“Learning” is another one of those abstract concepts which reveals significant complexity upon further examination. In the context of people, learning represents our ability to incorporate experience in a beneficial way; we can learn facts, skills, or social norms (among countless other things) through repeated (or one-time) exposure. The exact mechanics underlying the learning process are not yet fully clear, but we do know that our learning takes place in our brains, with neurons strengthening or weakening connections to other neurons based on firing patterns. However, the concept of learning extends beyond people – we now have algorithms which can learn, and we also see a process analogous to learning occurring in nature with evolution. Similar to human learning, both machine learning and evolution also involve leveraging experience to drive improvements in the system.
Taking learning as a broad category, one segmentation which stands out is “statistical” learning vs. “direct” learning. Statistical learning involves generating many possible solutions and filtering them out (together with some degree of memory or heredity), whereas direct learning starts with a single solution and then iterates in an “intelligent” manner toward better solutions.
Evolution is the prototypical example of statistical learning, with many varied instances of life competing for resources and the best arrangements propagating themselves (or really, their genes) more frequently. Computers can also leverage statistical learning, with systems like Alpha Zero employing essentially random simulation to identify the best chess or Go moves (with the moves that are part of a win getting labeled as better). One way to visualize statistical learning is as directed randomness – the properties of the system are set up so that progress will occur, on average.
Direct learning, on the other hand, is exemplified by neural network algorithms. These systems are randomly initialized, and are then updated “intelligently” based on how they respond to specific examples. For example, a neural network can be trained to recognize cats by repeatedly testing it on pictures of both cats and non-cats and updating the weights in the direction of higher accuracy (on the set which it was just tested on). Key to direct learning is this ability to establish a gradient; the gradient allows the updates to be “intelligent”.
The major difference between statistical and direct learning is in where they leverage large numbers; statistical learning is reliant on a large number of solutions, whereas direct learning requires a large number of examples. These dynamics mean that statistical learning is more exploratory in nature (you don’t need to know exactly what a good solution looks like at the start), while direct learning is better geared toward synthesis / generalization (as it requires many “perfect” examples from the start). Looked at differently, direct learning can be seen as dragging the system in a particular direction, while statistical learning is more passive and involves seeing where the system lands.
Interestingly, our brains appear to make use of both statistical and direct learning. The most compelling example of statistical learning occurs in young children when a pruning process takes place that cuts down total synapse count by 50-90%. In this process, each synapse can be viewed as a “solution”, with only those more effective solutions remaining in place.
Direct learning, on the other hand, takes place constantly, with the world itself (or really, our senses) providing continual “perfect” examples, which our brain then works to synthesize / generalize (resulting in our concepts).
On top of the learning taking place in individual brains, the structure of the brain itself has been generated via millions of years of statistical learning (by evolution), which has been enabled by the direct learning occurring in the brain of each organism involved. From a biological perspective, statistical and direct learning are clearly deeply intertwined. It will be interesting to see whether human-level artificial intelligence systems will similarly require both types, or if it’s possible to achieve with just one (particularly as direct learning techniques have been the main focus of late).