Dear Lucy,

U can read and follow the following points:

CNNs have been used a lot for non-image data. Few examples are:

Now, how to use CNN for non-image data at a high level can be visualised here in this wonderful blog post, for text classification - Implementing a CNN for Text Classification in TensorFlow

All the links are not working. Anyways, is there another application Different than “text” classification.

For example, classifying weather based on some features.

Hi Lucy,

as "Deep Learning" covers a lot of techniques, there are many examples in the medical area which do not necessarily use images. Some ideas are referenced here:

Article Artificial Intelligence in Precision Cardiovascular Medicine

Greetings, Sebastian

I use deep learning (tiny-dnn) with 1D sensor data for environmental perception and sound event recognition in robotics domain although I still run the neural network on handcrafted features, but a simple MLP-like network still outperform the old methods.

Although my youtube presentation about the C++ programming aspects of this, but you can see my network description and diagrams with comparison of my "deep" network with standard ML methods for sound event recognition:

https://www.youtube.com/watch?v=6flxcxogxbY

@casba Kertez

Thanks for your reply. sound events so the input data is basically signals?

@Lucy: I don't know what you mean about "basic signal", but it is simple audio data and even it is not processed like a time series, the sliding window in the temporal data is taken into account as independent data frames. It does not mean that I could not use some recurrent deep learning architecture for my data, I just have not tried yet. My other works were based on accelerometer and other simple sensors.

@Casba Ketesz

I mean what type of features are you dealing with?

imagine I have limited number of features like:

Features: name, timestamp, actor, and phase (start/complete).

Class: activity name

Can we still apply deep learning approach with such limited number of features?

@Lucy: In my problems so far I dealt with:

- Feature vector size: 10-80 features

- Dataset size: ~10000-850000 samples

Very big datasets with millions of samples and thousands of features are the traditional target of big deep networks, but small topologies are still suitable for moderate sized datasets. Based on my personal experiences, deep learning techniques can be used with small feature vector size (

Although it's a little different from the discussion happening here, you can check this: Article Deep Reinforcement Learning: An Overview

Specifically Sec.5