Showing posts with label Algorithmic Error. Show all posts
Showing posts with label Algorithmic Error. Show all posts

Artificial Intelligence - What Is Algorithmic Error and Bias?


Bias in algorithmic systems has emerged as one of the most pressing issues surrounding artificial intelligence ethics.

Algorithmic bias refers to a computer system's recurrent and systemic flaws that discriminate against certain groups or people.

It's crucial to remember that bias isn't necessarily a bad thing: it may be included into a system in order to fix an unjust system or reality.

Bias causes problems when it leads to an unjust or discriminating conclusion that affects people's lives and chances.

Individuals and communities that are already weak in society are often at danger from algorithmic prejudice and inaccuracy.

As a result, algorithmic prejudice may exacerbate social inequality by restricting people's access to services and goods.

Algorithms are increasingly being utilized to guide government decision-making, notably in the criminal justice sector for sentencing and bail, as well as in migration management using biometric technology like face and gait recognition.

When a government's algorithms are shown to be biased, individuals may lose faith in the AI system as well as its usage by institutions, whether they be government agencies or private businesses.

There have been several incidents of algorithmic prejudice during the past few years.

A high-profile example is Facebook's targeted advertising, which is based on algorithms that identify which demographic groups a given advertisement should be viewed by.

Indeed, according to one research, job advertising for janitors and related occupations on Facebook are often aimed towards lower-income groups and minorities, while ads for nurses or secretaries are focused at women (Ali et al. 2019).

This involves successfully profiling persons in protected classifications, such as race, gender, and economic bracket, in order to maximize the effectiveness and profitability of advertising.

Another well-known example is Amazon's algorithm for sorting and evaluating resumes in order to increase efficiency and ostensibly impartiality in the recruiting process.

Amazon's algorithm was trained using data from the company's previous recruiting practices.

However, once the algorithm was implemented, it became evident that it was prejudiced against women, with résumés that contained the terms "women" or "gender" or indicated that the candidate had attended a women's institution receiving worse rankings.

Little could be done to address the algorithm's prejudices since it was trained on Amazon's prior recruiting practices.

While the algorithm was plainly prejudiced, this example demonstrates how such biases may mirror social prejudices, including, in this instance, Amazon's deeply established biases against employing women.

Indeed, bias in an algorithmic system may develop in a variety of ways.

Algorithmic bias occurs when a group of people and their lived experiences are not taken into consideration while the algorithm is being designed.

This can happen at any point during the algorithm development process, from collecting data that isn't representative of all demographic groups to labeling data in ways that reproduce discriminatory profiling to the rollout of an algorithm that ignores the differential impact it may have on a specific group.

In recent years, there has been a proliferation of policy documents addressing the ethical responsibilities of state and non-state bodies using algorithmic processing—to ensure against unfair bias and other negative effects of algorithmic processing—partly in response to significant publicity of algorithmic biases (Jobin et al.2019).

The European Union's "Ethics Guidelines for Trustworthy AI," issued in 2018, is one of the most important rules in this area.

The EU statement lays forth seven principles for fair and ethical AI and algorithmic processing regulation.

Furthermore, with the adoption of the General Data Protection Regulation (GDPR) in 2018, the European Union has been in the forefront of legislative responses to algorithmic processing.

A corporation may be penalized up to 4% of its annual worldwide turnover if it uses an algorithm that is found to be prejudiced on the basis of race, gender, or another protected category, according to the GDPR, which applies in the first instance to the processing of all personal information inside the EU.

The difficulty of determining where a bias occurred and what dataset caused prejudice is a persisting challenge for algorithmic processing regulation.

This is sometimes referred to as the algorithmic black box problem: an algorithm's deep data processing layers are so intricate and many that a human cannot comprehend them.

Different data is fed into the algorithm to observe where the unequal results emerge, based on the right to an explanation when, subject to an automated decision under the GDPR, one of the replies has been to identify where the bias occurred via counterfactual explanations (Wachter et al.2018).

Technical solutions to the issue included building synthetic datasets that seek to repair naturally existing biases in datasets or provide an unbiased and representative dataset, in addition to legal and legislative instruments for tackling algorithmic bias.

