What are Underlying Assumption in AI?

In this blog, We will learn about the underlying assumptions in Artificial intelligence. To understand this topic better will also discuss the overview of AI, the role of the data, the computational demands of AI, and the challenges of AI related to accessibility and scalability. Also, we will look into how the biases are addressed in AI Technologies. The impact of assumptions is also explained in depth to understand each topic better.

Overview of Artificial Intelligence:

Artificial intelligence refers to the technology that deals with performing tasks that mimic the cognitive functions of the human brain. AI involves the computer system development that can enable them to perform tasks that usually require human intelligence.

Primary Technologies Involved in AI :

AI is a broad spectrum of distinct technologies that have various applications in different sectors. Some of the primary technologies involved in the AI are given below,

Machine learning:

Machine learning is mainly focused on the study of statistical algorithms that can learn patterns and valuable insights from the data and generalize new unseen data. This is a data-driven approach that uses data from the past to learn and predict on the new data.

Deep learning:

Deep learning mainly focuses on the simulation of neuron patterns similar to the human brain to perform complex tasks and make decisions as humans do. The “deep” refers to the multiple layers present In the neural networks in the models developed based on deep learning. The CNN, RNN, and LSTM are some example algorithms that come under deep learning.

Natural language processing(NLP):

Natural language processing, as the name suggests, enables machines to understand the natural language of humans as it is spoken. This is an important technology used in artificial intelligence to make communication between machines and humans easier, more convenient, and more interactive and user-friendly. This enables the machines to interpret and comprehend the human language.

Computer vision:

Computer vision enables the ability of machines to simulate the visuals like humans to analyze and interpret the environment visuals. This technology is used to classify the objects and extract useful insights from the images or videos. It is used in tasks such as object detection, image segmentation, etc. This also enables the machines to understand the images and interpret gestures.

Data Mining:

Data mining is discovering valuable patterns and insights from large datasets. It is used in machine learning to find hidden patterns in data and this technique is a combination of artificial intelligence and statistics. Data mining is an important concept in data analysis as it provides tools to interpret and analyze the data.

Reinforcement learning:

This technique enables the machines to learn from their past mistakes and not repeat them. This involves an agent that interacts with the environment and receives feedback and these feedbacks can be successes or failures of the results. They are considered the rewards and penalties. So these modify their decisions based on the feedback received. This is used in the decision-making process to get the most optimal outcome.

Components of AI applications:

AI applications generally involve three main components that help in assuring the proper development and implementation of AI systems. These components need to be well structured and should be validated. The key components of the AI application are given below,

Data:

The Data is the most crucial component for the AI application because the training of the AI system highly depends on the data. After all, the systems learn the patterns from the data which further helps them in making decisions or predictions accurately.

The AI systems are designed and developed in such a way that they learn from the data. Usually, the data is a combination of human reviews and pieces of information. The data used for training should not contain any algorithmic bias and should be ethical and fair. The data also should be free from noise and should be accurate avoiding the inconsistencies in the data.

Algorithm:

The algorithm can be defined as a set of instructions that give a step-by-step procedure for an AI system to solve a specific issue and give an output for the input data provided.

The Algorithms used in the AI field are usually complex and contain mathematical codes that enable the machine to learn from the new input data and develop an accurate output for that. So basically the algorithm is programmed to learn to perform a task rather than just performing a task.

Human Interaction:

Human involvement is a key component for the better development of the AI systems. Preparing the data programming the algorithms and also verifying the output needs the involvement of the human.

Human reviews are crucial while the data is collected and prepared to keep the data appropriate for the system. They are also important in verifying the output of algorithms for relevancy and accuracy. The absence of human reviews may lead to incorrect results of the AI systems that make the system inefficient, and more risky to use.

The Role of Data in AI:

As we saw earlier, Data is an important component in building an AI system.

The data lays the foundation for AI systems and enables them to learn, adapt, and evolve. In the context of artificial intelligence, Data equips the systems with the mandatory information that helps in understanding the underlying patterns and make decisions and also predict the outcomes.

The quality, quantity, and variety of data influence the algorithm’s and neural network’s effectiveness. It is important to know how the data is collected and processed and also how are they used in the development of AI systems. The data should not only be accurate but also fair, transparent, and unbiased.

