Deep learning has been used interchangeably with machine learning (ML), and ML and AI are put on par. As the data science is being practised today, we come across these terms every now and then. These terms have certain connotations.

Deep learning is a subset of ML and ML is a subset of AI. In fact, these three could be seen as three concentric, overlapping circles where AI is the outermost biggest circle, followed by ML and deep learning.

AI has taken over those tasks where human intelligence was deployed. ML and deep learning are both parts of AI.

ML is adaptive AI and is capable of working without human intervention or with least human intervention. Deep learning is a subset of ML, which uses artificial neural networks to mimic the learning process of the human brain.

Neural networks are computing systems with interconnected nodes whose working resembles the nervous system of neurons in the brain. These systems use algorithms which recognize hidden patterns, and correlations in raw data, cluster the data and classify it. They continue doing so over a period of time to keep learning and improving.

Neural networks use interconnected nodes or neurons in a layered structure resembling a human brain. It is called deep learning.

There is considerable research in neural networks fashioned after the neural network in the human nervous system.

Dr. Robert Hecht-Nielsen invented the first neural computer. According to him, a neural network consists of a highly inter-connected processing elements. These processors process the information in response to external inputs dynamically.

Human nervous system consists of billions of nerve cells or neurons, which are connected to other cells by axons. The stimuli are received by dendrites causing electric impulses which travel through the neural network. One neuron thus communicates with the other neuron to tackle the issue or stops the message without further forwarding it.

Artificial Neural Networks shortened as ANNs have multiple nodes, which interact with each other. These nodes receive data and perform simple operations on them. The output is passed on to their nodes. It is called activation or node value.

Each link is assigned a weight. ANNs have the capability of learning by altering weight values. Feed Forward ANN makes the information flow unidirectionally. It does not receive the information back through feedback loops. These are used in pattern generation, recognition, classification. They have fixed inputs and outputs. In FeedBack ANNs, there are feedback loops to address content addressable memories.

If the networks generates desirable output or good output, the weights of the nodes are kept as they are. If the output is not desired or poor output or with an error output, the system alters the weights in order to improve the results.

ANNs can be trained and are capable of learning. It could be supervised learning where the responses are provided by a teacher. ANNs make just guesswork and compares its own answers with those of the supervisor, and makes the adjustments.

In unsupervised learning, there is no example data set with known answers. Here a hidden pattern is recognised. There is clustering involved where a set of elements are divided into groups in accordance with some unknown pattern.

In reinforcement learning ANNs make the decision in response to the environment. In case, the observation is negative, the network adjusts its weight to make another decision next time.

Back Propagation Algorithm

This is the name given to the training/learning algorithm which learns by example. Here the algorithm is fed the example from which you expect it to learn. It changes the weights of the network to produce the desired output for a specific input after finishing the training.

Distinction between ML and Deep Learning

ML requires smaller data whereas deep learning requires big data. ML requires more human intervention to adjust and to learn. In deep learning, the computer learns from its own environment and past experience. ML requires short training, whereas deep learning requires longer training. ML is linear, whereas deep learning is non-linear and complex. ML training is on a CPU whereas deep learning requires GUI for training.

ML algorithms require human correction. Deep learning algorithms can improve their outcomes through repetition, with no human intervention.

Deep learning requires huge and at times unstructured data. It is an evolution of ML, and is a process that layers algorithms and computing units (or neurons) into an artificial neural network.

While ML uses simpler concepts such as predictive models, deep learning uses artificial neural networks.

Bayesian Networks (BN)

As in decision trees, these networks represent the probabilistic relationship of a set of random variables. These are also called Bayes Nets or Belief Networks.

In such networks, a node represents a random variable, e.g. the node cancer represents the proposition that the patient possibly suffers from cancer.

The edges connect the nodes. These edges represent the probabilistic dependence among the random variables. The strength of the relationship between variables is quantified by probability associated with each node.

The constraint in BN is that you cannot trace back to a node.

BNs are capable of handling multivalued variables simultaneously.

A knowledge engineer builds a Bayesian network. He defines a problem and identifies interesting variables. There are three values that these nodes can take a time, binary values, ordered values or integral values. Then arcs are created between the nodes. The conditional probabilities are assigned to each node to quantity the relationships among nodes.

