Algorithm Definition in Computer Science

Algorithm Definitions

• An algorithm is a set of instructions for solving a problem or completing a task, following a planned sequential order.
• A program consists of an algorithm and a data structure.
• Software requires more than just an algorithm; it also needs data to be computable.

Properties of Algorithm

• Finiteness
• Definiteness
• Input
• Output
• Effectiveness

Machine Learning

• The notion that we can program a computer to learn itself.
• Machine learning is a type of algorithm that can solve things that we don't know how to do.

Power and Datafication

• Datafication refers to the practices and relations through which data is constituted and made legible.
• Data is fundamental to algorithms, but it is often overlooked.
• Data positivism suggests that data has a unique epistemic role, but critics argue that it overlooks biases and errors in data-driven systems.

Algorithmic Organization

• The way an algorithm organizes and orders the world is never neutral.
• Algorithmic power refers to the epistemic challenges of machine learning, including the question of agency, responsibility, and accountability.

Criticisms of Artificial Intelligence

• The term "artificial intelligence" is often seen as an empty word, and some argue that it is not a technology but an ideology.
• Data-driven systems are often perceived as magically unknowable and absolute in their prediction patterns.

Machine Learning Types

• Rule-based learning: a set of rules is provided, and the computer completes the task.
• Machine learning: can solve things that we don't know how to do.
• Supervised machine learning: requires labeled data, which can be expensive.
• Unsupervised machine learning: finds patterns in data without labels, but requires a lot of data.

Algorithmic Bias

• Historical bias: a form of bias in the data.
• Data set bias: a poor data set that produces bias.
• Algorithmic bias: the amplification of gender, race, ability, and class discrimination by algorithms.

Black Box Effect

• Dealing with algorithms raises questions of responsibility and accountability.
• The complexity of algorithms makes it difficult to understand how they work.
• Companies often hide information about their algorithms from the public.

Correlation vs. Causation

• Correlation does not imply causation in machine learning.
• Machine learning replaces scientific proof and normativity with automated correlation.### Machine Learning and Labor
• Al is a new regime of truth, changing our understanding of knowledge and thinking, and we are starting to think that causality is not the only way to approach the world.
• Machine learning is based on prediction without causality, and workers are often treated like robots, with exploitation and lack of fair negotiation.

Worker Conditions

• Workers in companies like Amazon face stress, physical strain, and poor working conditions to meet picking rate quotas.
• Workers are considered less valued or trusted compared to robotic systems and are often treated like robots.

Human-Al Collaboration

• Human-Al collaboration often lacks fair negotiation, and workers are often exploited due to interchangeable roles and low compensation.
• Al development relies on the exploitation of human labor across the supply chain.

Amazon Labor

• Amazon presents a positive image to the public with incentives and slogans, but workers voice concerns about physical strain and scheduling issues.
• Al and algorithms intensify worker exploitation across various sectors.

Annotation and Hidden Labor

• Annotation is critical but often seen as temporary or unimportant by engineers focused on Al model building.
• Annotation tasks include labeling emotions in videos, classifying spam, and correcting Al errors.
• Annotation is often underpaid, except for chatbot training, which is a lucrative role requiring specific skills.

Leisure Society and Al

• Al is automating human labor at a speed that's never been seen before, leading to mass unemployment.
• There are two views on the impact of Al: the Utopian view, which sees increased leisure, quality of life, safety, and security, and the Cystopian view, which sees economic inequality, loss of purpose, surveillance, and control.

Types of Hidden Labor

• Data annotation: Labeling training sets, human workers have to do this in order to train the algorithm.
• Ghost work: Unpaid workers contribute to building, maintaining, and testing Al systems.
• Content moderation: Preforming crucial tasks like accessing harmful content, but are poorly paid and can suffer psychological trauma.

Asymmetric Distribution of Benefits

• The benefits and disadvantages of technology are not equally distributed, and big tech companies use branding to hide what they're doing.
• Algorithmic management is used to manage labor better, not replace it, and involves time management, decomposition, and surveillance.

Amazon Mechanical Turk

• Researchers buy labor from workers without knowing the conditions and pay, but get their tasks done.
• Exploitation of cheap labor in other countries (outsourcing) is a form of hidden labor and digital taylorism.

Al Development and Capitalism

• Capitalist Al is driven by profit, and technologies are being displayed for profit.
• Communist Al, on the other hand, aims to serve social wealth, with evenly distributed disposable time.

The Calculation Problem

• An algorithm that distributes knowledge: the free market is better able to handle the impossibility of one agent having all the information.
• Social organization and information distribution are key to solving the calculation problem.

