Patterns are everywhere, do you agree?

Edited using Canva, Background by Adrien

For some time now, I have been engaging in interesting dialogues with ChatGPT to fulfil my small and instant curiosities. Recently, while going through a research article on time series analysis of biosignals, a thought surfaced in my mind, and I couldn’t resist asking it –

Me: Patterns are everywhere, do you agree?

ChatGPT replied:

“… patterns are a fundamental aspect of many natural and man-made phenomena. Patterns can be found in everything from the arrangement of cells in living organisms to the movements of planets in the solar system, and from the structure of music to the behavior of financial markets….”

Interesting, right? Surprisingly, despite the fact that most of my research revolves around pattern recognition, this answer and the following sequence of thoughts stuck with me for long enough to write about it. The following article presents a deeper perspective into the presence and significance of patterns in human life. Further, to help CS students better connect with the discussion, a reflection on how modern computing systems leverage patterns is also included. So, let’s dive right in –

Why do Patterns matter?

As per Wikipedia — “A Pattern is a regularity in the world, in human-made design or in abstract ideas”. The early evolution of pattern recognition among humans was restricted to identifying favourable and unfavourable things and environments, possibly, due to limited cognitive capabilities. However, a significant step in the evolution of intelligence was developing the ability to document and utilise more complex patterns such as planetary positions, artistic patterns, seasonal patterns, behavioural patterns etc. Below are some important examples where pattern recognition has contributed to the development of early civilizations -

1. Learning to grow and cook food: Constantly hunting for food, complex digestion cycles involved in processing uncooked food and the health effects of bad food required early humans to invest a lot of energy, time and cognitive functions into meeting the needs of the human body. Many evolution theories believe that the development of the human brain can be largely contributed to our success in growing and consuming food that is readily available, safe and easy to digest.

It required developing the ability to recognize favourable food items, understand their natural habitat, develop agricultural processes & animal husbandries, recognize & predict seasonal variations, understand the impact of the environment & other living organisms on production, experiment with cooking styles & combinations; and the list can go on. Essentially, most of these activities involved pattern recognition.

2. Building social structures: Safety from threats and reproduction are crucial for the survival of any living species. Fortunately, humans were successful in building social structures that increased the probabilities of these traits manifold.

However, successful development of these social agreements involved exchanging ideas & emotions, establishing protocols & ensuring their acceptability, developing trade channels, and so on. Although today, many of us take these social structures for granted, their evolution required closely studying the repetitions and anomalies in the underlying patterns and adopting necessary measures.

Today, we utilize our ability to recognize patterns in almost every facet of our lives — to understand complex phenomena, predict their outcomes and plan actions accordingly. Most of the social, commercial and scientific breakthroughs which have led us to the modern world, involved the process of highly accurate identification, definition and utilization of various patterns. For example,

1. Discovery of the sub-atomic nucleus was made possible by keenly observing the deflection of alpha particles (in the famous Rutherford’s Gold foil experiment),
2. Implementation of large-scale socio-economic policies such as Pradhan Mantri Jan Dhan Yojana is backed by detailed analysis of data collected from several sources. Such analytics help government agencies to identify patterns (that indicate financial inclusion, income levels and access to banking services) and accordingly, plan implementation projects in various parts of the country.
3. Many large-scale scientific studies have indicated that the political stability and growth of existing democracies have been a driving factor in the increased adoption of democratic structures globally.

These examples might give the impression that understanding patterns is a complex process and should be used to solve large-scale or manually difficult problems. However, practically, most human capabilities that we can classify as intelligent come from our ability to use patterns.

How do we recognize patterns?

As humans, we are fascinated with patterns. Most things we appreciate and enjoy such as music, arts, literature etc. utilize our brains’ reactions to patterns (and surprises therein). Studies have linked the ability to recognize patterns with different parts of the human brain. However, a general acceptance is that our brain learns patterns primarily for two reasons — to avoid remembering raw data, and to predict what to expect next. These factors belong to two different aspects of learning, usually known as representation learning and probabilistic learning, respectively.

