce the output sound were something like this:
Find at most 200 “feature pixels” in the input image using trackingjs.
For each found “feature pixel”:
Calculate the average pixel value between the three RGB color values. This produces a single number per pixel between 0 and 255.
Normalize the pixel value from 0-255 to 20-500. This produces a base frequency for the output sound.
Create a sine wave oscillator using the Web Audio API for the pixel value.
Combine the oscillators to a single sound output.
While playing the sound, randomize the frequency of each oscillator slightly over time.
Using this approach, an image would be turned into a randomly changing output sound consisting of about 200 sine waves, each with a frequency between 20 and 500 Hz.
Here is an example output using the mountain from above as the input image (the red dots mark the found “features” of the image).
That might not sound terrible, until you realize that the sound is basically the same for any input image:
Mila might be a monster dog, but that output is just too dark :-)
There were a ton of problems with this implementation, to the point that it was actually outright silly. For example, the “feature pixel” selection mostly found edges/corners of the image, and using just 100 pixels as input corresponded to just 0.01% of all available pixels in the test images. Another problem was how the final pixel value was calculated from the average between the red, green and blue value of the pixel. Some colors arguably have more impact to the viewer than others, but this fact is not captured when taking the average.
Even with all its problems, the first experiment was a good first step, considering I did not know where to start before. It is possible that with a lot of tweaks, lots of new ideas and lots of time, this approach could start producing more interesting soundscapes. However, the downside of the approach was also that the music creation would always be guided by the experimenters: the humans. And I wanted to remove them from the equation.
Machine learning to the rescue
The second experiment ended before it even really started. It was clear that some kind of machine learning was needed to move forward, and it seemed that an artificial neural network might be the solution.
This was the idea:
Use every pixel of the input image as a single input node of the neural network.
Treat every output node as a single sound sample.
For the purposes of this blog post, everything that happens between input and output nodes of the network is largely hidden magic. With that in mind, here is how the network would look (P1 – Pm are the input pixels and S1 – Sn are the output samples):
To get an idea of the size of the network, consider this: the mountain test image from above is 1024 by 683 pixels, so the network would have 699,392 input nodes when using images of that size. Digital sound is just a collection of amplitudes in very tiny pieces called samples. The most commonly used sampling rate for music is 44.1KHz, which means that every second of digital music consists of 44,100 individual samples. For a neural network of this design to produce a five-second sound, it would thus require 220,500 output nodes.
The intentions were good but the implementation never happened. After having the initial idea, I started Python and tried to simply read and write soundfiles, but it didn’t go so well, and the weekend was nearly over, and, “oh a squirrel!”… and the code was never touched again.
Machine learning is great, but the motivation was suddenly lacking, and the project was put on ice. This was about two years ago, and the project was not revived until quite recently.
Deep learning has been steadily on the rise in recent years, often outperforming other machine learning techniques in specific areas such as voice recognition, language translation and image analysis. But deep learning is not limited to “practical” use cases. It has also been used to create art.
The A.I. Duet project shows another interesting use case for deep learning: the creator of the project, Yotam Mann, trained a model that can produce short sequences of piano notes based on the note input of a human. So if I played C-D-E, the software might respond with F-G-A although the result would most likely be slightly more interesting than that. 1
A.I. Duet is impressive, but it still has a big limitation: it only works for specific notes for a specific instrument. So while the result is amazing, what I really want is more complex arrangements and raw audio output. Even so, the above examples show that deep learning is a powerful and versatile machine learning technique, and it is now finally becoming more feasible than ever to achieve the goal of creating music using AI.
The bleeding edge, where the story ends
While doing some research on the latest state of the art for machine-generated sound, I stumbled upon yet another Google project called WaveNet. In an interesting blog post, the authors of WaveNet discuss how their research can be used to improve text-to-speech quality, but what is really exciting to me is that they also managed to produce short piano sequences that sound natural (there are some examples at the bottom of their blog post).
The big surprise here is that the piano samples are not just based on specific notes. They are raw audio samples generated from a model trained with actual piano music. 2
Finally! A tried and tested machine learning technique that produced raw audio. Reading about WaveNet marked the beginning of my final experiment with music generation, and is the entire reason this blog post exists.
I found an open source implementation of WaveNet, and to test the implementation, I wanted to start simple by using just one sound clip. For this purpose, I extracted an eight-second guitar intro from the post-rock track Ledge by Seas of Years3:
My hope was that by training the model with a single sound clip, I would be able to reproduce the same or a very similar clip to the original to validate that the model produced at least some sound. If this was possible, I would be able to train the model with more sound clips and see what happens.
Unfortunately, even with various tweaks to the network parameters, I could not manage to produce anything other than noise. Sharing an example of the output here is not even appropriate, because it would hurt your ears. The experiment ended with an early failure.
So what was the problem? I soon realized that even with this fairly simple example, I had been overly optimistic about the speed at which I would be able to train the model. I thought that I could train the network in just a few minutes, but the reality was very different.
