Current Flaws in Deep Learning: An Analysis
While Deep Learning algorithms have markedly improved the paradigm of Artificial Intelligence across domains like Natural Language Processing and computer vision, their performance comes with certain critical, potentially fatal flaws. This paper explains and analyses five areas of concern in neural networks and their design – the lack of necessary data, a lack of interpretability, software concerns during implementation, their biological plausibility, and the inability to encode knowledge. By citing and critiquing actual use-cases, challenges have been flagged. Finally, this paper makes a threefold recommendation- integrating traditional algorithms and explicit background knowledge into the newer methods, creating a hybrid design that amalgamates both supervised and unsupervised components and standardizing data collection across domains. The approaches suggested herein will make Deep Learning more sustainable and impactful by reducing computational resource requirements, making systems more biologically plausible and mitigating human bias.
Keywords - Deep Learning, Computer Vision, Natural Language Processing, Model Interpretability, Knowledge Encoding, Biological Plausibility, Software Development Challenges.