Photo by Carlos Muza on Unsplash
By Mehak Garg
AI has disrupted numerous industries such as communication, healthcare, and transportation by increasing access to information, creating surgical robots, and making self-driving cars. AI applications such as biometrics, deep learning platforms, and AI-optimized hardware are relatively new innovations that have the ability to make even bigger impacts in hundreds of industries. In order to ensure this impact remains positive, it is imperative that we consider the safety and ethics of AI before implementing it further in society.
In the past, we’ve seen how AI has been misused. For example, in mortgage lending, some AI systems discriminate against minority groups even though they were programmed with the intention of being fair. According to researchers at UC Berkeley, “Both online and face-to-face mortgage lenders charge higher interest rates to black and Latino borrowers, costing those homebuyers up to half a billion dollars more in interest every year.” In 2018, Amazon’s hiring system’s use of AI led to a bias where fewer women would be hired and have their resumes looked at. After the incident, they scrapped their system and rebuilt it to mitigate the issue. AI can even be discriminatory in our search bars. For example, women searching for a job on Google are less likely to be shown executive jobs or leadership roles compared to males.
AI ethics are a set of standards that can guide developers in using AI technologies to ensure that the final product is moral and ethical. Technology can be deemed moral if it treats everyone in society equally and doesn’t stereotype different groups based on their income, gender, or race. In order to successfully deploy a set of ethics to the field of AI, researchers must understand how these biases present themselves and the different ways AI technology can discriminate and be unsafe. There are three different levels of bias. The first level is a historical bias that already exists in the data set. The second level encompasses representation and measurement bias which are a result of how the algorithm is programmed. The third level includes evaluation and aggression biases which are a result of choices made when actually programming the algorithm.
In order to make AI safe to use and ethical, researchers have to optimize algorithms to limit the effect of these biases. Similar to how in real life there are multiple ways to solve a problem, there are multiple ways to program an AI. We can use different learning models for different types of problems. In fact, depending on the model, unsupervised learning might be more discriminatory than supervised learning. Supervised learning is when we train the machine as if it were in the presence of a teacher. Unsupervised learning is when we let the machine act on the data without any guidance.
To combat historical bias, researchers and others involved with the AI project should repeatedly check the data set to make sure it represents a diverse sample. Asking sample test questions while combing through the data set, such as which demographic gets the most loans or if females get more loans than men in a loan-centered dataset, can help you find potential areas of bias in the algorithm. As a final check, companies should monitor the real-world results of their algorithms like the demographics that are actually getting loans. Monitoring results regularly allows companies to act proactively and resolve these biases much more efficiently.
Fortunately, more and more companies have been prioritizing ethics and have outlined what an ethical algorithm functions like. Different organizations, private companies, and researchers have established five goals for every AI system. According to Anna Jobin from the Health Ethics and Policy Lab, an AI system is deemed as ethical or safe if it employs transparency, justice and fairness, non-maleficence, responsibility, and privacy. AI can make immeasurable impacts in fields spanning food production to the defense industry. Similar to how employing ethiccs in Amazon’s hiring system helped reduce biases, coupling AI’s potential with ethical guidelines will result in magnified impacts that can further benefit society.
Image Credit: Nikkei Asian Review
By Tia Jain
With about 2.5 million confirmed cases and nearly 170,000 deaths world-wide, the COVID-19 pandemic has caused a global catastrophe, as healthcare facilities find themselves struggling to treat overwhelming numbers of patients with insufficient supplies. Nobody would have thought that this year, words like "social distancing" and "quarantine" would become commonplace. In the midst of this global pandemic, many scientists are wondering if and how AI can be applied to help solve COVID-19. However, since AI is still a budding field, much of its potential still remains unexplored, particularly its application to healthcare.
Upon closer inspection of the pandemic's growth, the Harvard Business Review reveals that the main reason why the virus was not contained in its infancy is because our global economy and health care systems are "geared to handle linear, incremental demand," while COVID-19 grows at an exponential rate. Our national health system cannot keep up with this kind of demand, so naturally, many researchers have turned to technology, specifically AI, to find solutions that are both effective and have global coverage.
The first step in handling any virus is being able to diagnose it properly. According to the MIT Technology Review, COVID-Net does exactly that. Developed by Linda Wang and Alexander Wong at the University of Waterloo and the AI firm DarwinAI in Canada, COVID-Net is a convolutional neural network that can help spot COVID-19 in chest x-rays. A convolutional neural network (CNN) passes an image through a series of fully connected layers to get a classification. This makes image classification significantly easier; image classification is a process in computer vision that can classify an image into one of many pre-established categories according to its visual content. For example, a CNN used for image classification may be able to classify an image as a dog if it contains four legs, ears, a tail, fur, and other dog-like characteristics.
