Short & Sweet AI

How do you train neural networks to understand and simulate human emotions?

From Short and Sweet AI, I’m Dr. Peper and today I’m discussing how to train your AI.

We use 10,000 possible combinations of muscle movements in the face to create one facial expression. Add to this more than 400 possible voice inflections, along with thousands of hand and body gestures. All these combinations change continuously throughout a human conversation. Our brains process these complex, sometimes intense emotions, subconsciously, in microseconds, over and over again throughout the day.

Emotion and Datasets

The way AI can help us is to have machines that can effectively communicate with us and understand what we want. They need to recognize our emotional state, how we’re feeling, through our voice, facial expressions and nonverbal cues.

In order to teach computers how to understand emotions, AI researchers use machine learning and neural networks. Machines are very good at analyzing large amounts of data. We’re talking a dataset that has almost 8 million facial expressions. When a machine trains on that many variations, it learns to detect patterns in facial movements and even the nuances between a smirk and a smile. The machines can listen to voice tone and recognize sounds that indicate stress or anger. How does it do this?

Emotion Metrics

Using computer vision, the algorithms identify key landmarks on the face such as the tip of the nose, the corners of the mouth or the corners of the eyebrows. Deep learning algorithms then analyze the pixels of the images to classify the expressions. Combinations of these facial expressions are then mapped to emotions. Another program for analyzing speech evaluates not what is said, but how it is said, calculating changes in tone, loudness, tempo and voice quality to understand what’s happening and the emotion and gender of the speaker. These are called emotion metrics. And when tested against human emotions, the key emotion metrics have accuracies above 90%.

Many companies are working on emotion AI. Amazon has a network for speech based emotion detection. Another company, Affectiva, has a neural network called SoundNet, that can classify anger from audio data in 1.2 seconds, regardless of the speaker’s language. That’s as fast as a human can detect anger from a voice. Another company, Cogito, has a system which analyzes voices, of military veterans with PTSD, to determine if they need help.

FATE Flaws

But there are worries about this technology. Many people in the field raise concerns that these types of systems have FATE flaws. FATE flaws in AI stand for fairness, accountability, transparency and ethical flaws. For example, a study with one facial recognition algorithm, showed faces of black people are rated as angrier, than faces of white people, even when the faces of black people were smiling.

Lisa Barret, a professor of psychology, spent 2 years along with 4 other scientists scrutinizing the evidence, for the accuracy of emotion AI. They concluded that companies using AI cannot reliably fingerprint, emotions through expressions. However, she does think in the future, emotions can be measured more accurately, when more sophisticated metrics are available.

As she explained: “it’s intuitive that emotions are very complex. Sometimes people cry in anger, sometimes they shout, some people laugh when angry and sometimes, they just sit silently and plan the demise of their enemy”.

From Short and Sweet AI, I’m Dr. Peper.

As always you can find further reading, videos and podcasts in the show notes.

Humans are incredibly skilled at identifying the emotions in a conversation. We can “hear” a smile. And we correctly identify emotions in a voice even when we don’t speak the language. In fact more than 50 categories exist within the human emotions of surprise, joy, anger, sadness and fear. And each is conveyed through body language, words or tone. When you recognize these signals and respond appropriately, you have high emotional intelligence or high EQ.

AI has high IQ but low EQ

We know emotional intelligence and social skills correlate with a person’s potential for success in life. On the other hand, we live in a high IQ world surrounded with super advanced technology and AI systems developed to help us. But they have absolutely no EQ, no emotional intelligence. We need to build emotionally intelligent machines that truly understand human needs so we can have successful interactions with them.

Give machines emotions

The idea of making emotionally intelligent AI has been around for a long time. In 1997 an MIT Media lab professor, Rosalind Picard, published a book about computers and emotions entitled “Affective Computing”. Affect is a psychology term and refers to feeling, emotion, or mood.