The data can be of two types based on their nature: Structured and Unstructured data.

Structured data:

The structured data includes databases, and spreadsheets and is usually easier for systems to process as they are well-organized and formatted. They also enable easier data analysis and recognition of patterns. Structured data increases the ability of the system to interpret the data and learn effectively from them.

Unstructured data:

The unstructured data includes images, texts, and videos and is the most widely available data around us which may pose challenges for AI systems. This kind of data requires more complex techniques like natural language processing and computer vision. This has led to many breakthroughs enabling more natural interactions between machines and humans.

Computational demands of AI:

With the increase in development of the technology, there is a surge in the development of artificial intelligence systems which have experienced major transformation in past years.

A leading AI research organization named OpenAI has provided data that illustrates the computation demands of AI. According to the research, these demands are doubling every 3 months since 2012. This rate is faster than what was predicted in Moore’s law.

The necessity of parallel advancements in computation power to provide them the ability to learn vast datasets and make predictions has been escalating as there is an increase in the complexity and size of AI systems. The AI application has been used in various domains such as healthcare to autonomous vehicles. This demands high-speed processing systems that can handle large amounts of data in real time. This makes the computation demand a key factor in determining the effectiveness and scalability.

The AI is often referred to as a blessing in disguise. With the numerous amount of benefits of AI systems, there are some challenges that the usage of AI systems leads to. One such major challenge is that AI contributes to global carbon emissions as the energy requirements for running these systems are substantial.

Accessilbilties and scalability issues In AI:

Accessilbities issues:

Performance Inconsistencies:

The main issue in AI accessibility is the inconsistencies that occur in the performance. The AI models usually predict what they have learned during the training process and make associations with the current scenario to come up with a solution.

Acquiring the Inclusive data:

Finding the data of all the people can be very hard and disclosing the sensitive information of a person with a disability can be a disrespectful act. Not including certain people’s information can lead to underrepresenting other people in the dataset, leading to biased opinions in the AI systems. Synthetic datasets can lead to unethical and stereotyping behavior towards disabled people.

Scalability issues:

Data complexities:

The training of the AI systems requires a lot of relevant data records. These records of data have issues of data predictability and feasibility problems. Gathering relevant data is a very complex task because that data is contextual too. Data scientists usually spend most of their time cleansing and managing the data in traditional systems. So we need a proper system that can ensure data transparency. The cost of making AI models rises due to the huge amount of data complexities

Performance:

The technical performance that AI systems usually require is extremely intensive and also computer processing like matrix manipulation, linear algebra, and analysis in statistics. The results are obtained after repeated calculations which makes the process entirely time-consuming. This also shows the potential storage requirements and computer processing abilities are extremely intense. In the initial stage as there will be only small datasets it will be easier to maintain the performance of the AI systems. As the system progresses the requirements also become higher and performance may decrease.

Unpredictable behavior:

The nature of AI systems is prone to changes with changes in algorithms. So their volatile nature causes issues that are hard to identify and solve. There can be issues later after the testing processing is done in the system. Good contingency options can help in tackling unexpected behavior.

Security of the data:

Data security is another major challenge for AI systems as the cyber security features can’t be applied here. As all the data are stored in a single space here there are risks for privacy concerns. The safe data gets mixed with other data making them insecure. So proper privacy and security should be implemented in AI systems to protect the data.

Addressing biases inherent in AI Technologies:

The presence of bias can simply lead to unfair and unethical decisions that affect the accuracy of the systems. The mitigation of bias is an important process. The addressing of bias in mitigation strategies can be done in three ways

Preprocessing the Data:

The preprocessing of the data involves addressing bias before the training of the data. This technique ensures that the data used for training is representative of the entire population. This is done by using techniques like oversampling and undersampling, this helps in generating the representative data of the population that doesn’t under- or over-represent the data.

This also considers the historical patterns of marginalized groups but this technique can be time-consuming and may not be effective if the data is biased already. There are also some privacy concerns related to the retrieval of historical patterns. The implementation of this technique is done by using bias correction algorithms and data augmentations.

Model Selection:

This technique focuses on prioritizing fairness while the model is being selected. The selected model should be capable of analyzing the data properly and the selection is usually done by the researchers and they use methods based on the group fairness. This approach uses the regularization methods that penalize the model when the model makes discriminatory decisions. This method also uses techniques like group fairness and individual fairness.

