Deep learning, as it is primarily used, is essentially a statistical technique for classifying patterns, based on sample data, using neural networks with multiple layers. (Deep Learning: A Critical Appraisal)
There is no doubt deep learning benefits our lives a lot. However, through this week’s reading, we know that deep learning and AI has lots of limitations. Take health care sector as an example.
According to Healthcare Tech Outlook, AI applications making waves in healthcare today, including disease detection at early stages, drug creation etc. The potential of artificial intelligence for making healthcare better seems to be indisputable, but the medical limitations of present-day AI has to be acknowledged.
First, there is the risk of feeding the computer with underlying bias. The outcomes of deep learning only depend on previous income data, so the forecasting and predictive abilities of smart algorithms are anchored in previous case. Therefore, deep learning might be useless in novel cases of drug side-effects or treatment resistance.
Second, this can be especially problematic since machine learning apps usually run as a “black box” where the machinations of its decision-making aren’t open to inspection. If a clinician can only judge a prediction based on a system’s final outcome, it may either undermine the human opinion or simply prove worthless.
Third, in problems where data are limited, deep learning often is not an ideal solution. (Deep Learning: A Critical Appraisal) Human beings can learn abstract relationships in a few trials, but deep learning thousands, millions or even billions of explicit training examples. Nowadays, patients’ digital data is not enough to implement deep learning.
AI works best in closed-end classification problems given that there is enough data available and the test set closely resembles the training set. However, our world is always changing and unstable.
Therefore, AI is not as a universal solvent, but simply as one tool among many. Not only in health care, when we implement AI to run a conclusion, we need to be very careful about the results.
Zachary C. Lipton and Jacob Steinhardt, “Troubling Trends in Machine Learning Scholarship,” ArXiv:1807.03341 [Cs, Stat], July 9, 2018.
Gary Marcus, “Deep Learning: A Critical Appraisal,” ArXiv.Org, January 2, 2018.