Picard is credited with starting the field of computer science known as affective computing. It’s also called emotional artificial intelligence or emotion AI. Her book outlined how to give machines the skills of emotional intelligence so they can be genuinely intelligent and interact with us naturally. She believes computers should have the ability to recognize, understand, to even have and express, emotions. And by the way, this sounds very similar to what Ray Kurzweil has predicted in some of his https://www.abundance.video/videos/ray-kurzweil-peter-diamandis (conversations about the future).

The need for emotion datasets

In 2009 Picard and Rana el Kaliouby, a computer scientist from MIT, started an AI company called Affectiva based on emotion recognition technology. Subsequently, the company created a dataset of 7.9 million faces from 87 countries with recorded expressions for just about every human emotion. Above all, Picard and Kaliouby wanted to avoid biases in Affectiva’s algorithms. They therefore used a diversity of faces to pick up the differences in expressions from all ethnic groups, ages, genders and cultural backgrounds. Incidentally, I talked about the bias in large datasets in a previous flash talk on https://drpepermd.com/episode/imagenet/ (ImageNet).

Today Affectiva’s algorithms can detect human emotion from facial expressions and vocal cues. But even more, Kaliouby wants to train machines to recognize the subtle nuances in human emotions. Humans use a lot of nonverbal cues. Gestures, body language, voice tone all contribute to how emotions are communicated. For that reason researchers plan to develop emotion AI that is multimodal and can detect emotion the way humans do from multiple channels. Ultimately, Kaliouby wants to fuse digital technology with an ability to understand the humans using it.

The application of emotion AI

The power to detect human emotion has implications for every aspect of society. Emotion AI technology can detect mental and physical ailments based on how patients look or sound. In marketing it determines consumer’s reactions to commercials and TV shows. In the automotive world, emotion AI can identify distractions going on inside the car that could affect safety, such as arguments or a driver’s lack of focus. Finally, the biggest role so far has been in customer service. Call centers are already using emotion AI to identify the mood of customers on the phone.

DeepZen has released for purchase the first AI narrated audiobook.

From and Sweet AI, I’m Dr. Peper and today I’m talking about AI audiobooks.

In a previous flash talk, I discussed how we’re entering a https://drpepermd.com/episode/2-voice-first-computing/ (voice first )future. With smart assistants leading the way, we will request and consume information by speaking rather than type or read from a screen. We will type less on our laptops and smart phones and communicate more with voice. And as a result, people will consume more audiobooks.

Text to Speech

There are about one million books published each year in the US. Despite this only 40,000 books are recorded due to the costs. Audiobooks are time consuming and can cost up to $5000 per book to record. Not surprisingly then, companies have focused on perfecting AI to change text into speech through deep learning based systems. And there’s a whole history of machine learning breakthroughs over the last few years which has led to progressive improvement in the natural language processing algorithms. One of the biggest hurdles is AI generated voices sound flat and without emotion, in an almost comical way. Remember the Youtube https://www.youtube.com/watch?v=PTUY16CkS-k (Ben Bernanke video) of the financial crisis? Well, all that’s changed.

DeepZen

DeepZen, an London based AI company, released examples of it’s latest AI text to speech technology and they sound really good. The DeepZen team trained their algorithms on thousands of hours of narrator speech. As a result, the algorithm produces human sounding, highly emotive audio recordings using text from a book. Judge for yourself. Here’s a snippet of the audiobook, The Metamorphosis by Franz Kafka, generated by DeepZen’s text to speech technology.

Isn’t that fantastic? This is an audio recording generated by a machine from the text of a book. Because of this AI technology, it’ll be easy and cost effective to make an audio recording of any book out there. Eventually in all different languages.

