AI’s Deep Problem

Source: Public Domain Archive

Artificial intelligence is modeled to some extent on the human brain; and there’s a deep problem with this approach. Machine learning is a subset of artificial intelligence (AI) where computer programs automatically learn from data without explicit programming. Inspired in part by the human biology, deep learning is a machine learning method that deploys layers of artificial neurons, called nodes, in an artificial brain called a neural network. Neuroscientists and psychologists have yet to fully understand how the human brain works. Similarly, there’s a big problem with deep learning; scientists do not really know exactly how deep learning reaches its decisions. In both cases, complexity is at the root of the lack of transparency.

The human brain is complex; researchers estimate an adult male human brain to have 86 billion neurons on average [1]. Human neuroanatomy textbooks commonly gauge the number to be closer to 100 billion neurons. Similar to the human brain, deep learning consists of densely interconnected processing neurons, or nodes, arranged in multiple layers. Deep learning does not require explicit programming because it is designed to learn from vast amounts of input data. For example, Google’s deep learning program learned to recognize images of cats after being fed 10 million YouTube video thumbnails without hard coding or labeling the images [2].

To understand why deep learning is extraordinarily complex requires an understanding of the functional process itself. Neural networks find patterns in big data sets, and then develop the ability to conceptualize and generalize. Vast quantities of data are fed into the artificial neural nets. The first layer of nodes processes the data and then moves to the subsequent layers of nodes until the last layer is reached, and a single decision is made. In the processing, weights are mathematically computed for the nodes, and for the strength of the connection between nodes, like the brain synapses. The neural net creates a model with upward of billions, if not trillions, of parameters based on intricate connections between the nodes. It is this inherent complexity in the model that makes it impossible to determine exactly how deep learning produces its output.

The opacity of deep learning becomes problematic in several areas of ethics, legal, and quality control. For example, the auto industry is rapidly moving toward autonomous vehicles using deep learning technology. In the event of an accident, there is no definitive way to understand the reasoning behind the decisions made by neural networks in an autonomous vehicle. Who is at fault in such cases? The question poses an ethical and legal dilemma for all stakeholders, including the injured, passengers, insurance companies, and auto manufacturers. How does a consumer assess the quality of an autonomous vehicle without understanding the decision-making process of the driver?

Another example is the deployment of deep learning for image analysis in health care for certain cancers and diabetic retinopathy [6]. Would you trust the disease diagnosis of a deep learning model without knowing why it was made? A human doctor can explain his or her reasoning and logic when questioned by the patient. This is not the case with deep learning.

The extent of AI’s transparency problem is growing, and will only become more of an issue in the future as automation increases. The recent surge in commercial and research breakthroughs in AI have been largely due to increased computing power via graphics processing unit (GPU) accelerators to achieve massive parallel processing, versus a central processing unit (CPU) which processes information serially and sequentially [5]. Also contributing the rise of AI are decentralized cloud-based computing, and the availability of big data sets. Machine learning is used for speech recognition, autonomous vehicles, image processing, handwriting recognition, and more. The level of AI power and sophistication was demonstrated when Google DeepMind’s AlphaGo program, a deep learning model, defeated the world’s best human Go players [3]. Deep learning algorithms are part of the speech-recognition technology by Apple, Microsoft, Amazon and Google [4]. AI is being deployed throughout multiple industries globally, thus underscoring the importance of addressing its opacity.

Scientists and researchers are currently working on demystifying what’s often referred to as AI’s black box; no one knows exactly how deep learning arrives at its decisions. The irony is that artificial intelligence is modeled after the brain, and by doing so, inherits the unknowable complexity of human cognition.

Copyright © 2018 Cami Rosso All rights reserved.