AI

Why the future of deep learning depends on finding good data

Comment

Image Credits:

Ophir Tanz

Contributor

Ophir Tanz is the CEO of GumGum, an artificial intelligence company with particular expertise in computer vision. GumGum applies its capabilities to a variety of industries, from advertising to professional sports across the globe. Ophir holds a B.S. and a M.S. from Carnegie Mellon University and currently lives in Los Angeles.

More posts from Ophir Tanz

We’ve already taken a look at neural networks and deep learning techniques in a previous post, so now it’s time to address another major component of deep learning: data — meaning the images, videos, emails, driving patterns, phrases, objects and so on that are used to train neural networks.

Surprisingly, despite our world being quite literally deluged by data — currently about 2.5 quintillion bytes a day, for those keeping tabs — a good chunk of it is not labeled or structured, meaning that for most current forms of supervised learning, it’s unusable. And deep learning in particular depends on a steady supply of the good, structured and labeled stuff.

In the second part of our “A Mathless Guide to Neural Networks,” we’ll take a look at why high-quality, labeled data is so important, where it comes from, how it’s used and what solutions our eager-to-learn machines can expect in the near-term future.

Supervised learning: I wanna hold your hand

In our post about neural networks, we explained how data is fed to machines through an elaborate sausage press that dissects, analyzes and even refines itself on the fly. This process is considered supervised learning in that the giant piles of data fed to the machines have been painstakingly labeled in advance. For example, to train a neural network to identify pictures of apples or oranges, it needs to be fed images that are labeled as such. The idea is that machines can be groomed to understand data by finding what all pictures labeled apple or orange, respectively, have in common, so they can eventually use those recognized patterns to more accurately predict what they are seeing in new images. The more labeled pictures they see, the bigger (and more diverse) the data set, the better they can refine the accuracy of their predictions; practice makes (almost) perfect.

This approach is useful in teaching machines about visual data, and how to identify anything from photographs and video to graphics and handwriting. The obvious upside is that it is now relatively commonplace for machines to be equal or even better than humans at say, image recognition for a number of applications. For instance, Facebook’s Deep Learning software is able to match two images of an unfamiliar person at the same level of accuracy as a human  (better than 97 percent of the time), and Google, earlier this year, unveiled a neural network that can spot cancerous tumors in medical images more accurately than pathologists.

Unsupervised learning: Go west, young man

The companion to supervised learning, as you might guess, is called unsupervised learning. The idea is that you loosen the leash on your machine and let it dive into the data to discover and experience it on its own, look for patterns and connections and come to conclusions, without requiring the guidance of a chaperone.

This technique had long been frowned upon by a certain segment of artificial intelligence scientists, but, in 2012, Google demonstrated a deep learning network that was able to decipher cats, faces and other objects from a giant pile of unlabeled images. This technique is impressive and produces some extremely interesting and useful results, but, so far, unsupervised learning doesn’t match the accuracy and effectiveness of supervised training for many purposes — more on that in a bit.

Data, data, everywhere

It is in the chasm between these two techniques that we run into the larger issues that are proving to be confounding. It’s useful to liken these machines to human babies. We know that by simply setting our baby loose, without guidance, it’ll learn, but not necessarily what we want it to learn, nor in any predictable way. But since we also teach our baby by instructing it, then we need to expose it to large numbers of objects and concepts in an essentially infinite number of topics.

We need to teach our baby about directions, animals and plants, gravity and other physical properties, reading and language, food types and the elements, you know — the very stuff of existence. All of this can more or less be explained over time with a mix of show-and-tell and answering the endless questions that any curious young human asks.

It’s a tremendous undertaking, but one that most parents, as well as other people around the average child, take on each and every day on the fly. A neural network has the same needs, but its focus is usually more narrow and we don’t really socialize with it, so the labels need to be much more precise.

Currently there are a number of ways that AI researchers and scientists can get access to data to train their machines. The first way is to go out there and amass a giant stockpile of labeled data on their own. This happens to be the case for companies like Google, Amazon, Baidu, Apple, Microsoft and Facebook, all of which have businesses that, funnily enough,  generate breathtaking amounts of data — much of it laboriously curated for free by customers.

