Abzu was established in January 2018 with the vision of developing a technology capable of delivering the promise of Artificial General Intelligence (AGI). This means a system that can reason about problems it encounters, and independently come up with ideas and strategies based on prior experience.
Yes. The long-term goal of Abzu Labs is to develop a platform for Artificial General Intelligence.
Artificial General Intelligence is a dominating research area for all the actors in the wider high-tech industry. This is because there are virtually no limits to the applications of such a technology.
The domain is currently generating a lot of buzz, but fewer actual results. One area in which AI has made huge strides is in Deep Learning which has been very successful at solving problems in narrow AI, and which is making progress in AGI, too.
Abzu believes that the solution to true AGI lies in moving beyond the limitations of the current generation of artificial neural networks, into an approach which is much more inspired by self-organizing systems in nature, the pinnacle of which is arguably the human brain.
Abzu aims to be part of this revolution, and we believe that our unique technological approach will become a key component in the AGI systems of the future.
Abzu is an applied research company that also produces products in the form of targeted solutions as well as a general purpose platform.
Abzu is owned by the founding team behind the company. We’re a European group of specialists within High-Performance Computing, Computational Neuroscience, Deep Learning, and related fields. In Abzu, we have come together to realize our shared vision of artificial general intelligence.
As a private entity, we also have external investors, most notably PreSeed Ventures, which is Denmark’s largest and most successful innovation incubator.
We see Abzu as a pan-European company. Abzu is legally a Danish company, and our headquarters is in Copenhagen, Denmark. The founders of Abzu come from several different European countries and we currently have offices in Copenhagen, Denmark and in Barcelona, Spain. Our computing cluster is located at a datacenter in Falkenstein, Germany.
Abzu was originally funded by the founders and has been supported by investment from business angels and private venture capital.
Yes. If you are interested in investing in Artificial General Intelligence, you are welcome to get in touch with our CFO/COO Jonas Wilstrup, who handles investor relations.
Yes, if you are passionate about artificial intelligence, and have key competences in one of the fields of interest to us, such as high-performance computing, deep learning, neuroscience, or philosophy of mind, we’d love to hear from you.
Abzu is open to students and researchers looking to do projects with us. If you are a student who is interested in doing a project, be that undergraduate, master or PhD, you are welcome to get in touch.
Artificial General Intelligence is a cross-functional endeavour, so we’re actually interested in all sorts of profiles who have an interest in the field. We do have a strong engineering ethos, as we’re building a very large and complicated system. This does not mean that you need software engineering skills, but expect a work environment that values craftsmanship, attention to details, quality assurance and testing, and verifiable results very highly.
Deep Learning is excellent at finding complicated, nonlinear relationships between the input features and some predicted value. Given enough data, and the right Deep Learning architecture (and other hyperparameters) it’s possible to fit almost any problem.
The Abzu technology will perform on par with Deep Learning, given the same conditions. Where Abzu really excels is when some of the above assumptions do not hold. If for example:
– There is little data available, or even no data at all. Abzu will use prior experience from related learnings to suggest relevant models and start learning from there.
– The right model architecture is not known. Abzu will automatically try out a vast number of recombinations of existing skills to seek out a solution.
– It is unclear which features should even be included in the model. The recombination learning aspect of Abzu will ensure that context is taken into consideration when choosing which features to try. Abzu knows from prior experience which features may matter in a specific domain.
Maybe. At this early stage of our development, we have a pipeline of interesting problems that we’re using our platform to tackle. We select which projects and which partners to work with by a range of criteria, among which are how fit the project is for our platform, as well as ethical, commercial and geographical considerations.
If you have an interesting problem, don’t hesitate to contact Jonas Wilstrup. We’re always open to consider your project.
Well, yes. We try to take an ethically sound approach to the work we’re doing. We will not do projects that we don’t feel benefits society, and we interpret that rather strictly.
This applies to the obvious stuff: no killer robots, no manipulation of public opinion, or similar evil endeavours
Additionally, we have at least for now decided not to do projects that deal with personal data.
The Abzu Platform is not yet ready for general use. We use it internally to build various AGI solutions that help us evaluate and enhance our algorithms, scalability, and security.
If you have an interesting problem, you are welcome to contact us, as we are open to consider collaborative projects with partners
We aim to launch our platform for beta customers early 2019. The platform will allow selected partners to build and deploy AI models that utilize recombination learning to tackle problems of general AI. Full public availability of the platform is expected by the end of 2019
Yes and no. Deep Learning can be considered a subset of the approach taken by Abzu. It’s possible to deploy something that’s exactly equivalent to a Deep Learning network on the Abzu platform. But it’s also possible to launch models that break the limitations of Deep Learning in several ways, such as networks based on non-mathematical functions. Deeply nested networks of networks, and networks that discover the relevant connections between cells in various ways.
It may be that someday what we do will be considered next generation deep learning. For now, we choose to call it something else, namely machine cognition.
A combination. In Deep Learning, we have constrained ourselves to “neurons” that have a specific behavior (linear + activation function) exactly because these can be vectorized on GPUs. We believe that the real behavior of the components in self-learning and adaptive systems in nature are not thus constrained, and that this is one of the things holding Deep Learning back.
With Abzu it’s possible to launch Deep Learning style cells on a GPU. These cells are constrained in the same way, but enjoys the scalability of GPUs. This can then be combined with completely different types of cells, which can implement any kind of behavior to augment the learning.
Meta-learning means searching for hyperparameters within a model or set of models which are constrained to be Deep Learning, be that normal, convolutional, recurrent, or any interesting combination of these.
This can radically improve the system’s ability to come up with good models, specific numbers or sizes of layers, etc. And it can certainly save huge amounts of manual labor tuning the models.
As such meta-learning is an important tool, but it does not really address the problem of AGI as we define it. Our aim is to be able to teach the system one kind of relation or model, and then have the same system apply this skill to solve ever more complicated modeling challenges.