Data Science, Machine Learning, Deep Learning, or Artificial Intelligence?

In recent years, social media and newspapers have excitedly introduced us to the little wonders that certain new algorithms have accomplished. We hear of some algorithms as “Artificial Intelligence” (AI), while others are “Machine Learning” (ML). Meanwhile, another article would speak of “Deep Learning” and “Deep Fakes”, while others mention “Artificial Neural Networks” (ANN). In the midst of these, the term “Data Science” claims its share of the limelight. Often, people dismiss Data Science with a “Hah! Data Science is just statistics“. While it’s true that drawing a distinct line to separate the definition of these terms is a little difficult, it’s quite simple to understand them when we see the big picture.

What is Artificial Intelligence?

Put simply, AI is a very general term that describes synthetically created beings (in physical and/or software form) that are capable of human-like intelligence or higher forms of intelligence. Scientists and philosophers have made various attempts to identify and isolate the constituents of intelligence, but the human understanding of the concept is still abstract. As a result, early attempts at developing AI relied heavily on mathematical reasoning and symbolism (since those were the only tools we had for representing logic). Researchers like Rodney Brooks (the author of “Elephants don’t play chess“) soon realized that intelligent machines needed more than just logic. Interaction with the environment, common-sense reasoning, and a greater contextual understanding of the world via physical experiences are also important.

AI’s are broadly categorized into three types:

• Strong AI: This is the AI mankind wishes to eventually build. Machines that are as capable of general intelligence as us. Machines that are aware of their existence and capable of thought.
• Weak AI: The word “weak” was initially meant to indicate that the AI was a partial implementation of a “strong” AI. However, today the term is used synonymously with “narrow AI”, which represents the so-called “intelligent” algorithms that currently exist. However, no matter how intelligent a narrow AI may appear to be, it possesses no understanding of the tasks that it performs. Thus making it susceptible to failure. For example, a narrow AI trained to recognize images of ships will no longer be able to recognize ships if a single white pixel is added to the image.
• Super-intelligence: If a strong AI is made capable of performing research to discover new phenomena and also made capable of modifying its programming, it could become very powerful (intelligence explosion). The vast memory and speed that such an AI could possess, is considered a potential threat to humanity.

Methods of implementing AI

Various theories propose approaches for designing an AI. Some of these methods are:

• Logic-based: Some researchers believe that intelligence could be simulated using logic and knowledge representation methods. They created procedural representations of knowledge and reasoning.
• Anti-logic: Also called scruffy logic, this method uses empirical data or even logic if necessary, to build a working system that solves a purpose.
• Embodied intelligence: Since known biological intelligence makes use of movement and perception, some researchers believed that AI needed to be built as robots. The robot would eventually learn and acquire intelligence via the process of social interaction, exploration, and guided learning.
• Soft computing: Often, an AI doesn’t need to locate an optimal solution to a problem. ANN’s, fuzzy logic, and evolutionary algorithms provide methods of learning from data or experimental observation.
• Statistical methods: Bayesian theories, information theory, and Hidden Markov Models (HMM) were some statistical tools used for simulating analysis and decision-making processes. However, these methods were soon proven insufficient.

What is Machine Learning?

Within the vast scope of AI, machine learning is one set of techniques born from a need to help machines learn from data. The objective being, to create machines that can generalize prior learning experiences to perform well on previously un-encountered tasks. Some researchers see ML as a method of learning based on passive observations, whereas AI is viewed as actively interacting with its environment to learn.

Methods of learning

Various methods exist but these are the primary ones:

• Supervised: Algorithms are trained with human-defined input data and the corresponding desired output. As more training data is supplied, the algorithm becomes more capable of identifying patterns or categories in the data, by generalizing the past training data.
• Unsupervised: These algorithms are programmed to mathematically identify clusters, anomalies, or associations in data without needing a human to teach the algorithm using classification labels.
• Semi-supervised: Algorithms identify data better when unsupervised learning is assisted a little with the supervision of a human domain expert. The broader perspective of the human helps guide the algorithm in areas where it could have made errors.
• Reinforcement: Algorithms that learn from actions or decisions, by improving their accuracy based on a reward and avoiding actions/decisions based on a penalty.

