Leapfrogging Ahead in AI

Source: Deep Learning on Medium


Go to the profile of AI Research

While deep learning has propelled AI at the top of everyone’s mind, it is clear that the approach has severe limitations. Most notably, deep learning still requires large amounts of labeled training data and the models are often easily fooled by seemingly random adversarial patterns as shown below,

Source: https://arxiv.org/pdf/1412.6572.pdf

In the case of human intelligence, on the other hand, the brain seems to learn in a mostly unsupervised way with minimal amounts of labeled training data and perception is rather robust to random perturbations of the sensory input

While deep learning was inspired by ideas from neuroscience, it is often argued if well established fundamental mechanisms such as backpropagation are even biologically plausible (e.g., https://ieeexplore.ieee.org/document/118705).

White matter fiber architecture of the brain. Measured from diffusion spectral imaging (DSI). The fibers are color-coded by direction: red = left-right, green = anterior-posterior, blue = up-down. Source: www.humanconnectomeproject.org

There have also been another fifty years of neuroscience research that are not sufficiently reflected in the current deep learning community. Hence, we believe that the next generation of machine learning algorithms can only be developed by taking inspiration from the newest results in neuroscience research. It is our firm belief that only this will enable the next stage in the pursuit of strong AI and might even further our understanding of human intelligence.

When examining the capabilities of the brain, one of the most fundamental capabilities is the effortless generation of invariant representations of entities in an almost unsupervised manner. By observing the sensory input (visual, haptic, auditory, etc.) the brain is able to group entities (e.g. songs, objects, textures, words, etc.) independent of the transformations applied to them (e.g. changes in the key of a song). In our work, we follow the assumption (attributed originally to Vernon Mountcastle), that the cortex of the brain generates these invariant representations with a single cortical algorithm. However, is not clear at all how this cortical algorithm works or what data structure the brain is using.

To summarize, our main research goal is to propose a learning algorithm that can generate invariant representations in an unsupervised fashion inspired by neuroscience research.

Hence, it is our belief that we need to approach this challenge with an interdisciplinary team of mathematicians, computer scientists, and neuroscientists, which will be able to contribute to the research question of unsupervised generation of invariant representation from three different but equally important angles.

For more details on why we believe neuroscience might hold the key to Strong AI, we recommend a recent Innovator’s club talk titled Leapfrogging Ahead in AI by our Global Head of Data Science and Analytics Helmut Linde.