Original article can be found here (source): Artificial Intelligence on Medium
Complexities of Working with Deep Learning
Many pundits would argue that anyone with a laptop and an ability to write a few lines of code can directly work with deep learning.
Cloud vendors also provide the options and the APIs to configure your apps, work with deep learning algorithms, and to find meaningful results.
But is this really the case?
Let’s look at some of the complexities of working with Deep Learning/Machine learning projects.
Consider the case of image recognition
Deep learning has done well in this category. Data with their 2-D representations works perfectly well deep learning projects. As images can be represented in an array of data points so playing with them and finding patterns works with 2D images.
But can you extend this approach to all other problems? Let’s take the case of life sciences and working with complex 3-D objects, like molecules. You might be tempted to go the traditional route but that will definitely not lead to any valid findings.
Many practitioners have found that extending 2DL to 3D data is not straightforward. Such tasks are very much dependent on the data representation itself and the task at hand.
Still, a fast-growing community of researchers is working on such problems. With the new research, the gap is fast closing and soon 3D DL problems will have solutions.
Then the other biggest challenge is with the data
In most of the cases, companies grapple with the issue of having not enough data. This leads to using shortcuts and manipulations to augment data.
With such approaches, there are chances to get better accuracy with such tactics for training data sets. Such higher accuracies are not also always the best outcomes and could be the result of unintentional manipulations. Also, nothing much is guaranteed for the actual data set.
So, beware of the hidden cheats used to train the model and that’s where the role data scientist and ML engineers, to clean the data and also to understand the data, becomes super crucial.
Then comes the issue of data relevance or data freshness. Does the data that you had, still add value to the problem that you are trying to solve? Real-world keeps on changing and past data might not be reflective of the recent trend. This results in data and the corresponding models becoming completely irrelevant.
Also, you need to look at the data acquisition costs? Many times companies have gotten to a point where their models require additional data to be of any value. But the cost of acquiring such data might be too high.
Choosing the right algorithms
The next major step is the utility of the right algorithm.
That remains the domain of experts and that’s where the most scarcity of skilled resources comes from. Also given the computational resources being expensive the trail and error could compound the issue.
Still, the cloud operators are working on such problems on are coming up with options.
AWS claims to have come up with services, wherewith no machine learning experience you can use their AI-powered services to figure out the algorithm.
So, for Personalization, Advanced text analytics, Voice, Forecasting, Conversational agents, Transcription Document analysis, Image & video analysis, Translation you just need to plug in your data and their AI-service will do the rest.
Correlation doesn’t imply Causation
That is the fundamental underlying theme before venturing into finding solutions to the data problems.
Let’s take the simple cases of regression. You can pretty much take any two variables and come up with the correlation of fairly significant results. But does that really mean causation?
With the availability of big data, not all the statistical patterns represent actual findings. As machines can always detect noise or biases and spit out patterns that are of no relevance.
So, before diving deep, you need to understand the problem the objectives, and the kind of data you are working with.
Deep learning actual workings to many still remain a black box.
At times, no one knows, how the findings are being classified. And just taking results as true and valid without a deep understanding of the underlying mechanism and the data, is a massive issue.
So, beware, and work with experts.