Deep Learning: The Key to General Artificial Intelligence

2019-12-27

Deep learning algorithms are at the forefront of artificial intelligence and are also a major component of autonomous driving. However, despite the advancements in the field of artificial intelligence driven by deep learning in recent years, there are still fundamental issues with the technology itself and its underlying techniques, such as deep neural networks, that prevent them from replicating some of the most basic functions of the human brain. These challenges are well-known, and an increasing number of scientists acknowledge that these issues could pose serious obstacles to the future of artificial intelligence. Artificial neural networks have proven to be highly effective in pattern detection, and in most cases, increasing the size of the neural network and training it on larger annotated datasets can improve its accuracy. This has led to a "bigger is better" mentality among some AI researchers, who seek improvement and breakthroughs by creating increasingly larger AI models and datasets. However, so far, there is still no neural network that matches the structure of the human brain, which indicates a flaw in the current approach of "bigger is better" in artificial intelligence systems. Current deep learning systems "make stupid mistakes" and are "not reliable in the face of distributional changes," which is one of the main concerns of current AI systems. Neural networks are susceptible to adversarial examples, where perturbations in the data can cause the AI system to behave in an unstable manner. This can have fatal consequences, especially in sensitive domains. Deep learning has already created many useful system 1 applications, particularly in the field of computer vision. AI algorithms now perform tasks such as image classification, object detection, and facial recognition with higher accuracy than humans. Speech recognition and speech-to-text conversion are other areas where deep learning systems have shown good performance. However, even in domains where deep learning has made substantial progress, the effectiveness of system 1 is limited. As a result, system 2 has emerged, which can do things that system 1 cannot. Although it is much slower, this transformation from nothing to something is of undeniable importance. Bengio provides guidelines on how to improve system 2 of deep learning systems, which are highly technical and reference research papers and projects from the past few years. Some recurring themes in his speech provide clues for developers on the next steps to take. Causal relationships are a major drawback of current machine learning systems, with the core being the search for and matching of patterns in the data. Bengio believes that deep learning systems capable of combining and manipulating these named objects and semantic variables will achieve artificial intelligence with causal structures.

RECENT AI NEWS

RECENT AI TOOLS