Artificial Intelligence (AI), a branch of computer science research, has been around for more than two decades. While it was initially referred to as machine learning within the confines of computer science, the domain has expanded to include many new technologies. Today, AI is essentially a combination of machine learning, computer vision, natural language processing, speech recognition and robotics.
AI is currently able to do things that we often take for granted. From autocorrect in search results to better night-time photos with computational photography to predicting where floods will occur, AI is embedded into a number of facets of our life. In other words, AI is no longer a promising idea but one that can solve persistent problems like inequality and illness to climate change.
In general terms, AI is basically computer models built by using data available at the time that is capable of making inference on new data. Depending on the quality of the data, an AI model can vary in terms of its impact. In order to be impactful, AI models have generally been built with a focus on doing only one thing. That is changing with Pathways, a new AI architecture built by Google that allows computer researchers to train a single model to do thousands of things. Here is everything you need to know.
What is Google Pathways?
Google describes Pathways as a new way of thinking about AI, addressing some of the weaknesses of the existing systems but at the same harnessing its strengths. An AI system today is often trained from scratch for each new problem where the mathematical model’s parameters are initiated literally with random numbers. In other words, imagine being asked to forget everything you know in order to learn a new skill.
This is a complex process where machine learning models are trained by never extending the existing system to learn new tasks. From nothing, each new model is trained to do one thing and only one thing and can be trained to do a specific task. This leads to companies like Google building thousands of models for thousands of individual tasks.
Jeff Dean, Google Senior Fellow and SVP, Google Research, says that this results in longer time for models to learn each new task and need for much more data. In a shift from existing methodology, Dean explains Pathways as a new AI architecture where Google is training one model to not only handle many separate tasks but “also draw upon and combine its existing skills to learn new tasks faster and more effectively.”
How will Pathways change contemporary AI systems?
Pathways is not a revolutionary idea but instead an evolutionary one that is developed based on years of building AI models. The idea here is to build a model that learns by training on one task and then learn another task without requiring to start from scratch. Google gives an example of a model learning how aerial images can predict the elevation of a landscape being trained to predict how flood water will flow through that terrain.
This will result in a future of AI models that can be called upon as needed. Dean also explains how Pathways differs from contemporary AI systems, which can take in text, or images or speech. But Pathways architecture will enable multimodal AI models that encompass vision, auditory and language understanding simultaneously. This results in a model that is more insightful and less prone to biases and mistakes.
Another problem with contemporary AI systems that Pathways aims to tackle is density and inefficiency. The AI models built today are dense, where the whole neural network activates to accomplish a task. With Pathways, a single model can be sparsely activated and allow for a small pathway through the network to get into action as required.
As mentioned earlier, the Pathways AI model is also capable of learning and will learn how to route tasks through the most relevant parts of the model. As a result, Pathways makes an AI model that has a larger capacity to learn a variety of tasks, is faster and also more energy efficient.
For years, tech companies have wanted to build systems matching human intelligence. Pathways does not reach there but it definitely thinks like humans by activating fewer routes and learning new tasks by embedding existing skills. With Pathways, Google is advancing computer science research from single-purpose models to general-purpose intelligent systems capable of deeper understanding of our world and ability to adapt to new needs.