Learning models

The “learning” of ML algorithms is achieved by various mathematical models. Some of these are:

• Artificial Neural Networks (ANN): Biological neurons transmit impulses when they receive inputs from one or more neurons. The neurons do not get triggered for every impulse, but seem to respond to certain patterns of impulses and thresholds of impulse intensity. ANN’s are designed to mimic this behavior in a very simplistic manner. The learning of neurons is represented by fractional numbers (called “weights”) whose values are adjusted based on the various kinds of input data fed to the network and often even based on the expected output. The weights eventually form a mathematical representation of the data.
• Evolutionary algorithms (EA): Although EA is considered an independent paradigm, it is also used in ML. It consists of a population of parameters which “evolve” over time by varying their values. The “evolution” and “learning” are aided by each individual of the population. The individual imbibes characteristics of other individuals that accurately model the desired solution. Thus, each individual slowly becomes more accurate, while inaccurate individuals are killed and re-spawned.
• Regression: This method involves statistically analyzing the pattern of data and identifying a suitable mathematical function that can approximately represent the data.

What is Deep Learning?

The design of artificial neural networks began with a simple layer of input nodes, a threshold function, and an output node. It was designed to classify inputs of 0 or 1. This enabled it to function as a decision-making entity like an AND gate or an OR gate. Using more layers of nodes, the network could even represent an XOR function. The “depth” of an ANN is merely the number of layers of nodes it possesses. “Fat” networks have more nodes per layer. Such deep and fat networks were observed to be capable of representing very complex functions via their neural weights. These capabilities gave rise to various kinds of specialized deep learning architectures. Some of these are:

Types of deep learning networks

• Convolutional Neural Networks (CNN): Children begin drawing stick-figures of people and keep improving their drawings by incorporating more details as they grow older. CNN’s are specially crafted networks that are designed as clusters of layers that form a pipeline. Similar to how the children improve on details, the first layer of a CNN learns very basic characteristics of the data input to it. The next layers in the pipeline learn a higher-level pattern of the relationship between the inputs from the first layer. The process of learning more complex patterns from the simpler patterns of each previous layer continues. This process is believed to be similar to the function of biological neurons in the visual cortexes.
• Recurrent Neural Networks (RNN): Often, the length of input data may vary, or data in a series may be related to the previous data. In such cases, it helps if the ANN can “remember” a prior input and incorporate that memory when processing the next output. RNN’s are capable of maintaining such a memory over some time. They achieve this by having a copy of the output signals fed back to the intermediate layers of the network as additional inputs.
• Deep Belief Network (DBN): The DBN utilizes sets of special two-layer networks called Restricted Boltzmann Machines (RBM). Inputs fed to the RBM’s first layer are converted to numbers in the second layer. Using these numbers, the second layer tries to reconstruct the inputs. This allows the RBM to determine relationships between input features and their importance. The DBN consists of a pipeline of RBMs, where the reconstructed inputs of one RBM are fed to the next RBM until the entire network is trained. After this unsupervised training, the network is put through supervised training to give names (labels) to the patterns that were learned.

What is Data Science?

The field of statistics uses probability distributions, samples of data from a population, estimation, regression, and many such methods to work with data. Data science expands on mathematical knowledge and statistical methods by incorporating much larger sets of data (big data) and AI algorithms. Data science is not a concept to be compared with AI. Rather, the storage and analysis of data is an essential part of any learning process. The easy access to massive computing power and cheap storage has transformed every field of science by making it data-driven.

Data acquisition and tools used for storage

Data is first acquired via manual methods (for example, in forms filled by people), sensors, or software algorithms. The acquired data is then stored in an unstructured format or is processed and stored in a structured format. Some popular storage tools are:

• Graph data: Neo4J, Titan, ArangoDB, JanusGraph, SAP HANA, TerminusDB, RedisGraph, and OrientDB.
• Key-value store: BerkelyDB, Couchbase, Dynamo, Redis, Aerospike, and Ignite.
• Wide-column store: Cassandra, HBase, and Scylla.
• Document-oriented (semi-structured): CouchDB, MongoDB, Clusterpoint, MarkLogic and BaseX.
• Relational data: MySQL, PostgreSQL and MariaDB.
• Data lakes: Various cloud storage services allow data to be stored in a raw format. Google Cloud, Microsoft Azure, Cloudera, Oracle, and Teradata offer such storage.

Data transformation tools

Data often needs to be converted to a certain format before analysis. The following tools help with this process:

• Grammar-based: AWK, Perl, and XSLT (eXtensible Stylesheet Language Transformation).
• Domain-specific: Spark, Dataflow, Storm, and Kafka.
• Programming languages: Python, R, Julia, Java, C++, and Matlab.

