Beta Solutions Blog

Figure 1. Simple line fit to known sample data. At the heart of AI lies this same data-fitting principle.

AI solves a problem much like this, except with a fair bit more calculation than you would have used in high school. AI is tuned with potentially millions of data points, where each point could have many inputs and many outputs. For example, an input may be an image with each pixel effectively being an input. An output from this could be a simple yes/no decision for a question such as "Is there a person in this image?", or it could be multiple outputs for coordinates for each person found in the image. AI finds a solution model that best fits the original data set, with the model potentially using thousands of coefficients. AI can then use this model to predict the result of any future inputs. The output of a model is usually a probability rather than a direct yes or no answer.

Modern advances into "deep learning" have shifted complex problems into neural networks. Each node has its own function and solution, and the output of one node is then fed into further nodes. Splitting the problem into nodes allows it to be optimised in parallel using modern GPUs or cloud based server farms. There are currently very many architectures available for deep learning, some of which are pictured below.

Figure 2. Example neural network architecture. In between the inputs and outputs lie multiple hidden layers.

Training AI and Machine Learning

Since the early days of AI, board games have been a simple but effective way of determining AI "intelligence". One of the first AIs in the early 60s was, in fact, a checkers program that managed to beat experienced amateurs. Today AI is pretty much unbeatable in games like this and is able to teach itself games at astonishing rates. An example of this is Google's AlphaZero chess engine, which trained itself to play chess in under 4 hours to a superhuman level by simply playing itself over and over^[1]. Given more time, a 9-hour trained engine was able to defeat the most notable engine of the time, Stockfish 8.

Training and machine learning is at the heart of AI. AI needs a lot of previously categorised data to learn from. For board games this is relatively simple, as the engine can effectively play itself and learn along the way. There is no consequence to losing a board game. Clearly this is not the correct approach for all scenarios. In a fully autonomous vehicle mistakes are critical, and needless to say a trial and error approach to learning driving could be deadly. An autonomous vehicle AI actually learns from a human driver as an example. For an AI to be effective at driving it requires thousands of hours of observed inputs.

Extensive training is the reason why AI might not be applicable to all problems. Facebook and Google have access to millions of images, data points, etc. to train their AI algorithms. Obtaining thousands of data points and categorising these for your needs is very difficult and time consuming for a smaller company that's just starting out. However, as long as the function is non-critical, an AI system may be able to train itself over time, given a pre-built model of similar data.

Figure 3. A realistic picture on what AI is all about. A large part of AI "coding" lies in getting and categorising sufficient training data.

The Magic of AI

There are by now several examples of AI being used in medicine to predict patient current and future ailments. A prominent example of this is Deep Patient, an AI which has learned to predict conditions like liver cancer and is also strangely good at diagnosing complex disorders like schizophrenia. It seems to do this far better than most doctors can. How does it do this though? As of today, the answer is, no one knows. Deep Patient, much like most AIs, is a complex network of neural pathways that have been tuned through the use of a huge data set. The output is currently a result of optimised math that makes no sense to someone looking at the final system and is a prime example of one of the issues with AI. Although an AI has found a solution to the problem, we currently have no concrete way of verifying the end result's reliability, nor of any prejudices in the system due to biased data.

Google's Ali Rahimi, an award winner at the 2017 Conference on Neural Information Processing (NIPS)^[4], likened the current state of AI to the medieval magic of alchemy. Some things just seem to work. In medieval times these examples were metallurgy and glass-making. The processes around these examples worked like a charm when using the scientific knowledge of alchemy at the time, whereas other things like turning common metals into gold simply didn't. AI is much the same at the moment, where some architectures, processes and optimisations seem to work for specific problems, whereas other attempts at solving problems though AI do not. There is currently no full understanding of the processes that underlie machine learning. For alchemy, we now understand that magic is not the right explanation, but we still need to fully figure out the mysteries of AI.

An interesting attempt at understanding the inner workings of AI has been made by Google in its Deep Dream project^[2], which I'd encourage the reader to check out. It is a tool for finding things like faces in images, but can also be run in reverse to see what an optimised and biased AI might be looking for. The result is strange images that are now classified as a new art form.