Deep Learning has too many choices.

To start with there are so many methodologies starting with the original Perceptron schema proposed by Frank Rosenblatt in 1958 to the many others shown below. Deep learning is a crude model of the way the human mind is able to decode information from the senses, primarily sight and sound, using a collective set of neural impulses to identify an item as seen or heard. Many proposals were made starting in the 1940’s for so called machine learning. Only within the last ten years has compute power has been able to process the simplest images that a human can decode in well under a second.

Python and the “R” language have libraries to “do” machine learning. “scikit-learn”, Tensor-flow sponsored by Google and PyTorch popular with Facebook are the most popular ways to do Deep Learning.

My problem is there are too many patterns and too many libraries. Google has invested heavily in Tensor-flow 2.0 and offered up the open source so maybe that is the best choice. My dilemma is whatever language/library or methodology I use, it’s a heavy investment of my time to learn the library and apply it.

One other thing I should mention is kaggle.com. This is a meeting place for deep learners. Kaggle has a repository for your AI code and over a million other contributors. You can copy code from others and learn by example. You can access articles about Deep Learning. Or you can earn cash prizes for sponsored AI completions.

Plenty of resources are available to learn and practice Deep Learning. Some knowledge about Differential Calculus and Linear Algebra is helpful. Python has a numpy library of routines to handle mathematical constructs which is used heavily to support Tensors. Tensors are the basic unit of Deep Learning. Tensors are multi-dimensional matrices that can get pretty dense. Often times while learning, Deep Learning, you are lead down dead ends or given solutions that are already obsolete.

My personal interest is how deep learning is used in medicine, especially radiology. Vast amounts of data is collected during scans of our body tissue. A radiologist is expected to find that one subtle marker in a tissue sample that may be a cancer growth, for example. They have to review lots of patients and samples reliably without making mistakes. Machines are very thorough if trained right and don’t get tired. Having had cancer, this is very important to me.

In summary, what should I invest my time in. Maybe the answer is the problem I want to solve dictates what tool I should be using.

Computer engineer in Silicon Valley. Writing about topics that interest me and may interest you.