Many of us have been following Google’s rapid advances in the field of artificial intelligence, especially in areas like voice search and speech recognition. But it’s Microsoft whose latest announcement in this area is currently catching everyone’s attention.
Project Adam is Microsoft’s latest and greatest endeavor in the “deep learning” movement. Deep learning is a branch of AI in which systems known as neural networks are trained to more closely emulate the thought processes of the human brain. It has been used with great success in speech recognition, as well as searching for exotic subatomic particles and interpretation and classification of images. Microsoft researchers impressed many last month with Skype Translate, which was shown to provide on the fly translations in voice conversations.
It’s in the arena of image classification that Microsoft’s researchers and engineers are unleashing Project Adam. Like other deep learning systems, it employs a large network of distributed computing. A similar system was built by Google researchers in 2012. They unleashed a neural network of 16,000 computers, using a billion connections, on the internet. And they pointed it in a direction not unlike that of many human web surfers: photos of cats. It was given 10 million randomly selected thumbnails from Youtube videos and a list of 20,000 items. After a few days of processing, it had “learned” to recognize pictures of cats by means of a deep learning algorithm.
But this week, Microsoft claims to have outdone the earlier Google experiment. Project Adam, they claim, is 50 times faster, and is also dramatically more efficient, using 30 times less machines to perform the work. This was demonstrated in a test using the huge image database, ImageNet, which contains 22,000 types of images. Previously, only a few artificial intelligence models, including Google Brain, the technology that fuels the speech recognition system in the Android OS, Google’s photosearch, and the aforementioned cat pic project could handle the processing of a dataset this large.
This time, it was dog pictures which were the subject of the experiment. And Project Adam performed brilliantly. As expected, it can recognize an image of a dog. And it can also recognize breeds of dogs. And it goes yet a step further: it not only can identify the breed, for example, a Corgi, but it can also tell the difference between a Pembroke Welsh corgi and a Cardigan Welsh corgi, a task that not many human could accomplish.
But Project Adam doesn’t stop there. The Microsoft team tells us that it will be able to take a photo of a food item, analyze it, and provide information on the nutritional content of the food. Another projected application will be using a photo of a skin lesion to make an early diagnosis.
How It Works
According to Trishul Chilimbi, a member of the Microsoft team behind Project Adam,
In particular, these methods seem to work extremely well in tasks involving vision. And it appears that as you add levels or layers to the deep neural network, accuracy increases, up to a point. The sweet spot seems to be about six layers, which interestingly, corresponds to the number of layers of neuron in the human brain’s visual cortex.
In the example given of the photo of the corgi, each layer of processing identifies a different quality of the picture, starting with the shape of the dog’s body. Subsequent layers could identify fur and texture, followed by body parts, and, ultimately, high level recognition concepts like facial features. Visual comprehension increases at each step.
Mystery And Wonder
The most fascinating aspect of Project Adam, like other deep learning projects, is that its actual workings remain shrouded in mystery, even to those developing it. Though humans provide the algorithms for machine learning, they don’t actually program or instruct the system on how to break down the images into features. Somehow, in a mechanism not fully understood, the system figures out how to do this.
Chilimbi likens this to the early days of quantum physics, in which the experimentalists and the practitioners couldn’t explain what they were observing, and were thus, ahead of the theoreticians. Similarly, current deep learning researchers see the power and capabilities of deep neural networks, but really don’t understand the fundamentals of how they work.
What are your thoughts on Project Adam and deep learning? Are we truly getting closer to seeing genuine artificial intelligence? Do you think machines will soon surpass the capabilities of the human brain, and make us obsolete? Sound off in the comments below.