Emotion AI

DeepZen, and other companies like it, are at work on translating human emotion through machine or deep learning for other things besides recording audiobooks. It’s the field of emotion AI which allows machines to determine a person’s mood by the sound of their voice. And will create more human like interactions between machines and man. We can talk about that in the next Short and Sweet AI. I’m Dr. Peper.

https://literallypublicrelations.wordpress.com/2020/03/02/the-future-of-audiobooks-is-ai/ (https://literallypublicrelations.wordpress.com/2020/03/02/the-future-of-audiobooks-is-ai/)

https://news.developer.nvidia.com/inception-spotlight-deepzen-uses-ai-to-generate-speech-for-audiobooks/ (https://news.developer.nvidia.com/inception-spotlight-deepzen-uses-ai-to-generate-speech-for-audiobooks/)

https://www.wired.com/story/opinion-conversational-ai-can-propel-social-stereotypes/ (https://www.wired.com/story/opinion-conversational-ai-can-propel-social-stereotypes/)

https://www.wired.com/story/google-assistant-can-now-translate-on-your-phone/ (https://www.wired.com/story/google-assistant-can-now-translate-on-your-phone/)

Is there an upside to the coronavirus? Nope. But the outbreak did show how AI can be used to predict and accurately track a pandemic.

From Short and Sweet AI, I’m Dr. Peper, and today I’m talking about what everyone is talking about, the coronavirus or COVID 19.

There are so many ways AI impacts the cornavirus outbreak. Chinese drones are disinfecting public areas and track people who don’t adhere to quarantine. Robots decontaminate hospital rooms. Self-driving cars in China deliver supplies to medical workers. Facial recognition cameras search for people not wearing their mandated face mask. Infrared temperature scanners detect fevers in large groups of people. Doctors use AI software to find evidence of coronavirus in lung scans from patients who are ill.

Think about what happened with the SARS virus in 2003 and compare that to the coronavirus today. What you realize is over the last two decades AI has really advanced how we respond to pandemics.

Early Prediction with AI

Early detection means earlier disease containment. And AI can do it faster. Right from the beginning an artificial intelligence company sounded the alarm on the coronavirus. A Canadian company BlueDot used AI powered algorithms to analyze information from many different sources. They were able to identify places were there were outbreaks of diseases and forecast how they spread. The company sent out warnings to it’s clients to avoid Wuhan on December 31, 2019. The World Health Organization sent out a public warning on January 9, 2020, not until 10 days later.

100,00 reports /day

BlueDot used natural language processing and machine learning to analyze large amounts of data. The company uses an automated infectious disease surveillance program. The algorithm sifts through foreign language news reports, animal and plant disease publications, new releases from government and public health departments and much, much more. The data is vast: 100,00 new reports in 65 languages a day.

BlueDot’s algorithm doesn’t use social media data because the company says it’s too messy. Finding signs of the virus in a vast soup of rumors, posts about ordinary cold and flu symptoms and lots of speculation, requires as yet unavailable training sets for the algorithms. But BlueDot does use some unexpected sources such as global airline ticketing data. Using this information, the Bluedot physician and programmers correctly predicted in the first few days the virus would jump from Wuhan to Seoul, Taipei, then Tokyo.

Humans Validate Conclusions

Bluedot highlights how the best use of AI is to augment human understanding. After the data sifting is finished, epidemiologists take over to make sure, from a scientific stand point, the conclusions from the data make sense. As blueDot’s founder Kamran Khan points out “What we have done is use natural language processing and machine learning to train this engine to recognize whether this is an outbreak of anthrax in Mongolia versus an reunion of the heavy metal band Anthrax”.

But final supervision requires human input to validate the AI’s findings. Information from AI algorithms need humans to put it in context to take the next step. The field of artificial intelligence needs people who can operate at the intersection of AI and biology. It’s not enough to be an AI engineer. What’s needed is someone who can understand biology well enough to apply what AI comes up with.

Augmented Intelligence

The use of AI in prediciting the coronavirus pandemic shows us what many experts in artificial intelligence alr...

Data and AI are moving out of the cloud onto the edge of the network.

From Short and Sweet AI, I’m Dr. Peper. And today I’m talking about edge AI.

The Cloud

First there were mainframe computers which stored lots of data. Then data was stored on hard drives on desktop computers. Then laptop computers revolutionized our experience and we could access the internet to get unlimited data that’s stored in the cloud. And as you know, the cloud is a term referring to the massive number of computer servers which collect and store data from the internet. These servers are located in data centers all over the world. Now with mobile devices, we have access to data anytime, anywhere via the cloud.

