Researchers from New York University (NYU) utilized children’s learning processes to train artificial intelligence (AI). Published in the Science Journal, the method allows the AI system to learn from its surrounding environment instead of relying heavily on labeled data. The study was modeled after a child’s learning process.
To achieve this, researchers gathered a dataset through 60-hour first-person video recordings from head-mounted cameras worn by children between six months and two years old. This data was used to replicate the perspective of children in the AI model.
Next, a self-supervised learning (SSL) AI model was trained using the video dataset. The goal was to determine if the AI could learn about actions and changes by understanding temporal information from the videos, akin to the child’s learning process. SSL methods facilitate AI learning of patterns and structures in data without explicit labels.
The study also attempted to discern if AI required inherent biases or ‘shortcuts’ for effective learning or if it could comprehend the world via general learning algorithms—similar to how children learn. Even though the video covered barely 1% of the child’s awake time, the AI system picked up many words and concepts, demonstrating its efficiency at learning from limited but focused data.
The AI model demonstrated a high degree of competence in recognizing actions from videos and predicting missing segments within a video sequence – a process called video interpolation. This model was also more adept at recognizing objects under various conditions compared to those trained on static images, stressing the value of temporal data for building resilient and versatile models.
Wai Keen Vong, a research scientist at NYU’s Center for Data Science, emphasized the uniqueness of this approach. He pointed out that this study is the first to show a neural network trained on developmentally appropriate data from a single child, learning to associate words with their visual counterparts.
Study author Emri Orhan believes that self-supervised learning is critical for understanding complex learning processes and recommended its increased emphasis in AI Research. He also noted the growing interest in developing AI systems inspired by nature to create models or robots that think and behave like humans.
Despite the study’s promising results, Vong acknowledged some limitations, such as the language input to the model was text, not actual speech signals that children receive.
The research challenges traditional AI training models and promotes the dialogue on the most effective ways to mimic biological learning. It is an area of growing interest, particularly as colossal AI models like GPT-3 and GPT-4 start to reveal their limitations.