We can utilise MATLAB functions for this. In my case, I used some data on MATLAB and obtained a six-degree polynomial expression (you may add Gaussian noise to the data). As readers, you can repeat the same on MATLAB easily. Then, we can use this polynomial function to estimate the time taken for the ball to reach the floor when it is released from 50 m. We can also use the kinematic equation (s = ut +0.5at2) to calculate the time taken for the ball to reach the floor when released from 50 m. When we compare the result of the polynomial function and the kinematic equation, we can observe some errors between the data, but also a certain pattern that both data from the polynomial fitting and the kinematic equation follow.

Now, let’s recall what we did and did not do. In polynomial fitting, we did not use or input the parameter for gravitational acceleration (g = 9.81). We also did not obtain any kinematic equations. We only observe the height and the time to create a non-linear relation between the two variables. Therefore, we did not utilise UG or LM and their derivation. We only used polynomial fitting to find a relationship between the data, which happens to be non-linear. We are unaware of any relation between the mass of the iron ball and the time taken. We are also unaware of the impact on time when external forces, such as air resistance (speed and direction of wind), are exerted on the ball. We can also dispense the standard metric units like seconds, meters, and even more complex units (ms−2) and use more subjective measurements.

But do these issues matter when using the polynomial expression in real-life free-fall applications? We can add more parameters (mass and air resistance) to fit the polynomial equation and find a relation between the time and the distance (height). This is a fundamental principle of non-scientific thinking: thinking without any core knowledge (no scientific theory) of objects and how they are interrelated in the objective world [3]. It is a method of thinking that is purely based on creating non-theory relations between objects. Some can argue mathematics (abstract concepts) has been used here to create a non-linear relation. However, the mathematical operation used here is simple mathematics to find a relation between numbers. Another argument would be that this experiment belongs to the empirical observation in science. I agree that this is a scientific experiment, but what exactly is the conclusion of this experiment? Still, it absconds from a proper hypothesis fundamental to scientific thinking, which means this experiment is inspired by just day-to-day observation of free-fall objects. Also, it lacks a plausible scientific explanation of why this non-linear relation between time and height exists. This is not explained by a ground-breaking scientific discovery like the universal gravitation.

Part 2: A Box in Black

The OceanGate’s Titan implosion has been a major unfortunate incident reported by the media recently. The implosion has been attributed to different causes, from material failure to communication loss [4]. However, I was gobsmacked to learn that the Titan had made it to the Titanic wrecks in previous explorations. This fact has been overridden by the explanations of how scientific tests have not been adequately conducted during the manufacturing of the Titan. Had the Titan made it to the Titanic wrecks again, this would not have been news!

I am not implying that conducting scientific tests after the implosion is unimportant at all. But there should have been proper scientific conduct on how the Titan made it to the Titanic wrecks before, with possible failures of the submersible. Now, a “Box in Black” or absence of scientific explanation as to how the Titan explored the Titanic wrecks despite its failures that were discovered after the implosion, is present.

The power of science can sometimes be limited by the individuals who conduct it. This is because there would be biases that many would follow unconsciously and are mostly limited to current scientific explanations and experiments. Artificial Intelligence (AI) can be applied to fill this gap where certain biases and ambiguities need to be redeemed.

Let’s take a simple AI, or more precisely, the machine learning (ML) application. An image of a cat needs to be identified correctly from other animals by an ML algorithm. The first step is to acquire a collection of cat images and preprocess them to obtain data regarding rotation, flips, and colour contrast. Next, features such as shape (a cat shape), colour (different colours specific for cats), and unique features (fur, paws) would be extracted by the algorithm. Then, the ML algorithm would train the model to extract cat images from non-cat images. Humans can be involved in any of these processes; they can decide what features should be extracted. After the model is trained, it suddenly, from nowhere, can identify cat images.

Nevertheless, the beauty of ML lies in its architecture. It can occupy different architectures, but a Convolution Neural Network (CNN) would be the most appropriate one to utilise here [5]. CNN is a deep learning architecture where features are selected by feature extraction. It has several layers for functions such as preprocessing and applying non-linearity to emphasise certain features, downsampling, dropping out (retaining only prominent features), generating complex features (dense layer), and calculating the error. Previously, I mentioned humans can be involved in identifying which features need to be extracted, but in a CNN-based ML algorithm (unlike in a shallow learning algorithm, like linear regression, where features might be predefined by humans), the algorithm itself identifies them [6].

Figure 2: An architecture of CNN to identify written numbers [7]. Image by Analytic Vidya

“Artificial intelligence is growing fast, as are robots whose facial expressions can elicit empathy and make your mirror neurons quiver.” — Diane Ackerman

The main two aspects emphasised in Part 1 and Part 2 are,

1. Artificial intelligence (AI) is just a pattern recognition system.

2. The involvement of scientific theories is absent or extremely limited in AI.

The “Black Box” here mainly lies in the feature extraction and its relationship to the functions used and weights on each connection. In this ML algorithm, no individual can know why each weight was assigned to each connection between neurons to create a relationship between each feature. The features we are discussing could be much more complex, derived from surface features, such as shapes or colour. While it is possible to identify these features after execution, why they were considered more prominent is not evident. Additionally, the functions (such as ReLU, sigmoid, and tanh) employed for extracting features are paramount in deciding which features are prominent [5-6]. It is unclear why the functions remove some features while retaining others. The architecture will use different weights on each connection between neurons to identify and minimise the training error. Therefore, the prime target of the model is to minimise the error in assigning values on each connection. Complex features are just a result of reducing error rather than following specific guidelines to identify a feature. If we revert to the news on the Titan implosion, I always wonder if scientists really need to analyse the Titan based on scientific theories such as theories on fluid dynamics, pressure, material sciences, electromagnetism, and signal processing. Or else, they can just insert the data they have already acquired from the previous explorations of Titan and let an ML algorithm decide what features, say, to reinforce the hull, a submersible should comprise. Suppose the ML is input with data from previous explorations such as pressure, temperature changes, and the design deformations that occurred during exploration. Will it output a hull design based on creating complex connections between data?

There could be arguments, such as the AI model was developed on the basis of science and logic. While it has a regular architecture that follows, it does not mean the output of AI can be routed back to the input through its architecture. Most AI experts are now convinced that AI probably would not be entirely understandable forever. Some precautions for AI ambiguity exist, such as explainable AI (XAI) [8]. These models consist of simple architectures that are more transparent and meaningful in interpretation. I request you to explore the topics of XAI and its recent development to realise the differences between AI and XAI.

What would be the immediate impact of AI on the science? I would say nothing because scientific theories and methods are still popular. Theoretical physics has evolved from classical physics, and AI is not substantially involved in the experimentation of theoretical physics yet, especially in quantum mechanics. AI has the capability to replace many realms of classical physics, but that is in the future’s hands to decide. It should be noted that AI is already more utilised in economic science and business environments, where more uncertainties are present [9]. As an example, AI is being used for stock price predictions where prices are highly unpredictable. However, as I mentioned in Part 1, physics theories make precise predictions of the future based on formulas. Economic theories, on the other hand, are not natural science theories like physics. They involve individual psychology, societal aspects, government interventions, and global alterations like the recent pandemic, which are highly uncertain. AI is rapidly used for economic predictions based on different atmospheres. If AI can be successfully used in such uncertainties, can’t it replace the scientific analysis of different phenomena and explain them better than scientific theories?