The ensemble techniques can be used to combine various models but the individual fairness should be applied to all the models so that the combined model can reduce the bias. So this technique may lack consensus on what fairness constitutes.

Post Processing Decisions:

This technique involves the adjustment of the output of the AI systems. It alters the decisions made by the AI systems so that the decisions are more fair and ethical. This also includes the implementation of recalibrating or reweighting. This aims to achieve equalized odds and make sure that false positives are equally distributed among the demographic groups.

The changes made to the output of the model may affect the fairness among the demographic groups hence this can be complex and a huge amount of data is required for this. There is also a trade-off between the bias when the adjustment is made.

Impact of Assumptions on AI Development:

To understand the impact of assumptions on AI development better, we need to learn about the what are underlying assumptions and historical perspectives of AI and also about the epistemological beliefs, and recalibrating assumptions in AI.

What are the Underlying Assumptions in AI?

The Underlying assumptions in AI revolve around the idea of attempting the simulations of intelligence in machines like humans. It also states that algorithms can help the machine learn from the data.
The data used to train the algorithm can provide real-time insights and patterns that help them promote human-like cognition thinking in machines.
This assumes that machines can understand human thoughts, use the data to learn things, and apply logic and reasoning to solve problems like humans do.
The machine can also learn from their past experiences and receive feedback on the areas to improve. The machines can make their own decisions and make predictions unexpectedly or unpredictably.

Historical perspectives of AI:

The history of artificial intelligence begins with a lot of assumptions, myths, and confusion. The idea of artificial intelligence emerged when philosophers tried to describe human cognition as a mechanical manipulation. Later in the 1940s, a programmable digital computer was invented and this technology provided a practical demonstration of mathematical reasoning proving the point of the philosophers.
This device leads to the inspiration to build an “electronic brain” i.e. software that can think like humans. Then further workshop was held at Dartmouth College in the USA in 1956. Many researchers believed that there would be computer systems that would be as intelligent as human beings and funds were provided. Later, the researchers discovered the difficulty of the project. In 1974, due to huge criticism and pressure, the funds were stopped.
The theories based on artificial intelligence state that “The best model of intelligence is the human brain itself”. That is the human brain can be a perfect example to build AI systems. The idea of artificial intelligence was based on the assumption that the cognition thinking of humans can be mechanized. This assumption underlying AI is that human thoughts can be mechanized. This idea led back to exploring the formal methods of reasoning such as deductions made by ancient philosophers.
The main challenge for building AI systems is that thinking can be put into mathematical rules and how math is used to reason logically. Boole and Frege showed how math can be used in logical reasoning in the 1900s. Later it was found that math has its limits and there are few things that math can’t be used in. That’s when the truing machines came into the picture and showed that any kind of math reasoning can be done by following a set of rules.
Further, when researchers started to study they found that the brain was a network of neurons that network was an electrical network and this led to thinking that the construction of an electronic brain was possible. Later the scientist Alan Turing proved that any computation can be described digitally and circulated a paper on machine intelligence. In 1950 Turing published a paper named “ Computing Machinery and Intelligence” which specified the possibility of creating machines that can think and also introduced a concept called the “Turing test”.
The Turing test is a test conducted between a human and a machine where the machine and humans are placed in different rooms and they are examined, if the machine’s answer is similar to the answer of the human being. This test was the first serious step in the philosophy of artificial intelligence and is widely accepted. In 1943, Walter Pitts and Warren McCulloch showed how the network of an idealized artificial neural network can be used to perform simple logical functions.

Epistemological Assumptions in AI:

The epistemological assumptions are an important concept under the underlying assumptions in artificial intelligence as they involve the philosophical study of knowledge and help in shaping the underlying assumptions. They hold the idea of how the data is obtained and represented in AI systems.