For example, representing a fruit that tastes sweet(ish), looks like an oddly shaped sphere, has a juicy texture, and a red(ish) skin by a symbolic representation (yes, I am talking about apples), takes away the burden of remembering and using all these details. On the other hand, predicting the trajectory of a sting bee and correlating it with an earlier incidence helps us save ourselves from an unpleasant day.

However, instead of focusing on the neuroscience of how our brain processes patterns, we will proceed to look into the systematic study of patterns to keep the discussion within the context of computing.

Statistically, we study patterns as a collection of measurable and identifiable features that can be observed from a phenomenon. Usually, these features are recorded as quantitative (numerical) or qualitative (categorical) variables by utilizing appropriate data collection tools and methods. For example, to understand the pattern of rains in an arid region, we can record features like wind speed, wind direction, humidity, temperature, vegetation density etc. Such samples (each containing values for all features) are collected over long periods of time (or through several iterations/experiments), and eventually used to shortlist relevant features and their impact on rain.

Some difficult examples are given below to think about how we can break patterns into features. Let me know in the comments, and I can tell if you got them right.

a. How to tell if a Wind is a Storm?

b. How to tell apart a dog from a camel?

c. How to tell if a person is healthy?

d. How to differentiate gravity from friction?

Moving on, we can say that, the essence of all science lies in identifying the right set of features that impact the target phenomena, carefully recording those features, and applying well-defined procedures to report the underlying patterns.

Over the decades, researchers have used several algorithms and methods to learn these patterns. A brief timeline is presented below:

1. 1950–1960s: Inference-based methods driven by the principles of probabilistic and statistical analysis (such as Bayesian networks) were the most used methods in this period.
2. 1970s: Researchers started exploring the usage of structural properties in underlying data to look for spatial patterns. Template matching and Graph matching methods were successful in achieving great results at that time.
3. 1980–90s: While the idea of perceptron was reported in 1969, artificial neural networks were not considered useful until the development of the Back-Propogation algorithm, Hopfield networks and Convolutional neural networks in this period. It gave rise to a sudden increase in the use of computational methods in pattern recognition tasks.
4. 1990s: Due to the limitations of computational resources and training data, researchers started exploring other methods that could be used to learn patterns without requiring a lot of data or powerful computers. Methods based on fuzzy logics, support vector machines and decision trees were prominent in literature during this period.
5. 2000s: Biologically-inspired methods, such as genetic programming and swarm optimization algorithms, helped provide a newer perspective to the pattern recognition enthusiasts in this period.
6. 2010s-Now: The emergence of high-performance Graphical Processing Units (GPUs), parallel computing libraries and big data storage technologies during the first decade of 21st Century refuelled the use of deeper neural network architectures for pattern recognition tasks. The main advantage of such systems was their ability to learn directly from the data, i.e. no need for explicit feature extraction steps.

While some of the older algorithms are still in use due to their easier implementation and computational advantages, most exceptional AI innovations like ChatGPT are driven by deep learning systems only. Now that we have discussed enough about patterns, lets move to -

How do we use patterns?

Before we discuss their applications, let us walk through the well-known categories of patterns first —

1. Geometric patterns: These are described using spatial features such as shape, size, orientation, and position. Examples include patterns in images, such as circles, textures, curves etc.
2. Statistical patterns: These patterns are defined using statistical properties of the sample data collected such as mean, variance, and correlation. For example, patterns in data sets, such as trends, clusters, and outliers.
3. Structural patterns: These patterns can be used to represent hierarchies, networks, and dependencies. Such patterns are usually relevant in text processing to establish relationships between entities like sentences, paragraphs, chapters etc.
4. Temporal patterns: These patterns are described using features that involve some sense of sequence, duration, and/or frequency. Such patterns are usually relevant in time series analysis where we try to observe trends, cycles, and seasonal components in underlying data distribution.
5. Behavioral patterns: These patterns are described using behavioral features such as actions, reactions, and interactions between people or objects. Examples include patterns in user behavior, such as browsing history, search queries, and social media activity.

Now, if we can assume that — (a) necessary tools and methods are available to collect quality samples, (b) efficient algorithms are available, and (c) sufficient computational resources are available; theoretically, we can learn to recognize any pattern from any natural or man-made phenomena.