The first warning sign showed itself pretty quickly: every single step of the training process took more than 30 seconds to complete. In the beginning, I did not think much about this. Some machine learning models actually start producing decent results within the first few steps of training so I was hoping it would be the same here. However, after doing more research on WaveNet, it became clear that training a WaveNet model did not just require a few learning steps, it required several thousand. Clearly, training WaveNet on my machine was completely unfeasible, unless I was willing to wait more than a month for any kind of result.
Where do we go from here?
Machine learning has been rapidly evolving in recent years, propelled by software libraries like TensorFlow, and the technology is more accessible than ever for all kinds of developers. But there is also another side of the coin: in order to use the state of the art, we are often required to have massive amounts of computing power at our disposal. This is probably why a lot of high-profile AI research and projects are produced by companies like Google, Microsoft and IBM, because they have the capacity to run machine learning at a massive scale. For lone developers like me that just want to test the waters, it can be difficult to get very far because of the complexities of scale.
As a final example to illustrate this point, consider NSynth, an open source TensorFlow model for raw audio synthesis. It is based on WaveNet and on NSynth’s project page, it says:
The WaveNet model takes around 10 days on 32 K40 gpus (synchronous) to converge at ~200k iterations.
Training a model like that would cost more than $5,000 using Google Cloud resources 4. Of course, it is possible that a simpler model could be trained faster and cheaper, but the example still shows that some technologies are most definitely not available for everyone. We live in a time where there is great access to many technological advances, but the availability is often limited in practice, because of the scale at which the technologies need to operate.
So where do we go from here? Well, computational complexity killed my AI post-rock for now, but I doubt that it will take long before significant progress is made in this field. For now, I will enjoy listening to human-generated music. In a way, it is re-assuring that machines cannot outperform us in everything yet.
I had an interesting chat with Intercom support about what I perceived to be a security and privacy hole in their support messenger app, but it turned out that what I thought should be a great concern for them was happening “by design”.
Intercom is a popular customer relations tool, and one of their cool features is the chat messenger app. It adds a little chat icon to the bottom-right of a website and allows real-time chat with customers for help and support. We use it at Receiptful which allows us to chat directly with our users when they are signed-in to our app. It looks like this:
Chats are not private
A few days ago, I was using the Intercom chat app on a website that hosts some of our data. I needed to update some basic settings for our account and asked for help using the Intercom chat while I was signed-in to the service. A common use case for the Intercom chat is to allow support for both anonymous and signed-in users. What I found out is that there is no distinction between these by default.
When I signed out from the website, I noticed that my private chat session was still visible in the “anonymous” chat window. Even after restarting the browser and without signing in to the service, my private chat session was visible.
In other words: If I was on a shared computer, the next person using the browser would be able to see my private chat sessions, even though I signed out from the service where I had the chat in the first place.
Next, I tried to do the same thing on the Intercom website and it was the same deal: All previous announcements and private chats were visible from their frontpage without me signing in:
“This is, in fact, by design”
When I noticed that my private support chats were leaking into the anonymous part of their website, I reported it to Intercom as a possible security hole because I did not think that it was intentional that private chats were visible while being signed out. This is the response from Intercom support:
This is, in fact, by design. We track users using an anonymous cookie, and when they logout that cookie still exists, so we can use that to keep the conversations in the messenger. I think your concern though is interesting, and I’ll forward this as feedback to our Messenger team.
If you’d like to ensure that others won’t see the conversations, I recommend clearing your cookies with us after logging out.
Apologies for the confusion there, it’s clear that sometimes what we think is a good idea isn’t always agreed upon by others.
So the privacy leak is “by design” and I have to remember to clear all my cookies to avoid it. What a joke. Imagine having a private chat on Facebook that was still visible after signing out. That would be quite horrible. Intercom clearly does not see their support chat system as a private conversation, although it most certainly is. In the chats, both my real name and email are used and what is even worse: I can create a new conversation using the same chat window, thereby impersonating whoever was the last one to use the system.
Now to be fair, there is a documented API called Intercom('shutdown') which clears the user cookie and resets the state of Intercom. However, Intercom does not even use this API themselves and I cannot imagine many websites that do this. So leaking chats are probably quite common.
The bigger picture
I think what really bothered me is that I already knew what Intercom would say when I reported the issue. Before I got the above answer from Intercom, I wrote this message to my colleagues:
The problem with lack of privacy is systemic. In this case with Intercom, usability won over privacy. They thought it was a “good idea” to keep chat windows open even after the user had signed out of their service and in most cases, this decision does not present a problem for the user if they are not on a shared computer. But by asking the questions “should private chats be visible after the user signs out”, “what if the user is on a shared computer” and “how does this relate to the privacy of our users”, I think they would have arrived at a different conclusion.
As developers in an a world of increasing surveillance, we need to ask ourselves questions about privacy when developing our solutions. And if there is an obvious case of private information leaking to a non-secured area, we should most definitely not consider it to be “by design”.