Although extending their theory in these new directions will take some work, Serre and Poggio’s model has already begun to spread through both the AI and neuroscience communities at MIT. Electrical-engineering graduate student Stan Bileschi recently finished a doctorate that applied the model to scene recognition, which is the derivation of higher-order judgments – “it’s a farm!” – from the recognition of separate objects – a barn, a cow, a split-rail fence. Bileschi believes that general scene analysis will be critical to many real-world machine vision applications – surveillance, for instance.
Immediate recognition is the foundation of overall visual recognition, says Poggio, but it’s not all there is to it. There are many levels of recognition, and immediate recognition is one of the simplest. Depending on the context, an object might be identified as a toy, a doll, a Barbie, a reflection of American culture, a female, a representation of a girl with a weird growth disorder, and so on, down a long list. Similarly, in chess problems, recognizing the right move can take seconds or minutes or hours, depending on the configuration of the pieces. Presumably, as problems get harder, solving them requires recruiting higher levels of brain function – and that takes time.
An immediate-recognition model might solve the vision problems that have impeded the development of useful maintenance and construction robots. Or we might find that to be really useful, such robots need to be able to recognize both anomalies in the landscape and their causes. That type of recognition is clearly of a higher order.
The next step is to build recognition models that recruit more and more resources, and thus require more processing time. “We know how the model could be changed to include time,” says Serre. “This might bring us closer to thinking – just maybe.”