Tools for analytics

• Analytics: Tableau, Domo, Sisense, Microsoft Power BI, Google Analytics, Watson Analytics, Ahana, Starburst, and Trino.
• Visualization:js, Pandas, Infogram, Plot.ly, and Leaflet.

Machine learning / Artificial Intelligence

• ANN’s and Deep Learning: Keras, PyTorch, TensorFlow, and Caffe.
• AutoML tools: These tools automate the process of applying ML solutions to problems in the real world. Some of these tools are AutoKeras, IBM Watson, DataRobot, Amazon Lex, and Google AutoML.

The future of these technologies

Lessons from the past

Historically, there have been occasions when philosophers and researchers had been very certain that they were close to creating a human-like AI. This happened in the late 1950s (known as the “golden years” for AI) and in the 1980s. Even today, the output of architectures like GPT-3 (Generative Pre-trained Transformer) tends to inspire such hope. The reality, however, is that just as in the past, the fundamental obstacle that humans have not overcome is in obtaining a clear understanding of what “intelligence” is. Therefore, until the concept of intelligence is understood, data technologies and AI will require human intervention.

The future of Artificial Intelligence

The folly of utilizing logical and probabilistic methods of programming AI has already been realized. Many researchers strongly believe that the appropriate way for an AI to achieve human intelligence is to physically experience the world.

• Embodied Consciousness (EC): The future of AI could very well be robots that would learn representations of the real-world by experiencing the world using its sensors. Such an EC would require phenomenally greater amounts of processing power (perhaps via quantum computing) and a new method of storing vast real-world data that it acquires via event-based signals.
• Cyborgs: Intelligent technology will increasingly be used to enhance the bodily functions of people with disabilities. Brain-computer interfaces will also be used by civilians and military personnel for enhanced capabilities via sensors and/or exoskeletons.
• Robot swarms: Nano-robots could help treat diseases better with chemotherapy drugs delivered to destroy only cancer cells. Autonomous robot swarms can form an autonomous army. Robot swarms could also assemble structures (swarm 3D printing).
• Virtual Reality (VR) and Augmented Reality (AR): Technologies like Google Glass and Microsoft’s HoloLens have already shown how exciting this technology is. AR and VR will increasingly be used to make information available at our fingertips and improve communication with humans and devices.
• Software agents: It could eventually be possible to have meaningful conversations with AI agents. Although such agents would not fully understand the human world, their algorithms will be far better equipped to provide novel game-play during games. Such agents could even help cope with mental trauma or the simple difficulties of life.
• Art and entertainment: AI can provide artists with highly creative suggestions. Generative Adversarial Networks (GAN) and GPT-3 have shown that an imaginative AI may even be able to create entire movies, novels, art, or music on its own.
• Security: Image processing AI’s would become advanced enough to accurately warn security personnel or shop owners about shoplifters or suspicious behavior.
• Food: Robot chefs or AI-based cooking machines will be capable of cooking any desired recipe (or even invent new dishes). Algorithms will be capable of providing suggestions of what combinations of ingredients humans can use for cooking. AI will also help monitor and report food contamination.

The future of Data Science

Data science can assist every field of technology when used with AI/ML algorithms.

• Data processing automation: The sheer magnitude of data being generated and captured, also means that it is becoming difficult to store and process data. So architectures that understand user needs based on past experiences, would be created to sift through vast data stores. The process of loading, transforming, and extracting useful data would see high levels of automation. Intelligent algorithms would eventually be capable of performing supervised labeling of data.
• Healthcare: Data can help with better marketing, diagnosis, epidemiology, monitoring patients, and assisting aging or disabled persons.
• Agriculture and retail: Data patterns can help predict demand and supply, thus reducing wastage. Data can also predict user preferences, crop diseases, weather, soil quality, and movements of harmful insect swarms.
• Finance: Data patterns can protect from fraud and improve returns on investments.
• Education: Data can identify learning patterns, thus helping students in understanding concepts and identifying appropriate career paths.
• Ecology: Data can help identify animal migratory patterns and anomalies in Nature that can provide early warnings. Data patterns can also help identify areas where water tables or tree cover need improvement.
• Transport sector: Vehicles will use real-time data to plan paths, avoid accidents and maximize fuel efficiency.
• Space sector: Data will play a massive role in identifying habitable planets, planning missions to extract resources from celestial bodies, and discovering anomalies or properties of the universe.

Artificial Intelligence and Data Science share a great future. Although surveillance and regulations can limit the scope of applications developed, these technologies are poised to radically alter the way we live and interact with the world.