Sudoor Manthan- IIRS Science Portal
Evolution of Artificial Intelligence was driven by the need of automating difficult and tiring tasks. The initial developments in the AI domain were around building rule based systems, which later turned into the urge for developing machines that could think. Figure 1 summarizes the evolution of AI over time: starting from mathematical models (top left), and moving clockwise through perceptrons, multilayer perceptrons (MLPs), billion-parameter models, the rise of human-computer interaction, robotic arms, decision and analytical systems, and finally, chatbots.
Figure 1: Journey of Artificial Intelligence – from Mathematics to Intelligence
The advent of AI, which started with the development of symbolic and rule based systems in the early 1950s (Turing, 1950), soon hit a road block due to their reliance on static rules. This led to the first AI-winter because the systems could not generalize and adapt. In 1958 perceptrons were introduced for supervised learning, where biologically inspired neurons adapted their internal weights based on positive and negative examples. Although the mathematical modelling of biological neurons was a breakthrough, the advancements were stalled until the advent of backpropagation algorithms (in the 1980s) for training the network of neurons. By the 1990s and early 2000s, researchers identified machine learning as a critical component that can give decisive power to the applications of artificial intelligence. These years were the foundation years of deep learning, during which the researchers identified the capability of deeper networks in learning features effectively. However, the vanishing gradient problem limited the training capability, until Hinton came up with pretraining and other approaches, in 2006 (Hinton et al., 2006), to resolve the same. Almost the same time, recurrent networks and Long Short Term Memory (LSTM) came up as a solution to the sequence modelling, while LeNet showed that convolutional layers could learn the visual features directly from the pixels. By 2016, with the advancements in computing power, deeper networks were made practically viable, resulting in breakthroughs in various domains. A timeline of AI evaluation is summarized in Figure 2.
Figure 2: Understanding of Trends in Artificial Intelligence, from early systems to current trends
As AI systems got better, they left the research bench and jumped into the real world. Imaging the robotic arms you see in modern factories — part of the Industry 4.0 movement — quietly spotting even the tiny flaws, saving power, and fixing machines before anyone notices a problem. All of this happens because sensors pour huge streams of data into small on-site computers that can think on their own. The same breakthroughs are now woven into our daily routines. In the early 2000s, the widespread availability and rapid expansion of satellite and geospatial data created fertile ground for new technological applications. The surge caused the emergence of GeoAI at the intersections of advanced AI with remote-sensing and GIS workflows. Faced with that deluge, researchers convened the first GeoAI workshop at ACM SIGSPATIAL 2019, cementing the term and a community devoted to deep-learning-for-place, working towards automation of ways to extract geospatial information about places, features, objects on earth. Geoinformatics Department of Indian Institute Remote Sensing conducts research in the domain of machine learning and AI for earth observation and geospatial intelligence for automating extraction of geospatial information on various geographic features such as building, trees, clouds, ships, to name a few, and for automated modelling and forecasting various geographic phenomena and processes. A recent study conducted at Indian Institute of Remote Sensing compared classical object-based methods with modern deep networks for building detection. This study underlined the significance of end-to-end learning pipelines for automating information extraction from satellite images (Srivastava et al., 2024). Such automating pipelines are slowly getting woven into our daily routines. Technological advances have turned GeoAI into the engine behind precision agriculture, urban growth planning, infrastructure health checks, and real-time disaster response — proof that whenever spatial data volumes spike, the GeoAI toolbox expands to keep pace.
Today AI is no longer confined to research labs. Various cloud based APIs allow small business owners to deploy their chatbots in minutes; smartphone users keep the soul alive for AI research, as users translate speech or diagnose plant disease offline; artist co-create illusions with diffusion-based tools; citizen scientist label food imagery that a GeoAI model turn into actionable maps within hours. Such progress marks genuine democratization of intelligence, yet it also imposes duties of transparency, fairness, and sustainability, giving rise to Explainable AI (XAI). From Turing’s proofs to generative images, the arc of AI has always bent towards wider capabilities; the task now is to ensure the benefits bend just as decisively towards everyone.
References
A.M Turning, I.—Computing Machinery and Intelligence, Mind, Volume LIX, Issue 236, October 1950, Pages 433–460, https://doi.org/10.1093/mind/LIX.236.433
Hinton, Geoffrey E., et al. “A fast learning algorithm for deep belief nets.” Neural computation 18, no. 7 (2006): 1527-1554.
Srivastava, Vandita et al. (2024). Investigations on extraction of buildings from RS imagery using deep learning models. International Journal of Remote Sensing, 45(1), 68–100. https://doi.org/10.1080/01431161.2023.2292016