Application of Neural Networks

They are used in aerospace, automobile guidance systems, electronics, finance, production, medicine (cancer cell analysis, EEG, ECG analysis, prosthetic design, transplant time optimizer), speech, telecom, transport, software, time series prediction, signal processing and control, anomaly detection.

The first neural network was conceived by Warren McCulloch and Walter Pitts in 1943. They wrote a paper on how neurons may work. They created a simple neural network using electric circuits.

In 1975, Fukushima developed the first neural network which was multilayered. These networks have been used to perform diverse tasks.

Later, deep learning systems where developed when Big Data appeared which was both structured and unstructured.

Weights are numeric values that are multiplied by inputs. In backpropagation, they are modified to reduce the loss.

There is self-adjustment depending on the difference between predicted outputs vs. training inputs.

Activation function is a mathematical formula that helps the neuron to switch on or off.

There is an input layer , there is a hidden layer of intermediary modes which take the weighted input and produce an output through activation, and output layer.

Convolutional neural networks are used in image processing, computer vision, speech recognition and machine translations.

In facial recognition , the brain quickly first resolves whether it is a male or female face or whether it is a black or white face. It is a matter of perception. The perception could be multi-layer perception.

Such networks also consider Long-Short-Term Memory (LSTM). There are sequence models. There are modular models.

Types of Deep Learning Neural Networks

We come across three major categories of neural networks — convolutional neural networks (CNNs), recurrent neural networks (RNNs) and generative adversarial neural networks (GANs).

CNNs are convoluted, fully connected layers used to process images as these networks can extract essential features from images with less computation cost and time. They are used to classify images and detection of objects.

RNNs are feedback connections to learn patterns, and are used where the context of the previous result could be extended to predicting the next results. A common illustration is natural language processing or NLP.

There is long-and-short-term memory (LSTM) here. RNNs are used to recognise voice and in the analysis of time series.

GANs are these instances where sufficient training data is not available. GANs are used to generate similar data, similar to the input. There are two aspects here — the generator and discriminator. The generator provides data similar to the original/data pattern, while the discriminator distinguishes between the original and the duplicate data generated. Both these are trained parallel.

There are two types of ad agencies operating in the market. There are network agencies, which are part of the global networks and there are independent agencies.

Advertising agencies will have to share business accountability, since producing content is today a democratic activity — anyone can generate content at home. It diminishes the value of ads per se. Organisations and marketers are in search of agencies which bring to the table something beyond advertising. The agencies can cater to a few clients to rise to the occasion, and develop high-value relationships with the clients. The nature of the relationship expands — it is not restricted to advertising and marketing, but goes beyond and becomes a business relationship.

Independent agencies have ownership which has personal skin in the game. The relationships matter them. In addition, there is a diversion of talent from network agencies to independent agencies, since these provide a lot of autonomy and foster a sense of independence. The manpower has the freedom to do what they want to do in independent agencies.

Clients treat agencies as vendors. There is a pitch for every project. Multiple agencies vie with each other for a pitch. Clients choose a script they like and produce that. It is a transactional way of doing business. This model is highly aggressive. There is frequent pitching for one project. This model may not be suitable for all the agencies.

In this age of digital advertising, digital contributes more to business than creativity. Clients survive through e-commerce, performance marketing or SEO, Creativity has limitations. In digital, there is a dive into mechanics, into numbers and understanding the consumer behaviour on these platforms. Digital agencies arrive at a more solution-based approach.

There should be a swing back to quality of content. Today, the content is churned out just to remain visible. There could be a shift in content where quality starts to matter again. Digital is anti-craft and anti-quality. We can witness a comeback of craft and quality.

Agencies will not be just producers of advertising, but will elevate to the role of business partners. They will have to play a larger role.

All across the world, the network agencies are struggling, and so there is a larger role for independent agencies.

An art work is a laborious job. However, the advent of AI has made artworks available on internet with ease. It affects the career of the artists. An AI-assisted bot generates art work as a response to a prompt.

Researchers want to protect the original art works. They developed a tool called Glaze to thwart AI from mimicking the style of the artists. Glaze Beta 2 is available for download.