In terms of the training database for the model, COVID-Net, specifically, was trained on 6000 images from varying lung diseases. In general, the more data you train your model on or the larger your dataset is, the more accurate your results will be. In terms of the percent of total data that should be used for training versus testing, the ratio is typically 80:20. Reserving 20% of data only for testing prevents an common issue in ML called overfitting, which is when the model reflects the training data too well, consequently performing poorly on the testing data.
Once you identify who has the virus, the second step is stopping the spread of the virus. According to the MIT Technology Review, Andrew Ng's startup called LandingAI can alert people if they are not "social distancing," or staying six feet away from another person. Once embedded in a security camera system, a trained neural network identifies people from a scene in real-time and after being calibrated to real-world dimensions, a second algorithm computes the distances between them.
In addition to addressing COVID-19, AI can also be used for several healthcare-focused solutions. For example, doctors at OrthoAtlanta have recently begun to use Suki, which is "an AI-powered, voice-enabled digital assistant for doctors that is designed to ease the burden of documentation." This enables doctors to focus on treating patients, rather than writing down notes while the patient is speaking. Suki effectively understands voice commands and uses them to create clinically accurate notes that upon confirmation, are inputted into the electronic health record system. Clearly, as AI advances, more healthcare systems are leveraging its power, revolutionizing the way that doctors treat their patients and diagnose illnesses.
That being said, despite the abundance of benefits that the use of AI provides, it is important that consumers are always wary of the accuracy of their AI-generated advice. In fact, the most pervasive issue is that people might take AI-generated advice too seriously. For example, if an AI system tells someone that their symptoms pertain more to the normal influenza rather than the Coronavirus, the individual should seek medical help if they are still concerned or feel otherwise. It should be known that currently, AI is not advanced enough to substitute for a physical checkup by a human doctor. So while AI can help you understand or be a base guideline in interpreting your symptoms, it is in no means a replacement for humans; always take advice with caution.
Second, the use of AI presents ethical issues around storing personal data. Although LandingAI is an incredible application of machine learning to increase safety, it also raises many privacy concerns. Specifically, since LandingAI gathers data real-time from real people, there must be an appropriate consent process. It is unethical to film anyone without their consent, so before being implemented in any public setting, all users must agree to be recorded for safety purposes, or else their personal rights are being violated.
Every day, breaking news stories regarding COVID-19 are on headlines of every news station. With so much information being provided to us, it can be difficult to figure out which sources of media to trust. If you are interested in reading specifically about the latest coverage of the Coronavirus and tech, subscribe to the New York Times Coronavirus Briefing or the Algorithm from MIT Technology Review Coronavirus Newsletter for a reliable rundown of the newest updates! Also, make sure to practice social distancing and to adhere to the safety guidelines given by the CDC and the WHO.
A final question for the readers: If you could apply AI to either the detection, spread, or treatment of COVID-19, which would you pick? What ideas do you have in applying any machine learning principles to increasing social welfare and decreasing the growth in the number of cases? Also, do you think that the pros of using AI outweigh the cons? I'd love to hear from you. Feel free to share your thoughts in the comments below!
Thank you for reading and stay safe!
If you are interested in reading more about the research I mentioned in the article specifically, check out the following links!
Photo by Sean O. on Unsplash
Note: this post is geared towards middle- and high-school students.
In the past few months, many middle- and high-school students around the country have had their summer camps, jobs, or other activities cancelled due to the COVID-19 pandemic. If you’re one of them, you’re probably wondering: What now? Here at Allgirlithm, we want to help you make your summer fun and rewarding. Here’s a list of computer science and artificial intelligence-related activities you could try this summer:
1. Expand your Coding Portfolio
One thing you could do this summer is expand your coding portfolio! You could do so by working on side projects—for example, apps, games, or algorithmic puzzles. This will often expose you to new technologies and solutions, and improve your coding abilities and confidence. Here are specific ideas, along with some possible programming languages and IDEs:
And if you ever find yourself stuck on a coding or design problem, you can always consult StackOverflow or Reddit!
2. Create a Personal Website
Creating a personal website is one of the best ways to showcase the work you’ve done and highlight your accomplishments. If you’re completely new to web development, you could give Weebly and Wix a try—both are easy-to-use, WYSIWYG (“what you see is what you get”) website builders. Or, if you’re feeling a bit more ambitious, you could try Wordpress, a blog-focused service that allows for a bit more customization.
3. Conduct Remote and/or Self-Guided Research
Another thing you could do this summer is conduct research at home! Your research could be related to anything, but here are some CS/AI-related ideas:
4. Contribute to Open-Source Projects, including Coronavirus-Related Projects
GitHub has millions of public repositories for open-source projects you might want to contribute to! These include everything from fun and quirky web extensions to IDE plugins to machine learning software packages for technologies like Tensorflow.
Tons of different coronavirus-related projects have emerged in light of the pandemic. You could look through some open-source repositories—many of them are tagged “coronavirus”, “covid”, or something similar—and contribute your code or non-technical work to any that interest you. This Github repository has a list of coronavirus-related projects; be sure to check it out!