For a long time, scientists used to rely on manual methods of data collection that are usually constrained within a controlled environment (such as laboratories). Therefore, a huge amount of literature has been developed on the efficacy, applicability and ethical aspects of various data collection methods in almost all problem domains. However, due to the ubiquitous interventions of computing and the development of new sensing technologies, it has become easier to collect the features with increased precision and velocity. In fact, the data is available in such abundance that new challenges associated with their storage, retrieval and processing have given rise to new domains of research and sub-industries.

Now, one might wonder, if all the ingredients are available, why can’t we learn anything and everything? Some major issues are —

1. Not all patterns are derived from the same set of features. Finding the right features for understanding a pattern usually requires using everything we already know about the target phenomena.
2. There can be defects in measuring/identifying the features. Recording the desired features while controlling external factors, such as noise, bias, etc. is important.
3. There could be a lot of features that seem to be related to a pattern. Due to computational limitations, we may not consider all the features and finding which ones are most relevant is difficult.
4. Not all features are completely independent of each other. We need to check if there are any relationships between the selected features and whether these relationships contribute to our pattern.
5. It is uncertain if there are enough samples to identify the pattern. How to decide if different aspects of the pattern are sufficiently covered among the recorded samples?
6. How do we test our understanding of patterns, i.e. whether we have learned it right? The same validation strategy may not work well in all situations.
7. We have to continuously update our understanding of already learnt patterns over time. However, it may not be feasible to continuously record and process new samples. What if the new samples are not collected in a controlled environment? How much should new samples contribute to our current understanding of the pattern?
8. Once we understand a pattern, we should use the existing knowledge to understand new patterns. But, does, having a common set of features necessary for this transfer of learning?

Most existing and ongoing research in Statistics and Artificial Intelligence focuses on finding the solutions to these problems. I will try to discuss some of the currently accepted solutions for these issues in another article later. Let’s close this discussion with one last question -

If patterns are so helpful, can we create them?

Let’s start with a very common use case — writing emails. Does it require pattern recognition? Yes, it does. Indeed, writing an email that someone else can understand as intended, requires developing a superset of the following abilities –

1. Identify and write the characters & symbols of a common language,
2. Understand and use syntactic rules (how sentences are formed) & semantic structures (which sentences make sense) of the language,
3. Identify and use intonations, punctuations & literary styles,
4. Comprehend and retain the context of communication,
5. Predict and adapt to the emotional and intellectual background of the receiver.
6. Use the device and software involved.

All these steps rely on man-made patterns that simplify the exchange of ideas and information among ourselves. Building such patterns has always helped us create new knowledge and pass the same to the next generations. For example,

1. All respected artists and historians are known to retain a specific style over all their works making it easier to convey their ideas across larger audiences,
2. Most successful businesses consider the adaptability and reaction of the target audience towards associated user patterns when introducing new products,
3. Most of the modern science is based on the research methodologies abstracted from the major breakthroughs in respective domains.

And we can keep listing more.

Conclusion

Bottom line is that patterns are everywhere! As humans, consciously and subconsciously, we are so good at learning and creating new ones, that we rely a lot on them. With the evolution of our data processing capabilities and the emergence of new frontiers in computational intelligence, we are now able to build machines (computers, robots etc.) that are more capable than ever of learning these patterns. As a generation, we can consider ourselves fortunate enough to be part of this new revolution.

Food for thought

1. When machines started doing the work requiring manual strength, humans started getting weaker than before (generally). When machines start helping us recognize patterns, shall we lose our capability of working with them?
2. If machines will be able to identify complex patterns, they would perform better in most tasks that require taking actions based on patterns such as transportation, quality control, coaching etc. What will be the role of humans in such a society?
3. Does all human intelligence derive from pattern recognition? If machines become significantly better in pattern recognition than us, do we have any other skill/intellect which cannot be learnt by these machines?
4. Are we enough aware of patterns being learnt by the machines right now? Are we ensuring that they are used carefully?

Let me know in the comments! And if you learned something new from this article, please share it with others.