Microsoft has unveiled Project Adam, its deep learning neural network which is extremely good at image classification. Although you may think image classification has been done to death, Project Adam is a game changer- it claims to be 50 times faster than current leaders in the field. The level of granularity on the system is also pretty impressive; Project Adam can tell the difference between a Pembroke Welsh Corgi and a Cardigan Welsh corgi.
They’ve released also released a video, covering the basics of neural networks and giving an overview of Adam’s current capabilities.
Back in 2012, Google and Stanford released an unsupervised neural network which used 16,000 computer processors to classify images; scanning 10,000 Youtube image thumbnails, the system developed a neuron with a visual understanding of what a cat (amongst other things) looked like. Project Adam claims to twice as accurate using 30 times less computing power and 50 times faster.
Complete our SAP x Data Natives CDO Club survey now, and help us to help you
The secret behind this success? HOGWILD! HOGWILD!, in spite of its ridiculous title, is an intriguing way of executing Stochastic Gradient Descent (SGD) algorithms, with increased parallelism. Typically, running SGD algorithms requires “performance-destroying memory locking and synchronization”, to avoid collisions (two programmes using the same block of RAM at once). HOGWILD! uses SGD without any locking, and “allows processors access to shared memory with the possibility of over-writing each other’s work”. No locking means “optimal rate of convergence”, and drastically improved throughput.
Project Adam is currently available as a phone app which allows you to identify objects in photos. As the video outlines, future uses the Project Adam team forecast include taking photos of food and being able to access their calorific content, and even getting prognoses on moles and rashes, simply by taking a photo on your smartphone.
But Peter Lee, head of Microsoft Research, says Project Adam won’t be stopping at image classification. He hopes to find applications in e-commerce, sentiment analysis and robotics- and ultimately, hopes Adam evolves into the “ultimate machine intelligence”, capable of handling a vast array of input sources.