Artificial intelligence has entered the mainstream. We hear about it in the daily news, we use it every day on our devices. Have you ever wondered how it works? How do algorithms make decisions?

Not so long ago the edrone team had finally the opportunity to leave our virtual workspace behind and meet in the real world. We had an unbelievably good time, and that got me thinking about how it’s possible that a 110-person team, mainly created during lockdowns, is so close to each other.

Don’t just look at the benchmark figures when you evaluate your language model. A subjective judgement can also lead to interesting insights.

In real-life applications computational efficiency of ML models is as important as evaluation metrics

Recurrent language models are still alive and kicking. We release pretrained Polish and English ULMFiT models for Tensorflow Hub.

Continual learning is a machine learning domain that aims to mitigate catastrophic forgetting and enable models to be trained with an incoming stream of training data.

There’s a lot to talk about this time so let’s dive right into it. Here’s a list of everything that’s new, improved, optimized, easier-to-use, more powerful and generally better than before.

One of the many views regarding the philosophy of the mind is called „psychophysical dualism.” It was especially present in the texts of Descartes and postulated the separation of the psychological aspect of humanity from its physicality.

When it comes to processing natural language and developing smart voice assistants for eCommerce, sooner or later, you will encounter the term Inter-Annotator Agreement. Such an agreement is critical for the illusion of understanding words by a machine.

Facebook stuck out from the other giants, commonly referred to as the “big four” in terms of voice solutions. It’s a euphemism anyway because not everyone is aware that Facebook has devices with a voice interface at all.

At edrone, we collect tons of data. There is information about the order value among this data, and as you probably assume, there are lots of random values there. But in fact they aren’t entirely random.

The difference between probabilistic output vectors and target vectors is crucial for the network to learn. How can you measure this difference? What does the learning process look like? How do we indicate when our model is smart enough?