Subsymbolic AI is a branch of artificial intelligence that focuses on non-symbolic methods of problem-solving. It emphasizes continuous, analogue, and dynamic computational processes, leveraging learning and adaptation, instead of relying on pre-defined rules and symbols.
Imagine you’re learning to ride a bike. No one gives you a manual with specific rules to follow, like “tilt 30 degrees to the left when turning left”. Instead, you keep trying, and over time you get a ‘feel’ for it. Your brain learns how to balance, steer and pedal mostly by itself. In a similar way, subsymbolic AI learns to solve problems not through explicit instructions, but by experiencing and adapting over time.
Subsymbolic AI, also known as distributed AI, promotes a view of intelligence as emerging from networks of simpler, interconnected processes rather than being purely rule-driven. It operates utilizing learning algorithms that allow the model to improve from data, experience, trial and error, and interaction with the environment. Subsymbolic AI, therefore, is closely related to machine learning.
One common type of subsymbolic AI is artificial neural networks (ANNs). Modeled loosely after the human brain, ANNs consist of interconnected nodes (neurons) in multiple layers. Each connection has an associated weight, which changes (is “learned”) over time based on input data and feedback through a process called ‘backpropagation’. These changes in weight allow the network to ’learn’ and ‘adapt’ to problems it’s solving.
Another notable aspect of subsymbolic AI is its lack of symbolic interpretation. In symbolic AI, data or knowledge is represented using symbols, logical rules, or a combination of both. In contrast, knowledge in subsymbolic AI is encoded as numerical weight values spread across many interconnected nodes. Because the ‘knowledge’ is distributed across the network and not symbolically represented, it is often difficult to interpret the inner workings of subsymbolic AI models.
Finally, it’s also important to note that while subsymbolic AI can overcome limitations of symbolic AI—such as its difficulty in handling fuzziness, uncertainty, complex real-world problems, and learning from data—it also has its limitations. For instance, its lack of interpretability and transparency, the quantity of data needed, and the computational resources required.