Source: Deep Learning on Medium
Eight Deep Learning Pros and Cons
The Advantages and Disadvantages of Deep Learning Implementation Faced by Today’s Businesses and Organizations.
Deep learning has two key strengths that set it apart from other machine learning techniques. The first of these is feature learning. With other techniques, data scientists manually transform features to get the best results with a particular algorithm. This procedure takes time, and also requires a good deal of guesswork.
In contrast, deep learning learns higher-level abstractions from input data at many levels. The data scientist does not guess how to combine, recode, or summarize the inputs. Also, deep learning detects interactions among variables that may be invisible on the surface. It can detect nonlinear interactions and approximate any arbitrary function.
While it is possible to fit interaction effects using simpler methods, those methods require manual specification and more guesswork from the data scientist.
Deep learning automatically learns these relationships. Feature learning and the ability to detect complex relationships tend to make deep learning a good choice for certain kinds of data:
For problems like speech recognition and image recognition, the learner must distinguish between a huge number of discrete categories. (For example, a speech recognition application must distinguish among almost 200,000 words in English alone.)
Mathematicians call this property cardinality. Conventional machine learning techniques often fail at this task; deep learning can solve classification problems with hundreds of thousands of elements.
In problems such as video analysis, particle physics, or genomic analysis, a data set can have billions of features. Deep learning can work with massively “wide” data sets like that.
Labels or tags provide valuable information about a package of data. For example, an image might carry the label “cat.” For unsupervised learning, deep learning works with data that lacks informative labels, such as a bit-mapped image. Deep learning also harbors some disadvantages compared to other machine learning techniques.
Deep learning is a complicated procedure that requires many choices by the practitioner. These options include such things as network topology, transfer functions, activation functions, and the training algorithm. Methods and best practices are nascent; data scientists often rely on trial and error to find a working model. Consequently, deep learning models tend to take much more time than simpler and more mature techniques.
Deep learning models are difficult or impossible to interpret through inspection of model parameters. Such models may have many hidden layers, which have no “real world” referent. Data scientists evaluate the model by measuring how well it predicts, treating its internal structure as a “black box.”
Like many other machine learning techniques, deep learning is prone to overfitting, a tendency to “learn” characteristics of the training data that do not generalize to the population as a whole. Dropout and regularization techniques can help to prevent this problem. As with any machine learning technique, organizations should test and validate the model, and evaluate the accuracy with data that is independent of the training data set.
Training a deep learning model can require billions of computations. While it is possible to perform this task on conventional hardware, some industry analysts recommend specialized GPU-accelerated platforms. This hardware is not cheap; moreover, due to the demand for high-performance machines, some customers report back orders and extended delivery timelines.
Deep learning models are complex, which makes it harder to deploy in a production system. Due to the model’s opaqueness, organizations may need to implement additional measures to provide explanations to users.
Can you think of any benefits of disadvantages that your business or organization has faced as a result of deep-learning implementation?