BRAIN

Project Reference: BRAIN 

How does the brain work and what can we learn from it?
Contact: frederik.mallmann-trenn@kcl.ac.uk

Background:

If we contemplate the question, "What makes a human, a human?" A rudimentary yet illustrative response might be that a human possesses a head and a body. The head contains eyes and a mouth, and the eyes are equipped with an iris and a pupil, and so forth. Such are presentation is referred to as a "hierarchically structured" concept, as depicted in Figure 1. 

Interestingly, our brains seem to identify a human without necessitating recognition of all the components that typically define a human. To illustrate this, consider Picasso's art in Figure 2. The artwork is noticeably missing many details: there's no mouth, only one eye is visible, and we can't discern any legs. Yet, unmistakably, we perceive a human. 

Upon observing this Picasso piece, our brains instantaneously recognize a human despite the absence of certain body parts. We were fascinated by how the brain manages to achieve this feat. Specifically, we formulated models demonstrating how concepts are stored in the brain and how the brain can recognize and learn these concepts even when the input is noisy. 

Drawing inspiration from image recognition research [X1 find a more apt reference], which elucidates how artificial neural networks stores and learns information: they show that in a multi-layered neural network, basic concepts like colors and lines are stored in the lower layers. Conversely, more intricate information, such as shapes or objects, resides in the deeper layers. The deeper one ventures, the higher the level of concepts encountered.  

The question then arises - is this unique to artificial neural networks or does it extend to the brain? While giving a definitive answer is impossible, due to our current technological constraints on retrieving information from a brain neuron, Professor Lynch and Dr Frederik Mallmann-Trenn were able to show mathematically that this is possible in an abstract model of the brain. Under a series of assumptions, they prove that the learning of concepts can in a robust fashion, enabling the brain to recognize known concepts even amidst noise. The researchers show that, in their model, the target concept hierarchy gets imbedded in the brain (see Figure 2 rhs), every sub concept (e.g., head) gets is learned in the process.  However, their results do not hold if the noise level is too high – in this case recognition becomes unfeasible, which seems as natural if one considers certain other Picasso paintings... 

The initial research paper can be located in [1]. Recently, in [2], the researchers generalized theirwork and examined, among other things, the implications of when concepts overlap (refer to the Italian Menu example). The research suggests that if the overlap isn't excessive, we can still recognize noisy inputs - such as unsuccessful attempts at preparing Italian dishes. 

[1] https://www.sciencedirect.com/science/article/pii/S089360802100304X?via%3Dihub 

[2] https://link.springer.com/chapter/10.1007/978-3-031-32733-9_4 

[3] Bolei Zhou, David Bau, Aude Oliva, and Antonio Torralba. 
Interpreting deep visual representations via network dissection. 
IEEE Transactions on Pattern Analysis and Machine Intelligence,836
41(9):2131–2145, September 2019 

Objectives:

Work on related questions. What can we learn form the brain and can we use it for ANNs?