At AGICortex we challenge popular and recognized Machine Learning methods to push the technology forward. Not by 5%, but to present true technical breakthroughs contributing to the emergence of autonomous AI solutions with continuous general learning capability.
We believe that Artificial Intelligence of the future will be native to both physical and digital worlds.
Here are selected research areas on which we will focus in 2021:
Continuous object detection and classification with dynamic filter repository management
Current state-of-the-art techniques of object detection achieve around 55% accuracy, while using COCO dataset. It is impressive when compared to results around 20-40% that were possible in the previous years.
However it seems not good enough for applications in systems, where correct object recognition is one of the foundations.
In addition, the available solutions are resource hungry what makes it very hard to use them outside the cloud environment, without the significant drop in the delivered quality.
That is why we aim to perform long-term research on the very interesting concept of dynamic convolutional filter repository management. Such approach would not be possible to be realized with classic approach of Deep Learning.
The result will be the availability of large number of filters to pick from, but with utilization of only the most useful ones – based on the context of the situation.
We are also interested to equip systems operating in the physical world with the capability of continuous incremental learning. Therefore, we will extend our proprietary ML library with autonomous filter learning and object classification feature.
Optimization of neurosymbolic 3D neural networks architecture for real world use case scenarios
Most of the progress we want to show in this and the following years is related to a very different approach to Machine Learning. Instead of statically defined 2D neural networks with simple computational units (neurons) – we know that more complex architectures are needed.
We worked on the prototype of 3D neural networks for the last 6 years and now we want to optimize it for real world use case scenarios and business applications.
The difference between 2D and 3D neural networks is that the data flows not only from left to right and backwards – but in any possible direction. The third dimension here acts as a ‘modifier’ of what is happening inside the specific sub-network – literally turning on and off large parts of the neural architecture.
The computational unit is no longer a single neuron, but a group of them. You may think about capsule networks to find any similar concept – but we are way beyond that.
3D neural networks with a set of supporting modules form a digital equivalent of a brain. Able to learn and operate on its own – both with sensory perception and forming the adequate responses via virtual or physical body parts.
During this year we will focus on adapting the technology to use in products with a goal of analyzing the physical indoor environment.
This is only beginning. We will perform even more interesting research in the following years.
We look forward to cooperate with smart people regardless of age, sex, physical location, education(!) and anything besides your ability to solve non-trivial problems with computer code.
If you would like to become a part of this journey:
- Think how you would begin the work with dynamic filter repository management or what you could bring to our research efforts
- Apply here