Niels Bohr is quoted as saying “prediction is difficult, especially about the future”. This seemingly tautological statement hints at the point in the scientific process whereby hypotheses are formed which make predictions about measureable properties of the world. Of course, these properties were already there when the hypothesis was made so it is a prediction about the present. Challenging enough already, so imagine how much harder it is to predict an ever-evolving future.
Prediction was a key topic at the Neural Information Processing Systems (NIPS) conference which hosted a record 6000 delegates last week to discuss and learn about the latest trends in the rapidly developing field of machine learning. One of the more striking trends, discussed by influential keynote speakers from Facebook and DeepMind, was the graduation of machine learning systems from predicting the present to predicting the future.
This may not seem entirely new, but in many ways it is. Predictive analytics is popular industry buzz phrase and is typically used to describe classification of unknown data points into categories based on comparison with a large set of known data points. This is best interpreted as a prediction of the present, answering questions like:
- Does this image contain a dog/cat? (A popular toy problem because available online data sets contain a lot of images of dogs and cats but you probably already knew that).
- Is this email spam?
- Is this person credit-worthy?
The last example looks quite a lot like a prediction of the future if you think of it as asking the question “will this person default on a loan?” However, practically it is more like asking the question “does this person look more like people who have historically paid back their loans or people who defaulted?”.
The aspiration, articulated in keynote speeches by Yann LeCun of Facebook and Drew Purves of DeepMind, was a step change in the ambition for what is possible using modern deep learning methods. The ambition is genuine prediction of the future by learning enough about the structure of the underlying systems of interest to be able to predict things not explicitly seen in the data before.
The use cases showcased so far indicate tentative progress through some interesting examples:
- Predicting the future frames of a video – A compelling example showed a camera panning across a New York apartment and making a plausible guess about what the unseen furniture might look like.
- Predict the evolution of a dynamical system without explicit knowledge of physical laws – For example, which way will a teetering tower of blocks fall?
There remain significant challenges. Learning the structure of the rich and complex systems that are of interest requires access to appropriately rich and complex data sets. This is a particular challenge due to the voracious appetite of deep learning algorithms for data.
It seems likely that data from simulations as well as traditionally collected data will form part of the diet on which these algorithms will feed. This raises the need for more ambitious simulation frameworks. Leading the way in terms of complex simulations, DeepMind has recently open-sourced its simulation framework DeepMind Lab, which has a focus on simulating complex naturalistic environments from which artificial intelligence agents can learn.
Biology has inspired the neural network structure of deep learning algorithms; there may be more still to learn from nature.