Cloud Problems

But connecting to the cloud comes with problems such as the lag time between the cloud and the smart device, the cost of storing data in the cloud, and more pressing issues of privacy. We realize smart assistants are sending snippets of our audio back to the cloud for training purposes. And networks using the cloud can be hacked and data stolen.

The Edge

Meanwhile the https://drpepermd.com/episode/what-is-iot-and-why-does-it-matter/ (Iot) is connecting millions of devices to the internet through very fast speeds via https://drpepermd.com/episode/5g-fifth-generation-wireless-what-is-it/ (5G technology). But in order for these devices to respond quickly and efficiently, data is now moving out of the cloud and onto the devices or the edge of the network. And the AI is moving there too. If the data can reside on the device and be processed by artificial intelligence there, it doesn’t need to be sent back to the cloud. Edge AI is computing that takes place on the device rather than in the cloud.

No Dad Jokes

As an example, think of a security camera. It collects 24 hours of video data to send to the cloud for processing. This is can be quite expensive. For 23 of those hours nothing has happened. But with edge computing, the smart security camera knows to send to the cloud the one hour of video where something did happen.

Another everyday example is our coffee makers. With edge AI these smart devices need to recognize only 200 words to make coffee. It doesn’t need a cloud worth of data and AI being sent back and forth for the coffee maker to recognize the command “brew three cups of coffee”.

As Clive Thompson in a https://www.wired.com/story/edge-ai-appliances-privacy-at-home/ (Wired magazine article) explained: ” I don’t need light switches to tell Dad jokes or acheive self-awareness. When it comes to gadgets that share my house, I’d prefer they be less smart”.

Putting It Together

So how will edge computing, 5G and IoT work together? 5G creates a new layer of edge computing. A new network of IoT connected devices all interacting with each other within milliseconds through the AI and data located on the devices. enabled by 5G’s faster speeds and response times, and mmWave technology. IoT, 5G and edge AI can create private local area networks called “fog” compared to the traditional networks in the cloud. These local networks are more reliable and secure because they process data on the devices. Less data sent to the cloud means less chance of networks getting hacked.

Things are getting edgier because edge AI solves problems of efficiency, cost, and privacy.

Further reading, videos, and podcasts are linked in the show notes. And if you like this episode, reviews are always appreciated.

A consumer study found 58% of Americans don’t understand what 5G is. From Short and Sweet AI, I’m Dr. Peper, with an explanation of 5G.

What Is 5G?

5G stands for fifth generation cellular wireless. And it’s important to know it’s not just another “G” of 4G technology. It won’t work the same as 4G. 2020 is the year super fast 5G is expectd to have a slow roll out. That means the first 5G networks will be faster than 4G. But 5G won’t achieve it’s fastest speeds until companies complete the infrastructure. To get these benefits, users will have to buy new phones and the wireless providers will need to install new networking equipment.

How Fast Is 5G?

How much faster is 5G? Downloading a typical movie to your phone would take 17 secoonds with 5G compared to six minutes for 4G. 5G wireless will have connection speeds 10 – 20 times faster than the speediest home internet service. And it will be 600 times faster than typical 4G speeds on your phone today. 5G cellular technology uses something called millimeter-wave networks or mmWave. With mmWave, data can be streamed to phones at extremely fast rates but only over short distances. So a huge number of access points or small cell sites will be needed to transfer the signals instead of a few huge cell towers.

But is it only speed that matters?

There’s another speed, called latency speed, that’s perhaps even more important. The time between you asking Siri a question, searching the web, and getting a response will be faster. This is because of the lag time or latency speed with 5G is faster due to newer networking technology and more reliable signals.

Autonomous Vehicles

But does all this really make 5G revolutionary and justify the hype. Quite honestly, yes. Because it’s what this technology can now accomplish that’s so exciting. And this comes back to Iot, https://drpepermd.com/episode/what-is-iot-and-why-does-it-matter/ (the internet of things) which I discussed last time. The shorter latencies of 5G allow things connected to the internet to communicate directly to each other. MmWave technology allows thousands of things to be directly connected together at once.

