The algorithms driving artificial intelligence systems, including those similar to ChatGPT, are facing significant challenges in adapting to new information. A new study highlights fundamental flaws in current model designs, prompting tech companies to invest heavily in training models from scratch.
Today’s AI frameworks, primarily based on neural networks that mimic brain functionality, go through a structured development process. Initially, AI systems are trained, refining their artificial neurons through algorithms tailored to specific datasets. Subsequently, they operate using this training to react to new inputs. However, once training concludes, the models cannot learn or update themselves with new information.
This limitation necessitates retraining large AI models whenever new datasets become accessible, a process that can be prohibitively costly, especially given the vast size of current data pools.
Recent research from the University of Alberta investigates whether prominent AI systems can be developed to learn continuously post-training. The findings reveal that many artificial neurons become unresponsive when exposed to new data, effectively impairing the AI’s learning capabilities.
According to the lead researcher, a significant portion of the artificial neurons cease to function effectively, akin to having most neurons in the human brain rendered inactive. The implications are profound for the development of smarter, more adaptable AI systems.
The researchers executed a series of experiments leveraging the ImageNet database, which includes 14 million labeled images. In contrast to conventional methodologies that involve single-cycle training, they retrained models after each image pair test. This approach revealed a quick decline in the networks’ learning abilities, resulting in many neurons becoming inactive.
Additionally, the study explored reinforcement learning techniques to enable an AI to adapt, akin to teaching an ant how to navigate. Despite attempts to enhance continuous learning, the results indicated a marked decline in the ability to learn as retraining progressed.
While this challenge appears intrinsic to existing learning frameworks, researchers have proposed a potential solution. By developing an algorithm that randomly activates certain neurons after each training cycle, they observed improvements in learning efficiency, effectively reviving inactive neurons and restoring their capacity to learn.
Though initial results are promising, further testing on larger systems is necessary to validate the efficacy of this innovative approach. Addressing the challenge of continuous learning in AI is a high-stakes endeavor, with substantial financial implications for the industry. A viable solution could dramatically decrease the costs associated with training advanced AI models.
Topics:
