No, models trained on PrismRCL version 2.3.4 or earlier are not compatible with version 2.4.0 and later. However, you can still use and load these older models with PrismRCL 2.3.4. For additional information regarding compatability of models trained on 2.3.4, please read this blog post .

Our development team is continually working to enhance PrismRCL’s capabilities. We have now extended our support beyond PNG images and text data to include tabular data, broadening the range of data types our tool can process.

PrismRCL is an alternative to deep learning, and therefore the models it creates are specific to the algorithm. We are in the process of building a library of models that we plan to make available on demand to our power users through our API. Those pre-trained models may then provide a good starting point for some applications and use cases.

While cross-validation is not currently a built-in feature, it can be easily done in PrismRCL. Given a large dataset, PrismRCL can perform n-fold cross validations just like you would do with a neural network. Split your dataset into n training folds, and train on each fold and evaluate, then take the majority vote based the accuracies that you get from each training session.

When you run inference, you are doing so on a dataset not previously seen by the model. Therefore, the trained model does not have the class information of the inference dataset, which means that the confusion matrix and accuracy values cannot be calculated. You will notice that PrismRCL does output accuracy numbers and a confusion matrix when testing or evaluating a trained model. In that case, the application has access to the class information of the test dataset and can perform the necessary calculations. You can find those results in the summary file that is generated at the end of a training and evaluation session.

Since the application is not designed to be used via a GUI, error logging is also done via text files on the filesystem which makes integration with other applications and workflows a lot easier. In addition to logging errors in log files, it may not be a bad idea to out the errors in the command prompt.

RAM requirements will depend largely on the size of your dataset. Clearly, large datasets will require more RAM to be processed. This is also true for neural networks running on GPU hardware. Google Colab also requires a lot of GPU RAM for large datasets, and will often crash if the dataset is larger than the 55 GB usually allocated on a Pro Plus account.

Real-time updates are available in the logfiles. There you can see what the application is doing step-by-step. Having the status logfiles saved to the filesystem allows for easier scripting and integration with other applications which could make use of the log files.

The software is designed to run from the command line to make it easier to automate tasks and to run it unattended. Having a GUI means that you have to use it interactively, which is not the intent. Also, the design of the application has allowed us to build an API for it. Having the application work in this way made the API possible.

Yes, RCL text classification supports characters encoded in UTF-8, allowing for a wide range of languages beyond English.

In neural networks, every node of one layer is connected to many nodes of the next. In contrast, RCL employs minimal connections. Each node within a layer has no more than one parent. By moving away from the connectionist paradigm of neural networks, RCL makes significant gains in speed and transparency.

Where neural networks imitate the architecture of the brain, RCL has been designed to imitate the mind, specifically the function of drawing distinctions, as described in the tradition of Husserlian phenomenology. While neural networks tend to gravitate toward common patterns, RCL identifies patterns by distinctiveness and therefore illuminates patterns of greater subtlety and interest.