An individual art work or artwork portfolio is posted online to earn revenue. Generative AI can create art works in the same style. Glaze creates a cloaked version of the original image as a protective measure. After such cloaked version is produced, it can be posted online instead of the original work. It prevents AI to mimic it.

The work of the artist is first uploaded to Glaze. It makes a few alterations hardly perceptible for the human eye. These alterations are called ‘style cloak.’ The altered art work is called ‘cloaked art.’ It almost appears identical with original. The computer picks up the cloaked version. Now if the prompt asks the AI-bot to depict Husain’s horses, the generated images will be very different from the original work of Husain.

Glaze has its limitations. Some alterations made on flat colours and smooth backgrounds are visible. Glaze is is also not a permanent solution against AI simulation. AI too evolves, and Glaze is not future-proof. Thus art becomes vulnerable.

Glaze is useful as it allows artists to earn rather than the AI companies who charge a subscription fee. It is a beginning made to protect the artists from AI simulation.

AI has now advanced to a stage where a prompt can make it generate videos.

Generally, people consider the Big Tech promoters as the tech heroes. However, the reality is different. The World Wide Web was considered by the European Council for Nuclear Research (CERN) in Geneva. CERN is known for its research in physics. Tim Bernes Lee, a graduate in physics from Britain was employed in CERN. He conceived the idea of a system where the universities share their knowledge through a computer-based knowledge sharing system. It should be available free of cost and should be open to all. Thus the Web was born in Switzerland in 1989 and was fathered by a Britisher.

The US Defence establishment too promoted innovation. It formatted the rules for sending data from one computer to another. It forms the foundation on which the Internet is built. These rules are called TCP/IP and were invented by people employed by the U.S. Defence Advanced Projects Agency (DAPA) in the 1970s. Internet is thus the invention of the USA.

Wikipidia helps its 55 million users to get information about persons and things we do not know. It is a no profit organisation, founded by Jimmy Wales, an Albamion born in the US. However, he is a British citizen.

AI is the most pathbreaking innovation of the last 500 years. Geoffrey Hinton, Wimbledon-born, Cambridge educated and with a PhD from Edinburgh thought of the key idea of ‘artificial neural networks’. It is a mathematical technique that mimics the processing of the information in CNS. Dr. Hinton later moved to university of Toronto. He met there Youshua Bengio and Yann LeCun, both of French origin. They collaborated and implemented these ideas in practice in 2017. It gave a boost to AI. What other countries thought to be a foolish experiment was funded by the Canadian authorities.

These three innovation of world wide web, Wikipedia and AI have been made by people not from the Big Tech industry.

India has a talent pool of about 4 lac professionals knowing AI. It constitutes 16 per cent of the global AI pool. However, Indian manpower falls short in fulfilling the job requirements of the employers.

India is investing in AI and the growth in CAGR is 31 per cent. India’s share in global AI investment is 1.5 per cent. Bangalore happens to be an AI talent hub, ranked 5th in HBR review study. New Delhi, Hyderabad and Mumbai lag far behind.

AI has several dimensions — vision AI using sensors, cameras, neural networks, ML algorithms, facial recognition, conversational AI for voice and text-based conversations, sense AI culling information from emotions, expressions, uses in wearables, and decision AI for organisations.

The talent pool in demand must have skills in DB Query languages, deep learning, NLP, reinforcement learning, predictive analytics, scripting, robotics, neural networks and cloud computing.

Generative AI models respond to natural language questions. It interprets the questions of the users and the instructions too. The same are converted into machine language to produce the output. In GPT 3, the questions are converted into SQL or structured query language. The same is run on a database to fetch the required answer. It is necessary to learn the way the prompts are written. A prompt could be a general statement — compose a poem on Paris. A prompt engineer frames the question asking the model to write about the museums of Paris by making it specific or asking it to write about Paris as the fashion capital. Such prompts will produce the best output. The queries could be framed differently for different language models. GPT 3 and ChatGPT are preferred because these models keep client’s sensitive data concealed while generating a code. It is called tokenisation.

