Bayesian Deep Learning

One of the driving reasons behind Desupervised, is that we want to make AI available, safe, and useful to end users. To benefit from AI today, you need a large team of engineers and if using standard methods, the built AI models will not have any sense of uncertainty.

We call this the problem of uncertainty and we think using such AI models is outright dangerous. The approach we are aiming to take with Alvíss AI is to use Bayesian deep learning, an alternative to conventional machine learning where uncertainty estimation is a first-class citizen.

The Problem with Ignoring Uncertainty

When asking a question, in addition to the answer itself, we also want to know how likely the answer is to be correct. We would think it is ridiculous to trust all answers without considering how knowledgeable and confident each source is. That's the problem with many AI-based models today where all models appear equivalently confident in their predictions. When the models are not aware of what they do or do not know, there is no way for us to tell which one we should trust.

"Uncertainty" can be a scary word. Especially when it's used around automation. However, the information that uncertainty provides is crucial in order to make AI safer. Lack of uncertainty quantification exposes users of these models to overconfident prediction. Being able to quantify uncertainty means a model can communicate how credible it estimates its prediction to be. For example, if a model learns the growth rate of pineapples in Indonesia and then asks how it would grow on in Denmark, the model should be extremely uncertain about its answer given that it has no data on a pineapple growing in any similar weather conditions.

For a more detailed (and mathematical) description of the uncertainty problem, see this article by Yarin Gal.
Illustration of uncertainty coming from data (aleatoric) and model (epistemic). Read more about this topic in this article.
Each type of visual aid has pros and cons that must be evaluated to ensure it will be beneficial to the overall presentation. Before incorporating visual aids into speeches, the speaker should understand that if used incorrectly, the visual will not be an aid, but a distraction.

Planning ahead is important when using visual aids. It is necessary to choose a visual aid that is appropriate for the material and audience. The purpose of the visual aid is to enhance the presentation. Each type of visual aid has pros and cons that must be evaluated to ensure it will be beneficial to the overall presentation. Before incorporating visual aids into speeches, the speaker should understand that if used incorrectly, the visual will not be an aid, but a distraction. Planning ahead is important when using visual aids. It is necessary to choose a visual aid that is appropriate for the material and audience. The purpose of the visual aid is to enhance the presentation.

Deep Learning

What people today call Artificial Intelligence and Machine Learning, usually involve a technology named Deep learning. Deep Learning is a class of machine learning algorithms that rely on mathematical models that are very flexible with a huge amount of parameters, this means that it can be applied to a lot of different areas.

But this flexibility comes with a cost. Often, the model "remembers" the data it was trained on such that it becomes overconfident in what it predicts when exposed to new data. To combat this phenomenon the model needs to learn about doubt and the best way to teach it is to enable it to estimate uncertainty.

Bayesian Deep Learning

One solution to the uncertainty problem is Bayesian deep learning, which only recently became viable in computing. It combines the power of deep learning with prior domain knowledge. This enables all models to not only make predictions from the knowledge it has learned but also tell how confident it is about that prediction.