It would be folly to try to list them all here, but think of the billions of labeled and tagged images uploaded to the cloud storage of all these companies’ databases. Then think about all the documents, the search queries — by voice, and text, and photos and optical character recognition— the location data and mapping, the ratings and likes and shares, the purchases, the delivery addresses, the phone numbers and contact info and address books and the social connections.

Legacy companies — and any company of huge scale — tend to have a unique advantage in machine learning in that they have copious amounts of specific types of data (which may or may not be valuable in the end, but often are).

Data the hard way

If you don’t happen to own a Fortune 100 company with collections of trillions of data points, then you’d better be good at sharing (or have deep pockets). Access to lots of extremely varied data is a key part of AI research. Fortunately, there already is a large number of free and publicly shared labeled data sets that cover a mind-boggling array of categories (this Wikipedia page hosts links to dozens and dozens).

Depending on your fancy, there are data sets showing everything from human facial expressions and sign language to the faces of public figures and skin pigmentation. You can find millions of images of crowds, forests and pets — all kinds of pets — or sift through boatloads of user and customer reviews. There also are data sets consisting of spam emails, tweets in multiple languages, blog posts and legal case reports.

New kinds of data are emerging from the myriad increasingly ubiquitous sensors in the world, such as medical sensors, motion sensors, smart device gyroscopes, heat sensors and more. And then there are all those pictures people take of their food, wine labels and ironic signage. In other words, there’s no shortage whatsoever of data in its purest form.

So what’s the problem?!?

Despite this apparent cornucopia of data, in practice, it turns out that many of these collections aren’t so broadly useful. Either they are too small of a collection, they are poorly or partially labeled or they just don’t meet your needs. For instance, if you’re hoping to teach a machine to recognize a Starbucks logo in images, you may only be able to find a training database of images that have been variously labeled “beverages” or “drinks” or “coffee” or “container” or “Joe.” Without the right labels, they just aren’t useful. And the average law firm or established corporation may have millions of millions of contracts or other paperwork in its databases, but that data isn’t usable as it’s likely in a simple unlabeled PDF format.

Another challenge in terms of optimal data is making sure that the training sets used are both large and diverse. Why? Let’s explore the idea of training data with a simple thought experiment. Imagine we give a little kid, we’ll call him Ned, the task of recognizing Spanish words on flashcards. When shown a flashcard, all Ned needs to do is say “Yes, this is Spanish” or “No, this is not Spanish.”

Having never seen nor spoken Spanish before, this kid Ned is given 10 random flashcards in order to learn what Spanish words do and do not look like. Five of the cards have the Spanish words: niño, rojo, comer, uno and enfermos, and the other five cards have words from other languages: cat, 猫, céu, yötaivas and नभ. Ned is told he can have a huge bowl of ice cream if he can pick out each of the Spanish words from a new set of flashcards. After an hour of studying, it’s time to test.

On the first test Ned is shown a Spanish word: azul. Because the character “a” only shows up in the non-Spanish pile, azul is not a Spanish word as far as Ned is concerned. The second card has the Portuguese word for mother: mãe. Ned immediately shouts, “Spanish!” Again, wrong answer, but his training cards include only one card with a tilde, and it happens to be in the Spanish pile. A third card has volcano on it. The boy notices that the word ends with an “o” and, remembering his training cards, he confidently says, “Spanish.” A fourth card showing “منزل” doesn’t look like anything from either pile and we can see tears building as the boy watches his ice cream melt. Is this a problem with his reasoning skills or his training data?

One issue: data set size. The boy has spent all his energy memorizing just 10 cards. In training a complex model, such as a deep neural network, the use of small data sets can lead to something called overfitting, which is a common pitfall in machine learning.

Essentially, overfitting is a consequence of having a large number of learnable parameters relative to training samples — parameters being those “neurons” that we were exhaustively adjusting via backpropagation in our previous article. The result can be a model that has memorized this training data as opposed to learning general concepts from the data.

Think of our apple-orange network. With a small amount of apple images as our training data and a large neural network, we risk causing the network to hone in on the specific details — the color red, brown stems a round shape — needed to accurately differentiate between just the training data. Those fine-grained details may do very well to describe the training apple pictures specifically, but prove to be inconsequential, or even incorrect, when trying to recognize new, unseen apples at test time.