While such channels for redress are vital, one of the most comprehensive solutions to the issue is to have far more varied human teams developing, producing, using, and monitoring the effect of algorithms.

A mix of life experiences within diverse teams makes it more likely that prejudices will be discovered and corrected sooner.

~ Jai Krishna Ponnappan

You may also want to read more about Artificial Intelligence here.

See also: Biometric Technology; Explainable AI; Gender and AI.

Further Reading

Ali, Muhammed, Piotr Sapiezynski, Miranda Bogen, Aleksandra Korolova, Alan Mislove, and Aaron Rieke. 2019. “Discrimination through Optimization: How Facebook’s Ad Delivery Can Lead to Skewed Outcomes.” In Proceedings of the ACM on Human-Computer Interaction, vol. 3, CSCW, Article 199 (November). New York: Association for Computing Machinery.

European Union. 2018. “General Data Protection Regulation (GDPR).”

European Union. 2019. “Ethics Guidelines for Trustworthy AI.”

Jobin, Anna, Marcello Ienca, and Effy Vayena. 2019. “The Global Landscape of AI Ethics Guidelines.” Nature Machine Intelligence 1 (September): 389–99.

Noble, Safiya Umoja. 2018. Algorithms of Oppression: How Search Engines Reinforce Racism. New York: New York University Press.

Pasquale, Frank. 2016. The Black Box Society: The Secret Algorithms that Control Money and Information. Cambridge, MA: Harvard University Press.

Wachter, Sandra, Brent Mittelstadt, and Chris Russell. 2018. “Counterfactual Explanations Without Opening the Black Box: Automated Decisions and the GDPR.” Harvard Journal of Law & Technology 31, no. 2 (Spring): 841–87.

Zuboff, Shoshana. 2018. The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. London: Profile Books.

Artificial Intelligence - How Are Accidents and Risk Assessment Done Using AI?


Many computer-based systems' most significant feature is their reliability.

Physical damage, data loss, economic disruption, and human deaths may all result from mechanical and software failures.

Many essential systems are now controlled by robotics, automation, and artificial intelligence.

Nuclear power plants, financial markets, social security payments, traffic lights, and military radar stations are all under their watchful eye.

High-tech systems may be designed purposefully hazardous to people, as with Trojan horses, viruses, and spyware, or they can be dangerous due to human programming or operation errors.

They may become dangerous in the future as a result of purposeful or unintended actions made by the machines themselves, or as a result of unanticipated environmental variables.

The first person to be murdered while working with a robot occurred in 1979.

A one-ton parts-retrieval robot built by Litton Industries hit Ford Motor Company engineer Robert Williams in the head.

After failing to entirely switch off a malfunctioning robot on the production floor at Kawasaki Heavy Industries two years later, Japanese engineer Kenji Urada was murdered.

Urada was shoved into a grinding machine by the robot's arm.

Accidents do not always result in deaths.

A 300-pound Knightscope K5 security robot on patrol at a retail business center in Northern California, for example, knocked down a kid and ran over his foot in 2016.

Only a few cuts and swelling were sustained by the youngster.

The Cold War's history is littered with stories of nuclear near-misses caused by faulty computer technology.

In 1979, a computer glitch at the North American Aerospace Defense Command (NORAD) misled the Strategic Air Command into believing that the Soviet Union had fired over 2,000 nuclear missiles towards the US.

An examination revealed that a training scenario had been uploaded to an active defense computer by mistake.

In 1983, a Soviet military early warning system identified a single US intercontinental ballistic missile launching a nuclear assault.

Stanislav Petrov, the missile defense system's operator, correctly discounted the signal as a false alarm.

The reason of this and subsequent false alarms was ultimately discovered to be sunlight hitting high altitude clouds.

Petrov was eventually punished for humiliating his superiors by disclosing faults, despite preventing global thermonuclear Armageddon.

The so-called "2010 Flash Crash" was caused by stock market trading software.

In slightly over a half-hour on May 6, 2010, the S&P 500, Dow Jones, and NASDAQ stock indexes lost—and then mainly regained—a trillion dollars in value.

Navin Dal Singh Sarao, a U.K. trader, was arrested after a five-year investigation by the US Department of Justice for allegedly manipulating an automated system to issue and then cancel huge numbers of sell orders, allowing his business to acquire equities at temporarily reduced prices.