This technology will truly kickstart fully autonomous vehicles. Remember autonomous vehicles whichhttps://drpepermd.com/episode/14-self-driving-cars-are-we-there-yet/ ( I talked about in a previous episode) have no driver in the car unlike self driving vehicles with a steering wheel and a driver supervising. With 5G, cars are synchronized to traffic lights and each other. Future traffic has been described as being an elaborate street level ballet where cars flow like schools of fish in unison without colliding. There is even a new anticipated infrastructure called CV2X which stands for “cellular vehicle to everything”.

Virtual Reality

Because of 5G, IoT and something called edge computing, virtual reality becomes the new reality. With these 3 technologies, the perfect VR that exists today only in controlled scientific labs can now exist anywhere. These are perfect conditions for creating realistic representations of you with your specific tics and mannerisms. VR environments can be digitized and shared between users miles apart in real time. We will replace FaceTime with HologramTime. As we chat and interact with someone in their virtual world, they see us as holograms right next to them in their real world. We are one step closer to Ready Player One.

I’ve talked about IoT and 5G but the trifecta includes another technology called edge AI.

IoT, 5G and Edge Computing: these 3 different terms are everywhere and it’s time to talk about them. Today let’s discuss the Internet of things or IoT and see why it matters.

IoT

After all my research, the best definition I’ve found for the internet of things is by https://www.iotforall.com/what-is-iot-simple-explanation/ (Calum McClelland,) and, by the way, he has a lot of c’s and l’s in his name. Calum’s definiton of the internet of things is pretty simple. He says it means taking all the things in the world and connecting them to the internet. Voilà!

But what does that mean, all the things in the world? Well let’s look at the most popular by far IoT: smartphones. As Calum explains, with a smartphone you can listen to any song in the world. But not every song in the world is stored on your smartphone. In fact every song in the world is stored somewhere else. But your phone can ask for that song by being connected to the internet and then stream it to your device. Contrary to popular belief, your phone is not a super computer with unlimited storage. But it is connected to supercomputers and vast storage in the cloud.

Another example is your smart home assistant. When you ask it to play the daily news, it doesn’t have all the news programs stored in the device. But it’s connected to the cloud which does have that information. So it’s a thing connected to the internet or IoT.

Things Collect and Send

All the things in the world connected to the internet can be grouped into 3 categories. Things that collect and send information such as wearable health monitors. Things that receive intormation and put it into action such as your car keys. And things that can do both.

The things that collect and send information do so by sensors embedded in the smart device. Motion sensors tell us how many steps we’ve taken each day. Listening sensors can tell how many hours we’ve slept. Light sensors turn on when they sense the motion of someone in the room. But the important thing is this. The devices collecting and sending information via connection to the internet make our life easier.

Things Receive and Act

Things that receive and act on information, the second category of IoT, are smart devices such as thermostats which receive a command from us and then act on it and turn on the heat. Or alarm systems we can tell to unlock the house. Even refrigerators we can ask to show what food we have. Using these internet of things we can tell machines what to do even if we’re far away.

Helpful IoT Collects, Sends, Receives, Acts

And then things start to get awesome when IoT can do both, collect and send information and receive and act on information. An example would be a wearable alert system which use embedded sensors to detect when your body posture changes. The IoT device then sends that information to the cloud which analyzes your motion and determines you fell. The alert system then acts on that information to call 911. Ta-da!

So why is IoT grouped with 5G and edge computing. In my upcoming talks I’ll discuss how 5G will connect machines in diverse places such as factories, hospitals, schools and cities via IoT. And 5G will allow infrastructive to be retrofitted with artificial intelligence through edge computing as AI moves out of the cloud onto devices. Don’t worry! As always, I’ll make the explanations of 5G and edge AI short and sweet.

From Short and Sweet AI, I’m Dr. Peper.

We call them computers. They call them mentats. One is a machine, the other human. Both have superhuman intelligence.

From Short and Sweet AI, I’m Dr. Peper and today I’m talking about mentats.

