# What is the real difference between DSP and AI/data science?

I am very much into DSP, but lately everybody is talking about "Data science" and "deep learning", and from what I understand the whole thing about data science is to take a huge row of data, and extract meaningful data only from it.

Isn't it what you do with DSP, when you have a signal with noise and you extract the data from a raw of sampling? Or for example when you auto-correlate to find matches between signals? Do you use the same statistical tools in both cases?

• Maybe this question and its answers will be helpful. – Matt L. Mar 13 '16 at 19:50
• In DSP, one has often learned to deal with only one column (for 1D signals), with which you cannot do much in AI/data science, in general (yet) – Laurent Duval Mar 13 '16 at 21:37
• There is no difference in the statistical tools used for the same sorts of problems --- the tools were developed to solve specific problems, so why reinvent the wheel? Very many data science problems can be recast as DSP problems. However, data science spreads the definition of how the data was acquired... And generally wants more "business-oriented" answers than most DSP problems do. – Peter K. Mar 14 '16 at 1:46
• Votes and best answer validation are required – Laurent Duval Jul 28 '19 at 12:04

The purpose of both DSP and machine learning is to transform the input signal / data set into more meaningful information. This could be,

• removal of noise from an audio recording
• location of faces in an image
• classification of objects in an image
• etc

The main difference as I see it, is that in DSP the transform is designed by the engineer. The engineer will choose a set of signal processing operations that give the desired output. That choice is guided by experience and validated from the results of experiments.

In machine learning the transform is learnt. Typically, this requires a set of training signals with known outputs, over which the system will optimise its transform. The engineer still has to choose a machine learning architecture that can replicate the transform that is needed.

Often the two will be combined. And DSP is kind of machine learning anyway - the machine learning bit is done by the human and the test set is the sum of their existence.

Currently, one difference between data science and DSP seems to be in the amount of literature on computational efficiency at lower transistor counts and/or energy levels. A rack-full of GPU cores with a few kW of water cooling finds many different applications than something tiny that has to "extract meaningful results" using a hearing aid battery.

Another difference is that some well-defined problems have short clean mathematical closed-form DSP solutions that do not need huge data sets for training, plus even more testing.

When performing a tiny edit to the question, I looked at tags, and found that here, so far, only one tag contains data: data-request.

I consider the fields of signal processing, and image analysis, as quite composite. Almost everybody (my grand-mother as well) has some insight of what a chemist, a biologist is. In my time, explaining to my family what signal processing meant was a pain in the ASIC. I had to explain what a sensor was, and how a digital signal was composed. Then, after losing patience, I'd say I deal with mp3 and photoshop. Everybody then understands, but my so-called science suddently vanishes.

Signal processing, and image analysis, are composite. But not similar. They borrow from different other fields. In addition to the interest in low-level efficiency, I believe that a huge part of DSP is dedicated to more or less structured data, with the aim of extracting more or less "constructed" features.

Data science is an other umbrella term that encompasses even more than the two latter. From the wiki page:

[It draws] from many fields within the broad areas of mathematics, statistics, information science, and computer science, including signal processing, probability models, machine learning, statistical learning, data mining, database, data engineering, pattern recognition and learning, visualization, predictive analytics, uncertainty modeling, data warehousing, data compression, computer programming, artificial intelligence, and high performance computing.

Because it is fashionable, because with the advent of social media, people now know that there are data. And sounds, signals and images have becoming no more that "some data". Almost everybody gets what a data is: bits, unstructured. Or not.

This is funny: a lot asked me (in the past) to define a signal, no one ever asked me what a data is (while the question is more fundamental).

And everybody produces data, while few produced signals. The time was ripe. With the production of even more types of data, less structured, the features to be extracted became less and less "contructed", less "low-dimension" and hand-crafted, with interaction with machine learning.

Now, some of the features used in deep learning are less fine-designed than the usual ones, and the huge human-created tagged data compensate for that.

DSP code that i see and try to write usually has a lot fewer conditional execution constructs (if statements) than AI code does.