In 2015, there were two more software-induced market flash crashes, and in 2017, there were flash crashes in the gold futures market and digital cryptocurrency sector.

Tay (short for "Thinking about you"), a Microsoft Corporation artificial intelligence social media chatterbot, went tragically wrong in 2016.

Tay was created by Microsoft engineers to imitate a nineteen-year-old American girl and to learn from Twitter discussions.

Instead, Tay was trained to use harsh and aggressive language by internet trolls, which it then repeated in tweets.

After barely sixteen hours, Microsoft deleted Tay's account.

More AI-related accidents in motor vehicle operating may occur in the future.

In 2016, the first fatal collision involving a self-driving car happened when a Tesla Model S in autopilot mode collided with a semi-trailer crossing the highway.

The motorist may have been viewing a Harry Potter movie on a portable DVD player when the accident happened, according to witnesses.

Tesla's software does not yet allow for completely autonomous driving, hence a human operator is required.

Despite these dangers, one management consulting company claims that autonomous automobiles might avert up to 90% of road accidents.

Artificial intelligence security is rapidly growing as a topic of cybersecurity study.

Militaries all around the globe are working on prototypes of dangerous autonomous weapons systems.

Automatic weapons, such as drones, that now rely on a human operator to make deadly force judgments against targets, might be replaced with automated systems that make life and death decisions.

Robotic decision-makers on the battlefield may one day outperform humans in extracting patterns from the fog of war and reacting quickly and logically to novel or challenging circumstances.

High technology is becoming more and more important in modern civilization, yet it is also becoming more fragile and prone to failure.

An inquisitive squirrel caused the NASDAQ's primary computer to collapse in 1987, bringing one of the world's major stock exchanges to its knees.

In another example, the ozone hole above Antarctica was not discovered for years because exceptionally low levels reported in data-processed satellite images were assumed to be mistakes.

It's likely that the complexity of autonomous systems, as well as society's reliance on them under quickly changing circumstances, will make completely tested AI unachievable.

Artificial intelligence is powered by software that can adapt to and interact with its surroundings and users.

Changes in variables, individual acts, or events may have unanticipated and even disastrous consequences.

One of the dark secrets of sophisticated artificial intelligence is that it is based on mathematical approaches and deep learning algorithms that are so complicated that even its creators are baffled as to how it makes accurate conclusions.

Autonomous cars, for example, depend on exclusively computer-written instructions while they watch people driving in real-world situations.

But how can a self-driving automobile learn to anticipate the unexpected?

Will attempts to adjust AI-generated code to decrease apparent faults, omissions, and impenetrability lessen the likelihood of unintended negative consequences, or will they merely magnify existing problems and produce new ones? Although it is unclear how to mitigate the risks of artificial intelligence, it is likely that society will rely on well-established and presumably trustworthy machine-learning systems to automatically provide rationales for their actions, as well as examine newly developed cognitive computing systems on our behalf.

~ Jai Krishna Ponnappan

You may also want to read more about Artificial Intelligence here.

Also see: Algorithmic Error and Bias; Autonomy and Complacency; Beneficial AI, Asilo mar Meeting on; Campaign to Stop Killer Robots; Driverless Vehicles and Liability; Explainable AI; Product Liability and AI; Trolley Problem.

Further Reading

De Visser, Ewart Jan. 2012. “The World Is Not Enough: Trust in Cognitive Agents.” Ph.D. diss., George Mason University.

Forester, Tom, and Perry Morrison. 1990. “Computer Unreliability and Social Vulnerability.” Futures 22, no. 5 (June): 462–74.

Lee, John D., and Katrina A. See. 2004. “Trust in Automation: Designing for Appropriate Reliance.” Human Factors 46, no. 1 (Spring): 50–80.

Yudkowsky, Eliezer. 2008. “Artificial Intelligence as a Positive and Negative Factor in Global Risk.” In Global Catastrophic Risks, edited by Nick Bostrom and Milan 

M. Ćirković, 308–45. New York: Oxford University Press.

What Is Artificial General Intelligence?

Artificial General Intelligence (AGI) is defined as the software representation of generalized human cognitive capacities that enables the ...