Artificial intelligence may be able to learn from minimal amounts of data as efficiently as newborn chicks.
Just after hatching, many birds learn to identify and follow the first moving object they encounter 鈥� a process called imprinting, which can offer protection in the wild as it helps them stay near a parent. It doesn鈥檛 take much visual information for a bird to learn to prefer one object and follow it. Researchers wanted to know whether AI models called transformers could do a similar task with limited inputs.
鈥淣o one has ever tested if transformers are actually more data hungry than brains,鈥� says at Indiana University Bloomington. Transformers are generic learning systems that can be trained to perform a wide variety of tasks, making them useful in both AI chatbots such as ChatGPT and in computer vision applications, such as image recognition and autonomous car navigation.
Advertisement
鈥淭o directly compare learning algorithms to brains, we need to train them on the same experiences,鈥� says Wood. She and her colleagues first raised chicks in a box where the only visual stimulation came from a rotating 3D object presented on a screen. After the first week, they ran each chick through hundreds of test trials that showed that same object on one screen 鈥� presented from both familiar and unfamiliar perspectives 鈥� and displayed a second unfamiliar object on another screen. The chicks spent more of their time near the first object, suggesting they had imprinted on it.
The researchers then created a virtual simulation of the set-up and used a virtual agent to move through it while looking around and recording a first-person view. That provided tens of thousands of simulated images for training and evaluating four transformer models.
The AI models had just 300 milliseconds to learn from each simulated image 鈥� approximating how long biological neurons fire after being presented with an image. The researchers found that the AIs could learn to recognise a 3D object as quickly and accurately as the chicks 鈥� with about 70 percent accuracy 鈥� 鈥渆ven when learning from a sparse environment鈥�, says Wood. Her team presented this work at the Conference on Neural Information Processing Systems on 12 December in New Orleans, Louisiana.
The study is 鈥渁 great piece of work鈥� in comparing machine performance with biological brains, says at the University of Sydney in Australia. But he also noted the limitations of measuring animal attention and experiences, such as a chick鈥檚 emotional experience of becoming imprinted on an object.
鈥淲e might be able to say that the chick 鈥榮aw鈥� its imprinting object, but that will have a component of experience to it,鈥漢e says. 鈥淧articularly as imprinting is to do with identifying its mother, it would be unsurprising if that visual experience were combined with a suite of other components of experience: fear yielding to comfort, for example, as the chick comes to regard the object as its imprinted 鈥榤other鈥�.鈥�
arXiv