Technology but No AI

In the science fiction novel Dune, the use of computers and artificial intelligence has been outlawed. The author, Frank Herbert was far from mainstream science fiction when he created a future universe conspicuously lacking in artificial intelligence and robots. Don’t get me wrong, there’s a lot of technology in the Dune universe: lasguns, atomics, intergalactic navigation powered through the use of prescience created by a substance called spice. But there are no droids, no self driving transporters, no computer vision, no robot storm troopers.

In the novel’s back story we learn that in the past men had used thinking machines to enslave humanity. This lead to several centuries of war known as the Butlerian Jihad. In the end humans prevailed and defeated the men with the machines. And thinking machiines were outlawed, use of computers punishable by death. Humans had to enhance their own natural intelligence with rigorous discipline and advanced training. They learned to cultivate superhuman abilities by following a secret training method. They learned to rapidly analyse and process large amounts of data in great detail. Just as an olympic athlete masters the physical demands of competition, these humans honed and expanded their mental abilities. They became living supercomputers known as Mentats.

Mentat Training

The Mentat training encompassed many elements and different levels of ability. Their skills included logic, inference and extrapolation, insight, future planning, and detailed understanding of events. At the highest level they were skilled in wisdom and diplomacy, negotiated delicate matters, and could judge matters of life and death. They could make decisions similar to machine learning which is based on data and probabilities but they were not able to make intuitive decisions. Indeed, this inability to make decisions in the absence of data made them ineffective as leaders.

Dystopian science fiction stories of machines with superintelligence rising up against the humans who created them have become ever more popular. Is it because we’re living during a fourth revolution created by rapidly advancing artificial intelligence and we’re fearful of it? Some think Frank Herbert decided to not have thinking machines or artificial intelligence in Dune for a reason. Because he wanted to warn against AI and the dangers of a society run by intelligent computers. Is his future where humans have superhuman intelligence like computers really possible?

Augmented AI

The answer is, perhaps, if you remember Ray Kurzweil’s prediction discussed in a previous episode about his book, https://drpepermd.com/episode/13-the-singularity-is-near/ (The Singularity is Near). He believes in the future there will be something called brain computer interfaces. Brain computer interfaces or BCIs will connect all the information and data from the cloud and download it directly to our brains. There’s a photo of what it might look like on my instagram today. Then we will not have to ban artificial intelligence. Nor rely on some secret training program to enhance our mental abilities to supercomputer levels as in the novel Dune.

Perhaps it will be a combination of humans augmented by artificial intelligence with brain computer interfaces. In my opinion, it is possible our future will be like Frank Herbert’s future in Dune, where we will all become Mentats or human supercomputers.

From Short and Sweet AI, I’m Dr. Peper and today I’m discussing artificial intelligence and the fantastic, futuristic novel, Dune.

Intelligent technology, robots, space creatures and other supernatural worlds all live in science fiction. Much of it is dystopian with thinking, conscious machines rising up against humans and taking over and controlling the human race. But Dune gives us another option.

Dune takes place 20,000 years in the future, on Arrakis, a planet that is entirely desert. And as the political, ecological and religious battles of the great houses of Duke Leto and Baron Harkkonen and the Imperium play out, artificial intelligence is conspicuously missing. What gives? How can the best selling science fiction novel of all time not include AI?

Butlerian Jihad

Within the first few pages of the novel, we learn that in the past men used machines to control and enslave the human race. This had lead to a great war, the Butlerian Jihad. The war lasted several centuries until men defeated other men and the machines they used. In the book thinking machines, basically computers and artificial intelligence, became outlawed. Anyone recreating them was sentenced to death. It was a universal commandment: “Thou shalt not make a machine in the likeness of a human mind.”

An often overlooked but crucial point is that the machines did not somehow enslave humans by themselves. Rather it was men controlling the machines who enslaved other men. As the book explains “Once men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them.”