5. Participate in Online Hackathons and Coding Competitions
If you go to Devpost and scroll down, you’ll see a list of online hackathons taking place in the near future—for example, hack:now and HackDSC. Take a look at these and see which ones you’re interested in! Many of them offer cash prizes and tech gadgets for winners.
A number of websites host coding competitions on scheduled days throughout the year, as well as training programs and practice problems open to users anytime. Some examples include Topcoder, CodeForces, CodeChef, HackerRank, USACO, and USACO’s Training Gateway. Training for and participating in coding competitions will improve your algorithmic thinking and problem-solving skills, which will help you in many of your other endeavors!
6. Take an Online Course
Coursera, edX, and Udemy are all great websites offering a variety of online courses. Codecademy is also offering Codecademy Pro for free during the remainder of this semester.
If you want to try some college-sponsored courses, MIT OpenCourseWare, Harvard Online Learning, and Stanford Online are great places to start. One of the most popular CS MOOCs of all time is Harvard’s CS50 class, an introductory class in programming, algorithms, and data structures.
Note: many of these activities require Internet connection. Some Internet service providers are offering free or low-cost programs. Here are a few links to information that may be of interest to you; it may also help to check with your school district or city/state government for more information.
1. Your Guide to Internet Service During New Coronavirus (COVID-19) Outbreak
2. Get online during the coronavirus outbreak
3. Comcast, AT&T, Sprint offering free or low-cost internet for students amid COVID-19 crisis
Image Credit: Complex Magazine
By Ore James
With most teens stuck at home in the midst of a pandemic, Netflix and other streaming services have become somewhat of a refuge. As browser extensions like Netflix Party facilitate remote interactions with friends, the relative normalcy provided by streaming may explain its popularity among the quarantined - of course, the data science behind these recommendations also plays a role in keeping us glued to the platform.
Most familiar with Netflix have probably seen genres like “critically-acclaimed movies about friendship” or “comedies for hopeless romantics” mixed into their homepages. However odd these micro-genres may seem, there exists a solid method behind them - personalized recommendations rely on machine learning algorithms to keep subscribers engaged, aiming to prolong our binge-watching sessions.
Recommendation systems are simply platforms to suggest content based on existing user preferences. The Netflix recommendation system compresses its large streaming library into personalized, easily-navigable rows using machine learning.
Within machine learning, systems regularly rewrite their algorithms - or data-based instructions - according to user data. Essentially, the systems collect data from users, learn from the data, and apply what it learned to make decisions. As Todd Yellin, Netflix’s VP of product innovation, told Wired in 2017, “What we see from those profiles is the following kinds of data — what people watch, what they watch after, what they watch before, what they watched a year ago, what they’ve watched recently and what time of day.” In addition to this data, Netflix relies on man-made tags, which categorize the service’s content, to determine the types of media users prefer. Taking all of this into account, machine learning algorithms interpret these data sets to ultimately decide which content to recommend.
Machine learning can take on highly specific forms to maximize user experiences. Recently, for example, Netflix incorporated an artwork-based algorithm to further personalize recommendations. The algorithm uses user preferences to determine which artwork will appear next to movies and shows. Machine learning comes into play when the algorithm chooses the art - data suggesting a user likes horror movies, for instance, may compel the algorithm to choose dark and chilling artwork for “Stranger Things.” By incorporating artwork, the algorithm demonstrates data can combine with a sense of creativity to further increase user engagement.
Image Credit: UX Planet
The downside to the Netflix approach of highly specific suggestions, as many subscribers have observed, is the fact that data from a customer's watch history often fails to reveal their actual tastes. For instance, within the current system, users who accidentally click on a documentary may face homepages cluttered with docudramas they have no interest in watching. The potential for users to miss out on content - or, conversely, face a homepage of content they don’t want to see, presents a glaring flaw in recommendation systems that can ultimately harm the subscriber experience. Content creators have voiced similar concerns - when Netflix canceled sitcom “Luca and Bertie” after one positively reviewed season in 2019, show creator Lisa Hanawalt pointed to the algorithms as a cause for low viewership.
Of course, machine learning is just one subset of artificial intelligence. Deep learning is a sector of machine learning in which a machine uses artificial neural networks, inspired by the brain’s neural networks, to “train” itself to make more accurate predictions. In practice, deep learning enables artificial intelligence to “think” and learn. While the technology is more commonly used to identify photos or audio, services like Movix.ai employ deep learning to recommend movies by adapting to user preferences in real-time, aiming for more accurate movie recommendations. Netflix itself seems to be slowly moving away from strict machine learning, following competitors like HBO Max; in August 2019, the service began beta-testing a Collections section, which relied on humans, not algorithms, to group titles for users.
While Netflix’s complex algorithms currently sort its thousands of titles well enough to keep many of us on the platform, it’s clear the future possibilities of machine learning, deep learning, and creativity in streaming are endless.