Artificial intelligence will be able to predict our health thanks to a change in the concept of artificial neural networks, which will allow categorizing information and identifying patterns in a more efficient way.
The department of artificial intelligence of the University of Toronto tries to create a system of deep learning that can predict a patient’s health. However, the systems that store medical records are not prepared to provide the information needed to create the proposed breakthrough.
People visit the doctor for different reasons and at different times in their lives, this creates arbitrary data that artificial neural networks have trouble categorizing.
Duvenaud and his collaborators, from the Vector Institute, propose to redesign these neural networks to allow the IA develop prototypes, as proposed by the University of Toronto.
A traditional neural network is built by layers that work together to look for patterns in the data provided to them. Duvenaud’s proposal destroys these layers. To understand what this process entails, we must first understand how artificial neural networks work.
Neural networks are the foundation of deep learning
The most common process for training an artificial neural network is archive piles of tagged data. That is, if we wanted to create a system that recognizes animals, what we would have to do is feed a neural network with photos of animals that are synchronize with their corresponding names. Once all this data is synchronized, the formulas created to solve this puzzle they can be used again and again to categorize new animals.
However, a single formula covering the entire photo-name process is too broad, so it would generate inaccurate results. It would be like using a single formula to differentiate dogs and cats, you could create a formula based on their size or their ears, but this would create false positives and negatives, as there is no general rule that differentiates them.
This is where the layers of the neural network take importance. These layers separate the process into different steps. For example, the first layer could collect all the pixels and, based on a formula, select the ones that are most relevant for cats and dogs. The next layer could construct patterns extracted from these groups of pixels, and decipher which images have long whiskers or ears. The following layers will progressively identify more complex characteristics of the animals, until the last layer will decide based on the accumulation of patterns. This process, step by step, enables neural networks to create more accurate predictions.
The concept of layers has been of great help to the field of AI. Although it has also been a step backwards, since if you are looking for a model that continuously transforms, it will have to be divided into different steps. Going back to the medical history example, that would mean grouping medical records into temporary periods like months or years, which would be very inaccurate. Therefore, the best way to approach reality is to divide the records into more specific units, such as days or even hours. This division taken to the extreme would be the best neural network, with an infinite number of layers to register infinitesimal different options.
Is this idea practical?
If this starts to sound familiar, it’s because this problem tried to be solved in the past with the estimate. The hated calculation provides great equations to solve these infinitesimal options. In short, the calculation was developed to avoid the nightmare of creating units that were in constant change, that is, to avoid what artificial neural networks suppose.
But the magic of Duvenaud and all his collaborators resolve this conflict in the most obvious way. Duvenaud has proposed replace neural layers with calculus equations. This proposal would eliminate the network system to make it a block called ODE (Ordinary differential equations).
It is quite difficult to understand this concept, but Duvenaud uses an analogy so that those less knowledgeable in the matter can understand what this advance will mean. Duvenaud considers an instrument like a violin, with which you can slide your hand over the strings to play the different notes; now we consider one like the piano, where there is a specific number of keys to play the notes. A traditional artificial neural network is like a piano.
ODE would change the system to a violin. In this new system it is not necessary to specify a specific number of layers (keys) at the beginning of the process. ODE allows to establish a level of accuracy; and the model will find the most efficient way to train, based on that margin of error that has been established at the beginning.
The concept is not yet ready to be applied
One disadvantage is that artificial neural networks let you know how long it will take you to train to meet your goals, something that would not be possible with the new ODE system.
Duvenaud’s work is proof of the validity of the concept which, unfortunately, not ready to be applied yet. It is a proposal that must be tested with different experiments and confirmed when it can begin to be used.
It is believed that the most efficient application of this system will be, for the time being, timeline-based systems such as health systems.