Assumptions:

Formalism: Formalism is an assumption made by AI researchers in the field of Symbolic AI systems where they assume that intelligence can be captured through some predefined rules and logical systems. This is a key concept in the epistemological assumptions. This follows the logical structures and rules and helps in shaping how the data or knowledge will be represented and used in AI systems.
Data: The data is the most important source for AI systems. The systems usually learn the statistical patterns from the data we provide and also learn about the correlations between different datasets used for training. So assumptions on this raise questions on the data-driven methods limitations. Some limitations of these data-driven approaches are finding algorithmic bias and lack of fairness in the dataset. This highlights the importance of providing a proper dataset that can train the AI system to be fair and unbiased.
Storage of Representation and Reasoning: The next important concept involved in this epistemological assumption is the separation of information that is the data and facts from the algorithms used to manipulate these data. In traditional AI systems, the data and facts are stored in separate places and the algorithms are stored separately. This way of storing is assumed to simplify the problem-solving processes as the represented data or facts and knowledge are usually in structured formats and AI systems can easily manipulate and organize them. But this could leave some important aspects of embodied cognition and situated knowledge.

Perspectives:

Embodied cognition: Embodied cognition refers to the cognitive processes that are influenced by the physical body and their interactions with the environment. This raises the importance of connecting the physical world with AI systems. The AI systems should be able to connect to the physical world and not only focus on abstract representations.
Integration of connectionism: This helps in blurring the line between the data and knowledge and introduces new ways to learn from the data. In connectionist systems, knowledge is not represented in the form of symbols instead they are represented in the form of artificial neurons connected in a neural network. The main aim of these artificial neural networks is to mimic the neural network structure of the human brain to promote thinking and problem-solving skills like humans. The neural networks are the computational models mainly used in connectionism and they consist of a network of interconnected artificial neurons as mentioned earlier. This enables the systems to capture the complex patterns by providing them a proper training using the proper dataset.
Transparent and Accountable AI systems: The AI systems are assumed to be accountable and transparent. This is crucial to ensure that the AI systems function without any biases and are fair to all its users. This promotes trust and makes the systems more understandable. This way the assumptions will continue to shape the future of artificial intelligence systems.

Recalibrating Assumptions in AI:

The AI policies are always based on some core assumptions, these assumptions are based on a wide range of factors including ethical principles and technologies. To keep these core assumptions valid and updated with time the recalibration of the assumptions is important. So that the policies remain fair and ethical and also remain valid with advancement of the technology.
The main problem with these common assumptions is that these assumptions are more like opinions than facts. Most of the assumptions are constantly repeated and may not accepted universally. To make sure the assumptions used to make the AI policy are fair they are based on the theories that are passed globally. These assumptions usually have to engage with uncomfortable counterpoints if they are not honest and transparent.
One of the common assumptions made by the majority of people is that AI systems have very great potential to perform any tasks that originally required human intelligence. This assumption ignores the major fact that artificial intelligence has a limited capacity to duplicate human intelligence. They have only been able to replicate the narrow facets of human intelligence. The AI systems lack empathy, creativity, and common sense.
Another assumption is that the primary source for AI development and deployment is data. The collection of huge amounts of relevant data can increase the capacity of AI Systems. This assumption may be partially true but data may be an important factor but it solely can’t increase the capacity of the AI systems. What matters the most is the quality of the data and the diversity of the data to have fair and unbiased systems. The collection of the data can be unethical in some sensitive domains like healthcare, etc where the data collection can lead to security and privacy concerns.
The next assumption is, that to succeed in international competitions, all countries should develop their AI systems faster. Now creating AI systems faster to succeed in a competition may only be successful for a short period and that decision may not provide long-term benefits. Rushing in the development of the AI systems may lack the considerations of safety, fairness, and accountability leading to potential risks and unintended consequences.
The next assumption is encoding the ethical principles into AI. Now this requires much greater measures that are beyond just the technical fixes and to create an ethical AI system we need to incorporate societal norms, human rights, and legal laws to get a fair and ethical AI system.

Conclusion:

The overview of the article gave a brief idea of primary technologies in AI and its components. The role of data was also discussed and challenges related to accessibility and scalability were also explained. Further, the article elaborated on the addressing of bias in the AI systems. Then we learned about the impact of assumptions by learning about the underlying assumptions. The underlying assumptions in AI can be elaborated by taking a brief look at the historical perspectives, recalibrating assumptions, and epistemological assumptions in AI. This is essential for navigating the development, and deployment of AI properly. By learning about the philosophical, ethical, and societal implications of these assumptions we can create fair and transparent AI systems. We also saw how important it is to recalibrate these assumptions to ensure the AI system policy and goals stay updated.