Its capacity to support data-driven decisions and boost productivity offers exciting possibilities.
Deep learning continues to be of utmost interest to AI practitioners due to its ability to support data-driven decisions and the potential for analysis of unstructured data, resulting in significant productivity improvement and financial benefits. By processing speech, text, image, and video, deep learning has enabled us to develop more insightful perspectives.
Deep learning comprises a combination of algorithms that drill multiple layers to extract insights from raw data. It can be particularly applied to unsupervised learning tasks and to try and interpret unlabeled data, which is often much more than labeled data. Although there are exciting possibilities of deep learning, it has some facets that require a cautious approach.
Deep learning is still considered a blackbox as its recommendations are not always based on proven theories. The inferences it produces are primarily based on an empirical approach, leaving out other important factors that may impact the decision-making process. Currently, the inferences are drawn from the limited scope of training data that is made available to the system whereas the realworld data and its distribution could be vast. At times classifying unlabeled data as part of the familiar category makes deep learning architectures flawed. The biases of humans while making decisions are adopted by deep learning frameworks too, thus taking a toll on accuracy. There is also a need for alertness against digital manipulation by hackers.
So what is the future of deep learning going to be? It is important to acknowledge here that the definition of deep learning itself has been evolving as it embraces developments within the larger fold of AI. Gary Marcus, a pioneer in deep learning, has stated four possibilities for the future of deep learning, namely, unsupervised learning that would enable systems to determine their own objectives and do problem solving, integration of deep learning with symbolic systems that would make inferences more accurate, developing learning models through better insights from cognitive and developmental psychology and AI becoming multidimensional to deal with the complexities of the problem.
According to Yann LeCun, Chief AI Scientist at Facebook, the next phase of AI will not be supervised but reinforced. He has elaborated on self- supervised learning, i.e. the idea of learning to represent the world before learning to perform a specific task. As per Nikko Strom, a scientist, integrating symbolic reasoning and learning efficiently from interactions with the world are two major challenges that deep learning should be addressing. It is anticipated that deep learning systems of the future would be able to handle non-uniform data available in multiple formats. There are also exciting interdisciplinary initiatives underway that could have a significant impact on deep learning.
In summary, in order to get the most out of AI, particularly deep learning, we need multi-disciplinary research and experiments in various domains that could have far-reaching implications in developmental sectors as well as in mission-critical functions. While advanced research is underway in the western world, Indian institutions and the industry need to urgently come forward and initiate similar studies in India with a focus on addressing our unique needs.
Originally appeared in Financial Express