In our society today, the most infamous disease has to be cancer. It is incredibly difficult to find a person who has not been affected by this disease in some way, either by contracting it themselves, or knowing someone who has. Though the percentage of deaths associated with cancer has been falling, the incidence or number of cases each year has not. What this means is that we’ve gotten better at keeping people alive, but not necessarily preventing people from getting cancer.
There are a variety of reasons for this, but the main point of the matter is that cancer is complicated. Start with the simple fact that when we say “cancer” we’re actually not specifying which cancer. Also the general perception of cancer is simply that there is a type of cell which isn’t working properly, but recent studies have shown that even this claim is too general as there are many types of cancer cells and also surrounding non-cancer cells which contribute to a single cancer tumor in varying degrees. This makes it incredibly difficult to treat cancer and even get drugs to the tumor. If you want to read more about this complexity, see the article below.
So How do We Handle Complexity?
Unless you’ve become a 21st Century hermit, the terms “Artificial Intelligence (AI)”, “Machine Learning”, and “Deep Learning” should ring a bell. Simply speaking, AI is the broad term to describe devices which behave in an “intelligent” manner. Most people think robots but this applies to broader devices and algorithms such as that Google Search you so often use or the suggestions your keyboard gives you when you text someone. In general you can think of AI as the idea of giving a non-human device human-level intelligence.
This idea began shortly after WWII, circa 1950s, in which the work of Alan Turing, most notably in decoding the German Enigma machine, demonstrated the ability of computing devices to achieve highly complicated mathematical tasks. The Turing Test, which is a test to see if AI has become indistinguishable from human intelligence, is also named after him.
Under the umbrella of AI is Machine Learning. This term arose out of the 1980s and 1990s at the advent of the computing and internet boom. Usually this is characterized by the algorithms developed to achieve some semblance of artificial intelligence. In other words, Machine Learning was the beginning of actually making AI happen. We got things that could start recognizing letters and images on their own, albeit with a lot of time and a lot of data. However, these were heavily reliant on the engineers to choose the correct algorithm and provide good data. From this came that pesky online chess computer which always beats you and which only the highest level chess masters could defeat. At that point in time it was pretty remarkable but the computer hardware to make it practical for more complicated tasks didn’t quite exist yet. The biggest issue lay in the lack of data that existed and the fact that computers weren’t that powerful … yet.
Enter 2012, which was not all that long ago. This is about the time that Artificial Neural Networks (ANNs) became all the craze. Now it’s important to note that the concept of ANNs had been around since the early days of AI and that it is inherently rooted in biology. Basically the human brain is the best thing we have in handling complexity. The functional unit that makes a brain work is the neuron. Each neuron connects to many other different neurons and sends signals simultaneously back and forth so that you can remember things and be you. ANNs were inspired by this biology but instead of having each artificial neuron link to hundreds of others, it was heavily simplified where the neuron networks had discrete layers and had information flow, for the most part, in a single direction. However, for years the AI community shunned ANNs as they did not seem to achieve good results and on top of that, it was computationally intensive! We simply did not have the computing power to make anything but the smallest ANNs work in a reasonable amount of time. There also just wasn’t enough data around to train these networks.
So what happened around 2012? Basically this:
- The computing power we could put into a single device exponentially skyrocketed. This has only continued every year so at this point in 2017 we’re sitting at around 18–19 billion transistors in a single device.
2. Not only that, but our ability to store data also increased greatly. We began to hit the Terabyte territory of data storage and the transition to Solid-State Drives (SSDs) as opposed to optical Hard Disk Drives (HDDs) sped up recall times.
3. The price of this hardware also got cheaper, meaning that an average person could own more computing power than the entire Apollo Space Program had at its disposal.
An intrepid engineer at Google by the name of Andrew Ng, made the leap for deep learning at a point in time where the hardware components that had been needed, but missing, for Neural Nets to function efficiently finally existed. He created a massive ANN with many layers (this is where the term deep comes from) and fed an absurd amount of data into it and it worked.
The Deep Neural Network was able to teach itself how to recognize what was a cat based off of millions of YouTube videos. Soon it could distinguish things like different street signs and Facebook was able to utilize it to recognize your face with 97% accuracy. Since then DeepBlue and AlphaGo have defeated the best Chess and Go players in the world. It’s gotten even more ridiculous as AI can now defeat human players in complex video games such as Dota 2.
So What About Cancer?
Now that we also have the ability to acquire genomic data on the DNA and gene expressions of different cancer tumors (which is another discussion you can read more about here), the U.S. Government has started The Cancer Genome Atlas which aims to utilize & leverage the vast amount of cancer information by making the data public for researchers to sift through (while still maintaining HIPAA).
Since Deep Learning can teach computers to recognize and characterize features in a set of data, it is very possible that we can apply it to cancer data to learn more about who is most susceptible to cancer, what could be indicators that a cancerous tumor is forming, and even predict which genes if mutated could catalyze a tumor to form.
In a way, we could indirectly use biology to solve biology. The human brain is the best computer we have. It is energy efficient, highly adaptable, and can recognize trends and patterns instantaneously. It is a three-dimensional network of neurons with millions of connections, whereas computers are still, for the most part, two-dimensional planes of transistors. However computers are much better at handling and storing large sets of non-linearly related data and this is the growing problem with cancer data.
Each new discovery in cancer introduces another layer of complexity within a massive web of relationships. We know that DNA is important to cancer but it depends on which genes are expressed. We know that cancer cells are highly affected by their environment but also that this environment is constantly changing because the cancer cells themselves change the surrounding tissue. We know that cancer cells can induce other types of cells to also become cancerous, but that using one treatment to kill the first cancer cell line doesn’t always work on the other types of cancer cell lines. The list goes on but the one tool we need to figure out cancer may lie between our own two ears.
It is becoming increasingly clear that finding solutions and treatments for something like cancer will require a combined cross-disciplinary effort between medical professionals, engineers, computer scientists, and policy makers. With many of these individuals already utilizing the biological “computers” they were born with to solve this problem, we may very soon see the artificial brains joining the effort.
For a more detailed timeline of Machine Learning see here.
For information about neural research I recommend Dr. Brewer’s Neural Coding work here.
For a detailed description of the mathematics behind Neural Networks I recommend this free online book here.