...

When he’s not working on one of his several projects, machine learning developer Jorge Muñoz spends much of his time trying to keep up to date on the cutting edge of artificial intelligence research.

“Every month there is something new; every month there are people doing something interesting,” he said. “The field is getting really huge, so it’s getting really difficult to stay informed on what everyone is doing.”

The AI explosion of the last few years “as it’s actually started working” has created a huge demand for resources to learn about AI, Muñoz said. Edge cases notwithstanding — Muñoz mentioned a 17-year-old who coauthored a paper published by OpenAI — he thinks years of experience are still required to successful develop machine learning models.

“I don’t think [AI] is something like programming where you can take a crash course in 3–4 months and become productive — you need to understand how things work internally or otherwise they’re not going to work,” he said. “You won’t be able to create a new deep learning model but to copy what others have done. You see this all the time: people who have 99.9% accuracy during training but when they put their model in production it doesn’t work at all.”

Like others that Clusterone has spoken with, Muñoz thinks we’re still a while away from developing artificial general intelligence.

“I think there’s a lot of propaganda surrounding the field — it’s very misunderstood, like we will have autonomous robots in our houses very soon,” he said, “It’s very challenging, there are a lot of problems we need to solve first.”

Muñoz said he’s happy working on a few of those problems in the meantime. Right now, his major project involves applying semantic hashing to image analysis; given an image, the goal is to search for images with similar subject matter. Semantic hashing consists of using a neural network to assign items to memory ‘addresses’ such that nearby addresses contain semantically similar items. Muñoz’s client hopes to use his work glean information from images on social networks like Instagram for marketing purposes.

Semantic hashing with images is not an easy problem, Muñoz said — he’s been working on the project for half a year. He has a terabyte of labeled images, an important asset given the fact that dataset availability is often one of the major problems facing machine learning projects.

“Supervised learning works very well but you need to label the data,” Muñoz said. “That’s where you spend all the time.”

Muñoz started the project with Caffe, then switched to TensorFlow for its ease of moving from experimentation to production.

“As soon as Google released TensorFlow it was clear to me that TensorFlow was the way to go, because its production is more flexible than Caffe,” he said. “You can use it for production — you can deploy your models on the web or on mobile.”

Semantic hashing has largely been used with text data, but Muñoz hopes to extend the technique to his work with images. Muñoz said he thinks a productive line of research in machine learning involves taking models that have been successful in a particular application and trying to adapt them for a different application: can we apply a translation model that has some knowledge of word semantics in conjunction with image recognition to describe images?

Looking ahead, Muñoz said fundamental changes to machine learning techniques are needed to move towards more successful models, and ultimately general intelligence.

Right now, to apply deep learning to problems “we have to design [model] architecture, but the brain doesn’t work this way,” Muñoz said. “So we need to find a better learning algorithm to address this.”

Given that many of the key ideas and techniques in machine learning originally stemmed from neuroscience, Muñoz believes further replication of brain function is the key to achieving better AI.

“Deep learning with backpropagation is very efficient and very good, but I think if we want to move on to next steps we need to change that learning algorithm,” Muñoz said. “Try to get it closer to how the brain works.”

“We need more neuroscientists coming into the field to help us,” he said.

Muñoz’s interest in the brain relates to what he says is another eventual outcome of the AI field that fascinates him: machine consciousness. He was originally drawn into the topic of artificial consciousness while studying machine learning at university in Spain.

Machine consciousness is coming, Muñoz believes, though he said there’s a long way to go. One problem, he said, is that we don’t have a definitive test for consciousness at this point. We know some of the features that need to be present in a conscious agent — awareness of surroundings and body, awareness of own mind — but don’t have good metrics for measuring these features.

There are unresolved philosophical debates regarding artificial consciousness as well, ranging from in what sense it is even possible to what legal and moral rights a machine consciousness might have, Muñoz said.

“Once we do create artificial consciousness, there’s going to be a big debate: should we treat it like a human?” Muñoz said. “People are going to be scared, but we have to deal with it.”

To Muñoz, ongoing research in artificial consciousness is one of the aspects of the AI field that makes him excited to work in it each day.

“Personally, it’s very inspiring,” Muñoz said. “To move forward we need to do something different.”