Reconfiguration of Societal Structures

• Al transforms societal structures, power dynamics, labor, governance, ethics, and cultural practices.
• The debate reshapes the economy in response to profound changes by Al and automation, aiming to maximize Al benefits and minimize negative impact.

Bias in Machine Learning

• Machine learning data can be biased, and it's not just the data's fault, but also the machine learning pipeline.
• Bias can arise from the choices made during the pipeline, and it can be allocative, representational, historical, or deployment-based.

Types of Bias

• Allocative harm: When opportunities or resources are withheld from certain people or groups.
• Representational harm: When certain people or groups are stigmatized or stereotyped.
• Historical bias: When the world as it is or was leads to a model that produces harmful outcomes.
• Representational bias: When the development sample under-represents some part of the population.
• Evaluation bias: When the benchmark data used for a particular task does not represent the use population.
• Deployment bias: When there is a mismatch between the problem a model is intended to solve and the way it is actually used.

Fairness and Remedy

• Fairness algorithms modify the modeling process to satisfy particular notions of fairness.
• To remedy bias, we need to identify or diagnose bias in a system and develop methods to avoid bias and correct it when we see it.

Training Sets

• Training sets are the foundation on which contemporary machine-learning systems are built.

• Emergent bias arises in a context of use with real users.

• Prexisting bias can be entered consciously or subconsciously into computer systems.

• Technical bias arises from technical constraints and considerations.### Data Sets and Algorithmic Bias

• Data sets consist of a collection of images labeled and sorted into categories, but lack of representation and language models can lead to underperformance.

• Bias in data labeling can promote harmful stereotypes and representations, and algorithmic bias can occur due to interconnected factors, including aggregation of biased data.

• Bias cannot be simply fixed; it requires a comprehensive understanding of the social context, imperfect processing, and unclear definitions of fairness.

Model Cards

• A model card is a concise document that provides essential information about an AI model, including its development, intended use cases, and details about its performance.
• Model cards can help address bias in AI models by providing transparency and understanding of the model's limitations and potential biases.

Language and Bias

• Language models can perpetuate bias towards languages that are already well-represented, leaving out languages that are not well-represented on the internet.
• Researchers are working to create new data sets that prioritize underrepresented languages and address bias in language models.
• Joy Buolamwini's work on algorithmic bias highlights the importance of addressing bias in facial recognition algorithms, particularly for Black women.

Hallucinations and Environmental Effects

• Hallucinations in language models refer to the gap between statistical production of language and real-world correspondents.
• Large language models require substantial energy and result in high CO2 emissions, disproportionately affecting marginalized communities.
• Social movements can reshape language and norms, but large language models trained on static data may not reflect these changes.

Mitigation Strategies

• Prioritizing environmental impact assessments in model research is crucial to minimize the environmental effects of large language models.
• Budgeting research for dataset curation and documentation can help reduce bias and promote more accurate language models.
• Intersectional bias in language models can be addressed by prioritizing diversity and representation in dataset curation.

Foundation Models and Lovelace Effect

• Foundation models are a type of machine/deep learning model that starts with a large dataset and finds patterns, providing a general-purpose model for a wide range of tasks.
• The Lovelace effect refers to users perceiving computing systems as original and creative, rather than simply doing what they were programmed to do.
• Foundation models require a whole ecosystem, including large datasets and tons of data, to function effectively.

Dangers of Large Language Models

• Large language models have an unfathomable environmental cost due to the huge carbon emissions required to power data centers.
• The size of datasets does not guarantee diversity, and limited curation can lead to bias and unequal representation.
• Transparency and accountability are essential to critically assess the impact of large language models.

Pandemic and Education

• The pandemic has accelerated the shift to remote online learning, further exacerbating existing systemic inequalities in education.
• AI systems in education aim to present visions of a particular type of learner, but critics argue that this approach reproduces inequitable societal structures.
• A more radical approach to education is needed to address the root causes of inequity and promote more just and equitable outcomes.

Surveillance and Bias

• Surveillance technology in classrooms is widely accepted, but students rarely know or consent to it.
• These systems can give false positive rates for LGBTQ-related terms, claiming they are hate speech.
• Diverse voices need to be involved in the design process to address bias and promote fairness in AI systems.

Algorithmic Controversies

• Algorithmic controversies, such as the UK's use of algorithms to predict A-levels, have led to biased outcomes, particularly for disadvantaged groups.
• Biopolitical AI systems model students' socio-emotional states to influence desired behaviors, raising concerns about control and bias.
• Proctoring software has been shown to not recognize darker skin, highlighting the need for more diverse and inclusive AI systems.