What is it about babies that allows them to effortlessly distinguish between strawberries and rotten eggs, whereas computers struggle to do so?
The cutting-edge technique for teaching machines to identify patterns was brought to prominence by Geoffrey Hinton, David Rumelhart, and Ronald Williams in the 1980s. Their approach is straightforward: if you were training a neural network to differentiate between the scents of eggs and flowers, you would present it with various smell patterns while simultaneously informing the network whether each pattern corresponds to an egg or a flower, based on the activation of a specific neuron at the network’s output. In essence, you provide the network with the correct answers. Afterwards, you compare the actual output to the desired output and adjust the weights throughout the network in a manner that brings the actual output closer to the desired result. Repeating this process millions of times allows the network to learn to recognise patterns accurately, enabling it to distinguish between the smells of eggs and flowers.
Hinton and his team referred to this learning process as back-propagation, a technique that remains in use today. Essentially, AI developers send the error calculated at the end back through the entire network, determining the precise error contribution of each synapse and adjusting that synapse accordingly.
This method of learning, where a network is trained by presenting it with examples along with their correct answers, is known as supervised learning. Humans oversee this learning process by supplying the network with the correct answers. While many supervised learning techniques are more intricate, the underlying principle remains the same: correct answers are provided, and the networks are adjusted through backpropagation to update weights until the classification of input patterns reaches an adequate level of accuracy.
When we use mathematical and computer science terminology to explain backpropagation, we can say that it calculates the gradient of a loss function in relation to each weight in the network by employing the chain rule from calculus. Throughout the training phase, the system is given labelled examples (meaning we specify what each item is and which features correspond to which). Discrepancies between the predicted outputs and the actual results are sent backwards, enabling adjustments to the weights so that the network enhances its predictions incrementally. The network needs explicit feedback (target labels) and a substantial number of examples, refining its ability to identify not only patterns but also distinct categories.
This technique is utilised in areas such as image recognition, speech recognition, and autonomous vehicles. However, it does not accurately reflect the workings of the brain. The brain does not rely on supervised learning. For instance, a baby’s brain doesn’t require explicit labels to differentiate between berries and rotten eggs. Infants are born with a well-developed sense of smell, allowing them to identify familiar scents (like their mother’s milk) from unfamiliar ones based on direct experience and inherent biological wiring. For babies, positive feelings, comfort, and successful feeding act as reinforcement mechanisms rather than labelled feedback and error correction. Their ability to recognise patterns is closely tied to their innate models of the world, which are evolved templates that inform their initial perceptions and behaviours.