Applying Deep Learning in the domain of Signal Processing

Signal Processing is all about understanding patterns .

But what about applying deep learning on sensor data streams ?

There are few questions we need to touch before we go further...

1. What about current methods ?
2. Is deep learning applicable ?
3. What are the advantages ?

Well, in signal processing domain it’s common to see machine learning techniques .

• PCA - Dimensionality Reduction Techniques
• MFCC - Hand Made Features(This is a solid mathematical algorithm capable of extracting frequency components)
• SVM & Other ML algorithms with hand craft features

One thing common to above words is we must first design a set of features from sensor readings and then try to understand patterns . 

So sometimes it’s not good enough . Specially in modern day, we can have lots of data for one task with multiple sensors . Sensors become smaller, cheaper and easy to merge .

There are more sophisticated ways to extract information from sensors which uses Unsupervised Learning :

1. Restricted Boltzmann Machines
2. Auto Encoders
3. Variational Auto Encoders

Basically we use above methods to extract good representation of original data . Then we use extracted feature representations in a simple machine learning algorithm .

Is this the best way ?

No it’s NOT!

These feature representations are not task specific . Its like when your task is to recognize human activities using sensor readings its always desirable to extract something from a signal which, highlight the dynamic of activities .

The domain of signal processing heavily depends on the mathematical modeling . Specially in stabilizing , noise cancelling and filtering .

Sometimes we use filters like Kalman to get a stable readings in noisy deterministic environments .

Any way can we do amazing things with deep learning ?

Let’s find out.

This era belongs to the big data . No need to tell again and again data is the new oil. This is same for the domain of signal processing . For an example with IMU sensors how much data we get ?

It’s huge . Every smartphone has an IMU so we can get huge number of data in high velocity . Therefore, maybe with single IMU there can be lots of decisions we can take. Tracking , HAR(Human Activity Recognition) .

So applicability of deep learning is varying in many dimensions . Here I will try to note down few.

1. In modern day when it comes to sensor readings on one task it’s not getting only from a single sensor . There can be many sensors . For an example in IMU there can be 3. So we can have a better sensor fusing mechanism learn from its own data. Yes deep learning is well applicable in fusing these readings and getting a task specific feature representation.
2. What about the noise in the data. Normally noise is something inevitable in sensor data. So what if a model can learn how to deal with the noise from it’s own data .
3. How to work with long term sequential behaviors and how to use them when recognizing patterns .

There are many . So basically deep learning can learn hierarchical feature representation of signals . Combinations of CNN , and LSTM can do amazing things . Let’s talk about how to start on an another day . Till then I will put some fantastic papers that explain the application of DL in the domain of signal processing .

DEEP LEARNING FOR HUMAN ACTIVITY RECOGNITION

DeepSense: a Unified Deep Learning Framework for Time-Series Mobile Sensing Data Processing

See you!!