Another issue, and an important principle, is data diversity. Ned would have been a lot better off if he had seen a non-Spanish word ending in “o” or a wider range of Spanish accent marks. Statistically speaking, the more unique data you accrue, the higher the probability that said data will span a more diverse range of features. In the case of the apple-orange network, we want it to generalize enough so that it recognizes all images of apples and oranges, regardless of whether they were present in the training set. Not all apples are red, after all, and if we train our network only on images of red apples (even if we have loads of them), we run the risk of the network not recognizing green apples at test time. Thus, if the types of data used during training are biased and not representative of data we expect at test time, expect trouble.

The issue of bias is beginning to crop up in a lot of AI. Neural networks and the data sets used to train them reflect any biases of the people or groups of people who put them together. Again, by only training our apple-orange network with images of red apples, we risk the network learning the bias that apples can only be red. What about green apples, yellow apples and candy apples? If you extrapolate to other applications, such as facial recognition, the impact that data bias can have becomes glaringly obvious. As the old saying goes: garbage in, garbage out.

Building a mousetrap that thinks for itself

Short of hiring people to label data — which is a thing, by the way, and it’s pricey — or all of the companies of the world suddenly agreeing to open up all their proprietary data and distribute it happily to scientists across the globe (we’d advise against holding your breath), then the answer to the shortage of good training data is not having to rely on it at all. That’s right, rather than working toward the goal of getting as much training data as possible, the future of deep learning may be to work toward unsupervised learning techniques. If we think about teaching babies and infants about the world, this makes sense; after all, while we do teach our children plenty, much of the most important learning we do as humans is experiential, ad hoc — unsupervised.

More TechCrunch

After Apple loosened its App Store guidelines to permit game emulators, the retro game emulator Delta — an app 10 years in the making — hit the top of the…

Adobe comes after indie game emulator Delta for copying its logo

Meta is once again taking on its competitors by developing a feature that borrows concepts from others — in this case, BeReal and Snapchat. The company is developing a feature…

Meta’s latest experiment borrows from BeReal’s and Snapchat’s core ideas

Welcome to Startups Weekly! We’ve been drowning in AI news this week, with Google’s I/O setting the pace. And Elon Musk rages against the machine.

Startups Weekly: It’s the dawning of the age of AI — plus,  Musk is raging against the machine

IndieBio’s Bay Area incubator is about to debut its 15th cohort of biotech startups. We took special note of a few, which were making some major, bordering on ludicrous, claims…

IndieBio’s SF incubator lineup is making some wild biotech promises

YouTube TV has announced that its multiview feature for watching four streams at once is now available on Android phones and tablets. The Android launch comes two months after YouTube…

YouTube TV’s ‘multiview’ feature is now available on Android phones and tablets

Featured Article

Two Santa Cruz students uncover security bug that could let millions do their laundry for free

CSC ServiceWorks provides laundry machines to thousands of residential homes and universities, but the company ignored requests to fix a security bug.

18 hours ago
Two Santa Cruz students uncover security bug that could let millions do their laundry for free

OpenAI’s Superalignment team, responsible for developing ways to govern and steer “superintelligent” AI systems, was promised 20% of the company’s compute resources, according to a person from that team. But…

OpenAI created a team to control ‘superintelligent’ AI — then let it wither, source says

TechCrunch Disrupt 2024 is just around the corner, and the buzz is palpable. But what if we told you there’s a chance for you to not just attend, but also…

Harness the TechCrunch Effect: Host a Side Event at Disrupt 2024

Decks are all about telling a compelling story and Goodcarbon does a good job on that front. But there’s important information missing too.

Pitch Deck Teardown: Goodcarbon’s $5.5M seed deck

Slack is making it difficult for its customers if they want the company to stop using its data for model training.

