How do Machines Learn? The limits of Artificial Intelligence (AI)
Popular hyperbole might lead us to believe that it is only a matter of time before AI surpasses human intelligence and the role of knowledge will pass beyond the reach of humans into an opaque world of machine computation. This is a disturbing prospect, but is it really true?
It has been suggested that human intelligence and human reasoning are deficient and inadequate, with poor learning mechanisms evolved in environments ill-suited to the modern world. The proposition is that machine intelligence can escape these deficiencies and lead to the creation of unlimited reasoning power far exceeding the abilities of the human mind.
But there is no evidence today of machines exceeding humans in their ability to reason and acquire knowledge. In reality the task of training machine learning systems is largely an exercise in human intuition. From cleaning and preparation of data, classification for supervised learning, training set selection, dimensionality reduction, feature extraction, algorithm choice, through to scoring, tuning and retraining, the process presupposes a certain expectation of what the “correct” answer should look like. It is not surprising therefore to find that the results approach but do not exceed that of a human expert.
Whilst big data brute force machine learning certainly holds the prospect of assisting humans in processing data and decision making, and has been successful in specific domains, it is not clear that this represents progress towards the goal of an autonomous machine intelligence.
Wolpert(‘No Free Lunch’ theorems 1997) has shown that without prior constraints any learning algorithm is as good as any other when averaged across all similar problems. Given the choice of any possible algorithm, it will always be possible to contrive a model that fits any particular set of data. This is the danger of applied statistical analysis; without a pre-selected hypothesis it always possible to find a correlation somewhere against some part of the data.
Problems with a universal artificial general intelligence (AGI) are not therefore technical, they are philosophical and relate to the question, what is learning? How do we learn? Can we learn more than we are taught? What is the relationship between learning and knowledge?
A data model may contain valid abstractions and show interesting correlations, useful for data visualisation or for data analytics; it may even have testable predictive power over observed data. But this is not enough to say we have acquired knowledge. Human knowledge implies hypothesis and hypothesis requires a causal model. Causality may be a fiction but is a necessary fiction which enables us to make a coherent and useful theoretical model of the world.
Causal models are what allow humans to build bridges, invent non-stick frying pans, fly to the moon, discover gravitational waves, postulate the existence of the Higgs boson and devise experiments to find it. These are not just abstractions but causal theoretical models refined by experiment and tested against the real world.
Until we understand the mechanisms by which human knowledge grows we cannot address the question of whether an AI system could eventually exceed the abilities of human intelligence, or what that system would need to look like in order to achieve those abilities.
To progress with the goal of developing autonomous learning machines we need to look at how humans acquire knowledge.
The brain* has enormous capability for pattern detection and pattern matching of sensory data and its multi-layered highly connected structure enables it to develop higher and higher hierarchical levels of abstraction, which ultimately form the conceptual basis for artistic and scientific thinking. This pattern detection and abstraction is not passive but is actively driven by a natural curiosity to seek for a ‘better’ model of the world – more elegant, more powerful, more profound.
Although there is evidence of supervised learning, where pre-conditioned groups of neurones may provide control signals to learning elsewhere, unsupervised and reinforcement learning appear to be the main mechanisms for learning in the brain**. The brain has the ability to construct an inference model, make predictions, compare with observed data, calculate discrepancies and iteratively improve on its match between prediction and observation.
Further, the brain is designed to remove connections that are not reinforced, leading to controlled ‘forgetting’, which helps to eradicate transient correlations and prune incorrect models which do not stand the test of time or are not useful. (Machine systems conversely are subject to catastrophic forgetting, whereby the learning of a new task destroys learned knowledge about a previous task).
Finally, it is evident that the brain has mechanisms that allow it to induce causal models from observed data and test these models by experiment.***
The power and sophistication of modern machine learning systems is increasing all the time. Whilst recognising their achievements, we should not think of these systems operating without the intervention of humans. A machine algorithm that incorrectly categorises some parts of a solution has not necessarily failed, it is the expected outcome of operating within the limits of a data driven model.
It is the collaborative effort between machine and human and the mutual discovery of effective causal models which will yield the greatest successes in the future application of machine intelligence.
END
*There is a vast wealth of neuroscientific research into this field to which this simplified description does not do justice.
**Friston. Learning and inference in the brain. Neural Networks 16 (2003) 1325–1352 “Although supervised learning schemes have an established utility in helping understand some aspects of functional architectures in the brain they are not candidates for models of representational learning. This is because their supervised aspect means the generative model is already known.”
***Goodman, Ullman, Tenenbaum. Learning a Theory of Causality. Psychological Review (2010)