Air India’s CEO has announced that the company wants to use GPT 4 to enhance the customer experience. As it is the Frequently Asked Questions (FAQs) do not cover all the queries of the customers. Here GPT 4 will take over, and replace the FAQs. The airline intends to remain ahead of the curve by implementing generative AI.

Air India can go forward and can allow the entire graphic user interface (GUI) to be taken over by GPT 4.

In addition, there are pilot briefings before they undertake a long range journey, say from Mumbai to San Francisco which takes 14-15 hours. Such briefings to pilots run into 150 pages. A pilot is expected to understand it a few hours before a flight. Artificial Intelligence (AI) can be used to summarise the salient features of the pilot briefing.

It is the new age technology on the lines of AI and ML. It is not restricted to crypto currency only. There re other applications of blockchain which improve transparency and security. It basically is a decentralised database cryptographically secured. The database is called a ledger. It enables parties to execute transactions and confirm them with no outside intervention of a central clearing agency. It is thus an efficient system for quick transactions.

There is a computer network on which the blockchain relies. Such computers are called nodes. These nodes could be controlled by one party or multiple parties. In case there is multiple party control, there comes a decentralised network. The data is more reliable as it is not controlled by a single individual or organisation.

If new data has to enter this the blockchain, all nodes must agree to it. Thus data cannot be tampered. It is accurate and reliable. All data in the blockchain is digitally signed — who has given entry to the data or the current owner of the asset.

Blockchain networks can be public or private. In private blockchain, there is access to only verified partners.

Blockchains are used for cross-border transactions. They can be used along with other technologies such as AI or ML.

Blockchains have led to new things such as Web 3.0, DeFi and NFTs.

In simplest possible terms, generative AI uses algorithms to process data and generate new output. In other words, it generates new data that is similar to the original training data. It gets this capacity through a process of learning patterns in the input data and uses that learning to generate new data that syncs with those patterns.

Generative AI thus uses pre-trained, large language model that provides the users output in the form of text, images and other content. This is done in response to text-based prompt.

Such chatbots could be of immense help in customer service. These can answers queries. They can summarise and condense policies. They can prepare promotional material and product manuals.

Salesforce, a California based company, has announced Einstein GPT to create personalised content, emails and targeted messages.

Generative AI can auto-generate code for the programmers.

AI is predicted to generate 10 per cent of all data and 20 per cent of test data.

Generative AI can be used in banking, finance and insurance. It can be used in new drug development and fashion design. It can assist conduct of meetings — prepare summaries, transcriptions and content. It can be used in conjunction with syntax algorithms.

On integration with organisation’s systems, generative AI can take commands such as ‘prepare the report’, ‘refine this offer’ or ‘create an application.’

The data fed as input to this model must be decent and unbiased. It needs protection.

Generative AI facilitates automation, augmentation and acceleration.

Of course, in precision tasks where being error-free is important, the model is not suited. There are legal issues too while it uses copyright material.

AI-based bots Generative Pre-trained Transformers use artificial neural networks and are capable of simulating text and speech of human beings, responding to queries, generating content and translating from one language to another-language.

GPT-1 was discussed as a natural language understanding model in a paper by OpenAI in 2018 and was just ‘a proof of concept’ never released for the masses. GPT-2 was released in 2018 by OpenAI to ML-community, and had limited application in text generation. In 2020, OpenAI released GPT-3, with 100 times more parameters than GPT-2 and used a vast data base for its training. There were additional changes that led to GPT-3.5 which now powers ChatGPT. GPT-4 is the latest. It is more creative and collaborative. It can tackle difficult problems with greater accuracy. It can edit and iterate with users. It responds not only to text but also to images. Though better than previous versions, it is not perfect. Its results of factual accuracy are better. It could be steered well, and does not cross the guard-rails. GPT-4 is capable of generating captions, classifications and analysis for images. It can handle thousands of words of text from the user. Thus it better understands the context. It can ferret out documents and analyse them.

It has limitations of biases and responds to adversarial prompts. There is still more work to be done.

It is a big leap from GPT-3.5 as it is iterative in conversation. In complex tasks, the difference between GPT-3.5 and GPT-4 can be noticed.

Microsoft Azure trained the model. The new Bing-AI already uses GPT-4. It will soon be available for paying users. It could be integrated to participating organisations’ products.