Alphabet Inc’s Google announced the introduction of Pathways, a new AI solution that combines the abilities of multiple ML solutions and brings them on a single AI system.
According to Jeff Dean, SVP-Google Research and Health, Google Senior Fellow and also Google’s head of AI, ML models are overspecialized at individual tasks and rely on one form of input. To synthesize them to several levels Google has built Pathways. This solution will enable a single AI system to generalize across millions of tasks, to understand different types of data and with higher efficiency. He explains that the solution is, “advancing us from the era of single-purpose models that merely recognize patterns to one in which more general-purpose intelligent systems reflect a deeper understanding of our world and can adapt to new needs.”
Sign up for your weekly dose of what's up in emerging technology.
Dean claims that Pathways is a solution to the three limitations of today’s AI models.
– AI models are typically trained to do only one thing.
– AI models mostly focus on one sense.
– AI models are dense and inefficient.
Dean argues that today’s AI systems are trained from scratch for new problems and the mathematical model’s parameters are initiated with random numbers. Each new model trains from nothing to do only one thing only, rather than extending the existing learning, which makes the process a lot more time consuming. Their solution pathways allow training a single model to do multiple things. The model can have different capabilities and stitched together to perform new and complex tasks. This he claims is getting closer to the human brain.
The solution can enable multimodal models that encompass vision, auditory, and language understanding simultaneously. The announcement states that “Pathways could handle more abstract forms of data, helping find useful patterns that have eluded human scientists in complex systems such as climate dynamics.”
The next-generation AI includes a single model that is ‘sparsely’ activated. This means that small relevant pathways through the network can take the task rather than the whole system. Such an architecture with larger capacity and variety of tasks can be fast and much more energy efficient.
This is the second solution by Google AI to bring multiple solutions to work together. Earlier this week Google AI proposed a method called Task Affinity Groupings (TAG) to determine which tasks should be trained together in multi-task neural networks.