Slack under attack over sneaky AI training policy

A Texas-based company that provides health insurance and benefit plans disclosed a data breach affecting almost 2.5 million people, some of whom had their Social Security number stolen. WebTPA said…

Healthcare company WebTPA discloses breach affecting 2.5 million people

Featured Article

Microsoft dodges UK antitrust scrutiny over its Mistral AI stake

Microsoft won’t be facing antitrust scrutiny in the U.K. over its recent investment into French AI startup Mistral AI.

20 hours ago
Microsoft dodges UK antitrust scrutiny over its Mistral AI stake

Ember has partnered with HSBC in the U.K. so that the bank’s business customers can access Ember’s services from their online accounts.

Embedded finance is still trendy as accounting automation startup Ember partners with HSBC UK

Kudos uses AI to figure out consumer spending habits so it can then provide more personalized financial advice, like maximizing rewards and utilizing credit effectively.

Kudos lands $10M for an AI smart wallet that picks the best credit card for purchases

The EU’s warning comes after Microsoft failed to respond to a legally binding request for information that focused on its generative AI tools.

EU warns Microsoft it could be fined billions over missing GenAI risk info

The prospects for troubled banking-as-a-service startup Synapse have gone from bad to worse this week after a United States Trustee filed an emergency motion on Wednesday.  The trustee is asking…

A US Trustee wants troubled fintech Synapse to be liquidated via Chapter 7 bankruptcy, cites ‘gross mismanagement’

U.K.-based Seraphim Space is spinning up its 13th accelerator program, with nine participating companies working on a range of tech from propulsion to in-space manufacturing and space situational awareness. The…

Seraphim’s latest space accelerator welcomes nine companies

OpenAI has reached a deal with Reddit to use the social news site’s data for training AI models. In a blog post on OpenAI’s press relations site, the company said…

OpenAI inks deal to train AI on Reddit data

X users will now be able to discover posts from new Communities that are trending directly from an Explore tab within the section.

X pushes more users to Communities

For Mark Zuckerberg’s 40th birthday, his wife got him a photoshoot. Zuckerberg gives the camera a sly smile as he sits amid a carefully crafted re-creation of his childhood bedroom.…

Mark Zuckerberg’s makeover: Midlife crisis or carefully crafted rebrand?

Strava announced a slew of features, including AI to weed out leaderboard cheats, a new ‘family’ subscription plan, dark mode and more.

Strava taps AI to weed out leaderboard cheats, unveils ‘family’ plan, dark mode and more

We all fall down sometimes. Astronauts are no exception. You need to be in peak physical condition for space travel, but bulky space suits and lower gravity levels can be…

Astronauts fall over. Robotic limbs can help them back up.

Microsoft will launch its custom Cobalt 100 chips to customers as a public preview at its Build conference next week, TechCrunch has learned. In an analyst briefing ahead of Build,…

Microsoft’s custom Cobalt chips will come to Azure next week

What a wild week for transportation news! It was a smorgasbord of news that seemed to touch every sector and theme in transportation.

Tesla keeps cutting jobs and the feds probe Waymo

Sony Music Group has sent letters to more than 700 tech companies and music streaming services to warn them not to use its music to train AI without explicit permission.…

Sony Music warns tech companies over ‘unauthorized’ use of its content to train AI

Winston Chi, Butter’s founder and CEO, told TechCrunch that “most parties, including our investors and us, are making money” from the exit.

GrubMarket buys Butter to give its food distribution tech an AI boost

The investor lawsuit is related to Bolt securing a $30 million personal loan to Ryan Breslow, which was later defaulted on.

Bolt founder Ryan Breslow wants to settle an investor lawsuit by returning $37 million worth of shares

Meta, the parent company of Facebook, launched an enterprise version of the prominent social network in 2015. It always seemed like a stretch for a company built on a consumer…

With the end of Workplace, it’s fair to wonder if Meta was ever serious about the enterprise

X, formerly Twitter, turned TweetDeck into X Pro and pushed it behind a paywall. But there is a new column-based social media tool in town, and it’s from Instagram Threads.…

Meta Threads is testing pinned columns on the web, similar to the old TweetDeck

As part of 2024’s Accessibility Awareness Day, Google is showing off some updates to Android that should be useful to folks with mobility or vision impairments. Project Gameface allows gamers…

Google expands hands-free and eyes-free interfaces on Android