Dune 2020

Frank Herbert who wrote Dune seems to be saying that if we let AI do our thinking, we can be controlled by the people who control the AI. As I discussed in a previous episode, we see this happening with the https://drpepermd.com/episode/why-is-bias-in-ai-important/ (bias) being programed into the computer algorithms. Algortihms that are used to make decisions which affect us on a daily basis. Herbert wrote Dune in 1965. This year there is mounting excitement and tremendous interest in the book. More than half a century later, it’s being made into a movie that has a fervent following. Dune fans and devotees are saying “Let us please get the Dune movie we all deserve.” Should that include a wish for thinking machines that can’t control us?

In my next episode I’ll discuss Dune’s alternative to artificial intelligence.

From Short and Sweet AI, I’m Dr. Peper

https://vocal.media/futurism/how-frank-herbert-s-dune-warned-of-the-rise-of-artificial-intelligence (https://vocal.media/futurism/how-frank-herbert-s-dune-warned-of-the-rise-of-artificial-intelligence)

https://steemit.com/philosophy/@zyx066/dune-and-the-thinking-machines (https://steemit.com/philosophy/@zyx066/dune-and-the-thinking-machines)

https://en.wikipedia.org/wiki/Dune_(novel) (https://en.wikipedia.org/wiki/Dune_(novel))

https://drpepermd.com/episode/why-is-bias-in-ai-important/ (Why is Bias in AI Important?)

Why is bias in artificial intelligence so important? Many people don’t realize that the algorithms used in AI today have a great impact on our daily life. These software programs decide whether we’re invited to a job interview, or eligible for a mortgage or undersurveillance by law enforcement. Organizations make these decisions with algorithms trained using datasets. If the datasets only reflect a few groups such as college educated people or people from certain socioeconomic backgrounds then the decisions will be biased.

bias in = bias out

The researchers who developed the datasets did not make the AI systems this way on purpose or out of malice. It was more unintentional and unconscious. People who create the algorithms have their own experiences and blindspots and the data reflects this. And you have more bias in the algorithms when the AI teammembers who program the computers are not a diverse group.

A quick example of bias in artificial intelligence is voice assistants like Alexa that’ve been trained on huge datasets of recorded speech from white, upper-middle-class, Americans. As a result the technology doesn’t understand commands from people with different accents and expressions.

ImageNet Roulette

In September 2019 a program called ImageNet Roulette caused a Twitter storm. People uploaded their selfies to the online program which used ImageNet to create labels. The labels attached to the selfies ranged from benign things like “face’ or “a person of no influence” to more troubling labels such as “first offender” and “rape suspect”. The project showed the dangers of feeding flawed data into an AI algorithm.

ImageNet is a 15 million image dataset that unlocked the potential of deep learning, a type of artificial intelligence used for everything from facial recognition to self-driving cars. This massive dataset is routinely used to train deep learning algorithms. But ImageNet Roulette’s creators wanted to crack ImageNet open and to show how biased the images are. And how the flawed dataset can lead to many flawed algorithms. As a result, a massive effort was launched to remove the most offensive labels and make the images more diverse.

AI Needs Diversity

Fei Fei Li, the computer vision expert who created ImageNet, has become a champion to make AI less biased and better for humanity . She left Google to lead Stanford’s new Institute for Human Centered AI. She’s testified before congressional hearings about the need to make changes to ensure there are diverse people engineering AI. And she’s founded AI4All, a summer program for girls in high school to develop more diversity in artificial intelligence.

Ten years after the launch of ImageNet, Li believes AI research needs to include people in neuroscience, psychology, philosophy and other disciplines to create AI with more human sensitivity. As she has said: “There is nothing artificial about AI. It’s inspired by people. It’s created by people and most importantly, it impacts people. It is a powerful tool we are only just beginning to understand, and that is a profound responsibility.”

As always, links to further reading, videos and podcasts are in the shownotes.

From Short and Sweet AI, I’m Dr. Peper

https://www.wired.com/story/ai-biased-how-scientists-trying-fix/ (https://www.wired.com/story/ai-biased-how-scientists-trying-fix/)

https://www.scmp.com/magazines/post-magazine/long-reads/article/2183463/bias-bias-out-stanford-scientist-out-make-ai-less