This webinar on deep learning moves on from building and training a simple neural network to implementing special deep learning architectures for sequential data, called recurrent neural networks.
Sequential data is all around us. Language as a sequence of words, time series as a sequence of numerical values like stock prices or sensor data, or signals as a sequence of samples from a sound wave, to give you just a few examples. This kind of data has special requirements when it comes to deep learning architectures.
What are the goals of this webinar?
In this webinar we start with inspecting different requirements of sequential data and how deep learning models can handle them. We’ll look at different use cases, which have all been implemented without code, including demand prediction with multivariate time series and a number of text based applications.
Who is this webinar for?
We welcome anyone interested in deep learning to join us! It will be of particular benefit to data analysts, data scientists, and deep learning developers who want to take advantage of the KNIME GUI to build, train, test, and deploy deep learning networks.
Who are the speakers?
Join Kathrin Melcher, data scientist at KNIME and Rosaria Silipo, principal data scientist at KNIME and head of the Evangelism Team, who wrote the book “Codeless Deep Learning with KNIME”, which is available via Packt Publishing.
Q&A from the webinar
Check out the questions that were asked during the webinar (click arrow to expand answer) and post more related questions in the comments
Are RNNs also suitable for anomaly detection within sensor signals timeseries?
Yes, they are useful for time series analysis
For RNNs can one encode multiple inputs features at each time steps?
Sure. You can use a full input vector at each step
If I use an input vector at each time step. How does that change my RNN? What kinds of nodes / setup do we need?
You need an input layer with as many inputs as many vector components. See the recording of the webinar for some examples.
In network design, why do we choose TanH rather than some other activation function like sigmoid?
It depends on the problem. Usually we use a sigmoid when we want to interpret the output as a probability. Or a ReLU when we want to produce a number. For the hidden layers there is a lot of trial and error.
We have been talking about creating networks from scratch. Is there a way to do transfer learning with KNIME?
Yes, this is possible. You can read a pretrained network and usethe Keras Freeze Layer to decide which layers you want to retrain
How do you handle negation for the sentiment analysis? (E.g., the movie was not nice.)
The network will take negations into account by remembering in the cell state that the word “not”, for example, was recently in the sequence
In time series analysis, is the 'position' of data respect to time important, so that the analysis is not "position independent" as in text mining case (e.g. prediction of next word). Can/how RNN hand this issue?
The position independent only refers to the ability of the network to extract the same information from all positions in the sequence. Through the hidden state, which connects the network copies from the different time steps, the network is able to learn dependencies between different time steps, so for this it takes into account where in sequence a certain value occurs.
Can I perform classification and use it with other multiple parameters to estimate how long before a task will be performed, using times from previous task ordering and performing times?
This is possible, if you have the data available.
Do these tools need GPU? If yes, what are some ways to get KNIME setup with GPU?
To speed up execution, especially during training, you can use a GPU support. This is also possible in KNIME Analytics Platform. Therefore you can select in your deep learning preferences to create and use a GPU environment.
Can I use an optimizer node in time series using LSTM?
Yes, you can use if for example to optimize the number of units in the LSTM.