This post requires working code and I need to figure out how to use the transform. Work in progress.

Today I was trying to process some very long string. The string actually was a list of pair of values to be translated to a dictionary.

This is a cross-post of the post-mortem from an open-source effort I have made within my organization. The link is here.

I’m sleepy so I will keep it fast

What is to be done in speech recognition? The field is perceived as solved as there are practical implementations of ASRs that can be used in human machine interfaces. Is there anything to be done?

I learned, actually yesterday, that I have Thinker type of teamworker, with Source-seeker splash. This means my forte is at coming up with new ideas while I typically am not good at implementing thm.

Deep learing is about learning features, this is the only tool we know that generated sensible features The classifiers we use are actully pretty simple. The features are the power.

The most visible recent achievements in the field of ML involve multitude of data and powerful models processed by multiple nodes. This general strategy is more feasible for large organisations, because we don’t really have well-grounded understanding of processes involved in deep learning, but it is quite easy to just buy more hardware. However, not every aspect of DL can be made up by scale - some aspects, like training time, are of particular interest of individuals just because it is hard to obtain the right amount of compute. One such things is a solution used in winning DAWNBench, where training times were several times lower than the best solutions presented by any other participant. The winning solutions used a phenomenon called superconvergence, which is based on a particular method of training ANNs using a variation of plain SGD, called Circulating Learning Rate. One characteristing thing about this approach is that the need for having a good learning rate is eliminated, as the CLR method tries several different learning rates during the training process. Since such a solution to iterate fast while solving the problem, and little need for parametrisation, I found the method of interest for my personal use.

Hi! Recently I’m struggling with integrating the Plain Old(tm) WFST with CTC predictions, I wanted to deliver a simpler project in the meantime - that is, to recognize spoken digits in Polish. Solutions to the problem were conceived somewhere in the ancient era, but I’ll use dead simple code for Keras.

Hi! Recently I’ve got C# on my CS course and to familiarize myself more with the topic, I wanted to do something both in line with my interests and that would let me learn more about .NET platform. In 2018 it seems to be rather friendly programming environment, as it turns out, with a very straightforward way to start. I’ve decided to implement basics of backpropagation algorithm, something basic, but nevertheless educational. Here we go:

What would happen if we wanted neural network to predict weight of the proper network? We can actually do this on runtime. Calculation of weights as a Dense of inputs means that we effectively combine each input variable with each other (and we do this twice). Addition of activation function may change this to a more elaborate, non-linear relationship.

In this text, I present an introduction to progressive neural networks, which is an interesting multi-task architecture; I also introduce an example implementation in Keras.

Humans are already bested by machines in particular tasks. Sure, the digital intelligence is still vulnerable to even one-pixel disturbances in the data, but the progress in the field is impressive, taking for example ImageNet challenge, where algorithms have lower error rate than best human annotators. Fortunately, we shouldn’t be afraid of waking up in sci-fi world soon. First, the algorithms are fragile, but second - they’re nowhere near of being intelligent.