As a large number of enterprises look towards adopting AI, only a few even understand what they are really talking about. AI algorithms have largely remained black boxes, which has led to many unexpected outcomes. It is therefore of paramount importance to look under the hood and try to understand what AI entails.
There are various ways to look at AI–it can be viewed as an algorithm, i.e. it can be classified into supervised learning, unsupervised learning, reinforcement learning, deep learning, and such.
It can also be viewed as a means of applying data such as vision, text, timelines, and such. There are numerous other perspectives from which AI can be viewed such as a Process (Cleaning & Labeling, Feature Engineering, Training the AI’s), How AI is applied to various domains such as Healthcare, Education, Agriculture, and many others.
An important way of looking at an AI, that we will be focusing on gaining a better perspective on its workings, is AI Architecture. What are the different layers that go into making an AI stack to render it successful for use for complex ecosystems such as Smart cities, a Telecom network, refineries, etc?
“When you think about the evolution of products, we have been building mixers, washing machines, and cars, that is Product 1.0, then we started building Google Search and Netflix and Facebook and all that, that is Product 2.0, and now we are heading towards a future where we are going to build Product 3.0,” explained Dr. Shailesh Kumar, the chief data scientist at Reliance Jio.
The Asth-ang AI
To better understand the layers of the AI stack, one must understand the complexities of these ‘products’ or ‘complex ecosystems’ that it supports. Kumar elaborates on the concept as follows – These ecosystems consist of billions of entities (customers, assets, facilities), trillions of interactions (Transactions, Social network, content, etc.), millions of decisions (pricing, scheduling, logistics, etc.) that factor in the growth and existence of these ecosystems.
Such a complex body needs a well-structured AI/API that smoothens the workings of these systems. The AI stack, as explained by Dr. Kumar, consists of eight layers. Starting with Layer 0 – Digitization, which is not really a part of the AI stack but it is equally necessary as it involves the conversion of raw data (printed material) into digital form, which is an essential part of using, and even starting, an AI stack.
Then comes Layer 1 – Interpretation, which is a common and broad layer, as it involves converting the raw data into semantics for better classification. “What follows the digitization process is a data-driven AI that focuses on interpreting low-level data such as Speech data, Face data, X-Ray data into high-level semantics such as Text, Emotion, Anomaly, and such. What is important to understand is that there can be multiple interpretations for a single input,” Kumar said.
This holistic layer assists in predicting things such as emotions, gender, text, etc. giving it a very individualistic approach. Spotify is a brilliant example of this as it classifies music based on danceability, energy, emotion, and other such aspects.
Layer 2 is Causality, which studies the causes or the reason behind the observed data based on domain knowledge and information. This one is quite important as it helps in building the AI architecture for the ecosystem. While it is important to note that not all effects have the factor of causality behind them, for example – someone’s preference or their behavior may not always have a reason behind them, sometimes things just are.
There are also different types of causality such as multi-factor causality, which entails taking into account all the possible causes behind certain effects – for example, a disease may be caused by several sources such as the weather, genetics, lifestyle, etc.
Another type of causality is cumulative causality, i.e., an effect is the result of not just one but the cumulative effort by different sources. A great example of this is that the demand for a product can rise or fall due to the cumulative effect of factors such as the holiday season, marketing campaigns, competition, etc. Once the causal factors have been identified, it becomes easier to address and deal with the issue.
Layer 3 – Prediction, involves predicting what is going to happen in the future. It is quite similar to Layer 1 in its modeling techniques, but what is different is their roles. What follows is Layer 4 – Explanation, as predicting is not enough, what is needed is an explanation, especially in the fields of healthcare and such.
Layer 5 – Controllability of the stack entails dividing the input into two parts, the observable and the controllable part of the input. As stated by the speaker, this distinction is quite important as it allows the enterprise to ascertain the controllable variables that help them maximize their numbers.
Layer 6 – Simulation involves playing with the controllable variables to get the desired results based on the predictions. Layer 7 – Optimization involves improving the simulated results to the benefit of the enterprise.
Layer 8 – Adaptability involves adapting to the uncontrollable variables and optimizing as such. The thing to remember is that each layer is just as important as the next and for the AI stack to function to its full potential, all the layers need to be followed in a proper manner.
AI’s inflexion point
“Artificial Intelligence has seen a huge inflection point over the past decade from being used for research & development to being utilized for consumer technology,” feels Rafee Tarafadar, CTO at Infosys.
With the mounting utilization of the technology on the consumer end, there is great pressure on enterprises to scale their AI technology for better & optimized results.
“This has resulted in a second inflection point where the tech is crossing borders from consumer tech to enterprise tech. Enterprises investing in cloud acceleration, digitization, and data transformation laid the foundation for scaling the AI and was a driving force for this transition for the AI tech”, Tarafdar added.
There are a few things that need to be fixed, as explained by Tarafdar, when it comes to the pilot AI projects taken up by the G2000 companies, which are democratizing the tech i.e. making the tools and solutions available throughout the organization, which requires special skills and tools like GPUs and TPUs. The third thing is that the outputs delivered by the AI models need to be interpretable.
“AI is writing software that achieves results that no human-written software can write,” Vishal Dhupar, Managing Director, South Asia at NVIDIA said while emphasising on the importance of AI for modern enterprises.
Computation needs for state-of-the-art AI models have gone up to over 30,000 times in the past 5 years, which means that the computing requirements need to be doubled every 2 months.
For example, the Human-Level Reading Comprehension of the NVIDIA AI Megatron BERT is 24x better than that of an average human. Half of Facebook’s users prefer their AI Chatbot, BlenderBot, for customer service. Caltech researchers developed a unique system, Neural Lander for drone deliveries that resulted in 6x safer landings.
With such a wide range of utilizations, right from personalized consumer experience to being the backbone for complex ecosystems such as a telecom network or a smart city, the potential of Artificial Intelligence is unparalleled and finally recognized.
The elaborate structure of the tech allows for an amazing user experience that covers all angles and caters to different industries effortlessly. While there is still a long way to go to achieve Artificial General Intelligence in its true capacity, the growth observed and achieved by data scientists and engineers is exponential.
