This dissertation addresses the opportunities and three main challenges of Artificial Intelligence capabilities and promotes the development of high-performance, low-cost AI-inspired architectures.
The author of this dissertation focuses on three main questions regarding Artificial Intelligence (AI): if the enhancement and augmentation of low-quality data can improve the performance of AI systems; if it is possible to reduce the number of parameters in a neural network while increasing its accuracy and robustness; and if the integration of visual perception and cognition of the expert into current architectures demystify the “black box” element of AI. The author discusses the benefits of AI, that AI systems are capable of learning from their experiences, adapting to new inputs, and can perform tasks similar to those performed by humans; and argues that with improvements in methodology and the availability of large databases, AI systems are capable of achieving exceptional performance on a variety of complex tasks. The author also identifies the three main challenges of AI, including: the high complexity and energy demands of current AI models that make it difficult to deploy them in resource-constrained environments; the poor quality of data impacting the performance of AI system and leading to inaccurate predictions; and the lack of understanding of the underlying process that reduces the trust in and the verifiability of the decisions made by the AI system. The author suggests that this dissertation will contribute to the development of efficient models that will accelerate the adoption of AI in medical and forensic applications.