Artificial Neural Networks (ANNs) can solve an extremely wide variety of problems, including virtually any problem reduceable to functions. Often ANNs are used to obtain forecasts, validate data, detect anomolies, research customer purchases, and manage risk. Also, ANNs can accomplish more complex tasks. Classifying images, making product recommendations, natural language processing, and even recognizing diseases from various scans are possible with ANNs and enough labeled data. However, ANNs do have some significant drawbacks.
ANN Hardware Dependence
ANNs usually require the use of expensive graphics processing units (GPUs) that allow parallel processing. They also need a lot of RAM. The use of cloud computing can solve this issue in many cases. Another issue is the expense of using cloud computing resources.
Personal computers may not have the hardware to tackle ANNs quickly or serve the models to make predictions.
ANN Sharing Difficult
Once you have trained an ANN, sharing it is difficult.
ANN Overfitting Likely
ANNs are prone to overfitting on the training data. One of the primary reasons that overfitting is likely is because an ANNs size and structure are largely chosen through trial and error.
Size and Structure Chosen by Trial and Error
The size and structure of ANNs are often selected based upon experience as well as trial and error. ANN size and structure choices are critical in coming to an accurate solution.
ANN Solution Not Guaranteed
ANNs do not guarantee convergence on a prediction or solution.
With the right choice of parameters (known as hyperparameters in ANNs), an ANN can approximate a target function until it reaches a satisfactory result. However, such a solution is not always obtainable.
ANN Training Takes a Long Time
ANNs take a relatively long time to train. In the world of computer problem solving, ANNs are somewhat slow. The greater a GPU’s power and the higher the ram, the less time it takes to reach a solution.
On the other hand, the larger the dataset and the more parameters, the longer the ANN tends to take to converge on a solution.
Often, if using cloud-based computing, the longer the training takes, the more it costs because users rent cloud computing resources by the hour.
Black Box Solution
ANNs by themselves are black boxes solution finders. Meaning that once a solution is found, it is difficult to figure out how that solution was attained.
According to the authors of the book “Deep Learning for Coders with Fastai and PyTorch”, layer by layer and even node by node weight comparisons can give mathematical insight. When combined with visualizations, ANNs may be made more comprehensible. “There is a vast body of research showing how to deeply inspect deep learning models and get rich insights from them, ” Jeremy Howard and Sylvan Gugger insisted.
I will discuss the inspection of ANNs in future articles.
Howard, J. , Gugger, S. (2020). Deep Learning for Coders with fastai & PyTorch. O’Reilly Media.
Mijwel, M. (2018, Jan. 27). Artificial Neural Networks Advantages and Disadvantages [Blog post]. Reviewed by Springer Nature. Retrieved from https://www.linkedin.com/pulse/artificial-neural-networks-advantages-disadvantages-maad-m-mijwel/
Shah, J. (2017, Nov. 16). Neural Networks for Beginners: Popular Types and Applications [Blog Post]. Retrieved from https://blog.statsbot.co/neural-networks-for-beginners-d99f2235efca