Deep Learning for Social Media Data Mining - Q&A

Q. What is Deep Learning?
Deep learning is a branch of computer vision where we teach a computer how to understand what’s inside an image by feeding the computer examples. At Ditto labs, we use this technology to empower marketers to sift through terabytes of photos to find the pixels that matter. Think about this – every day, billions of photos are shared on social networks, and we process them in real-time! The only way to deal with such massive amounts of data is with distributed computing solutions in the cloud building and running complex machine learning models. Luckily Ditto handles all the difficult work of perform image-recognition at scale and synthesizes valuable marketing insights for you automatically.

Q: What do you do?
At Ditto Labs we are on a quest to automatically see and understand as many things as possible in social media photos. Our machine learning software platform and infrastructure involve hundreds of servers running continuously in the cloud.

Q: What is the team at Ditto like?

Our team at Ditto Labs is a fantastic blend of engineering and research. We have people in Cambridge and Toronto and are growing quickly. Our research focuses on learning methods and structured prediction and their applications in image recognition specifically tailored to pictures of brands in social media photos.

Q: Why do you do what you do?
Computer vision requires a unique combination of creative thinking, statistical modeling, and massive amounts of cloud computing. The insight and modeling is just mathematics and the implementation is developer chops (and sometimes sheer hacking). We transform abstract mathematical concepts into real pieces of efficiently working software.

Q: What can deep learning detect in images?
A convolutional net is a way of organizing the neurons in artificial neural network in such a way that it can be applied to images. A Network is a tool of machine learning – which is a set of techniques that allow machines to do tasks that, as yet, only humans can perform. Just like synapses in the brain, the way these artificial nets work is by adjusting the strength of the connections between neurons. Each neuron learns something small, so by using thousands of neurons with a series of complex connections, we can learn more insightful things about images. Today a conv-net running on a mobile device can recognize objects in real-time (~60 recognitions every second).

Q: Why the term “deep” learning?
It reflects something real: shallow learning systems have one or two layers of neurons, while deep learning systems typically have five to 20 layers. It is not the learning that is shallow or deep, but the architecture that is being trained and the number of neuron connections involved.

Q: How much of machine learning is undiscovered?
Lots. Current ML theory lacks fundamental insights of biological learning. The type of deep learning that we are doing most of the time is called “supervised” learning. It works like this: You give a correctly labeled example to the system, it adjusts its parameters, and you repeat with a few million examples, after days/weeks of compute time the network figures it out the pattern with a high degree of accuracy. Now our animal brains don’t learn this way. And indeed they are much better than state of the art “unsupervised” learning solutions.

Q: What is the future direction of deep learning?
Deep learning has been successful at vision and speech, i.e., average human strengths. In this sense, expert topics like medicine and health care seem to be an area where DL will shine next (and sadly IBM’s Watson is failing here). On a philosophical note, if you consider the most ambitious questions in science today, you’ll typically hear “How do complex living systems emerge?”, “What is the unified field theory of physics?”, and “How does thought arise in the brain?”. Machine Learning and artificial intelligence, the field we work in, is inextricably linked to this last question. The true physical and computational nature of intelligence is still a mystery, yet everyday strides are being made in our understanding of what intelligence, both human and artificial, is really about. It’s exhilarating to be at the forefront of this.

Q: Anything else?
If you aren’t tapping into what all of this means for brands and business request an api key from Ditto today.