In biology, the genome acts as a constraint, it has a limited capacity to transmit information from one generation to the next. This forces the evolutionary system to prioritize the most useful and general rules for building a functional organism from birth. Similarly, the compression process in artificial networks helps identify the most important features, creating an efficient and adaptable system.
Many animals are born with impressive behavioral abilities. For example, spiders can spin webs, whales can swim, and certain monkeys have an innate fear of snakes. This may seem like magic, but it is a reflection of how evolution shaped these behaviors.
Animals that can perform essential tasks soon after birth are more likely to survive during their vulnerable first days, increasing their chances of reproducing.
These abilities are called innate behaviors. They do not need to be learned; they are already “programmed” in the brain from birth. This is different from learned behaviors, which require experience and practice. However, these two categories are not completely separate.
Innate behaviors provide the basis for learning, and an animal’s abilities are the result of the interaction between the two.
But how are these innate behaviors “written” into the brain? They are encoded in the genome, an organism’s complete set of genetic information. This is where things get puzzling: the genome has a limited capacity for information.
It needs to convey instructions for building complex brain circuits, but it doesn’t have enough space to specify every detail. This creates a problem known as the genomic bottleneck.
Imagine that you had to explain how to build an entire city using just a small instruction manual. That’s how the genome works. It can’t specify every connection between the billions of neurons in the human brain.
For example, the worm Caenorhabditis elegans has 302 neurons and a genome with enough capacity to specify each of those connections. The human brain, with about 86 billion neurons, requires much more information than the genome can provide.
To solve this problem, the genome does not provide a complete "map" but rather a set of rules. These rules define how neurons connect during development. Neurons can be instructed to connect to their nearest neighbors, creating simple grid-like patterns.
These cells can use chemical signals to “find” their destinations, such as axons following tracks of surface markers. In addition, the brain organizes structures into “columns” that can be replicated, reducing the need to specify each connection individually.
These rules allow complex structures to form from relatively simple instructions, such as the organization of the visual cortex or the formation of receptive fields.
However, these rules do not directly explain more sophisticated innate behaviors, such as the fear of snakes or the ability of a spider to build a web. These behaviors require neural networks that can perform advanced computations.
To understand how this works, scientists have drawn inspiration from artificial neural networks (ANNs), used in artificial intelligence (AI).
Natural neuron vs. artificial neural network. Source: Promact
Just like the brain, an ANN has weights that determine how the artificial “neurons” are connected. These weights are like instructions for performing tasks. Researchers have shown that it is possible to compress these instructions into a much smaller artificial “genome.”
This means that even with a limited amount of information, an ANN can perform complex tasks from the start, without additional training.
This compression is analogous to the genomic bottleneck in the brain. Evolution “compacts” genetic instructions to create efficient circuits that allow animals to perform essential tasks soon after birth.
While the genomic bottleneck may seem like a limitation, it can be an advantage. It forces evolution to prioritize the most important circuits, which can be adapted for different tasks. This acts as a regularize, a mechanism that simplifies circuits and makes them more robust.
The researchers propose that evolution works on two levels:
Internal learning: Animals adjust their behaviors throughout their lives, based on experiences.
External evolution: The genome is refined over generations to encode better circuits.
This approach could inspire new algorithms in AI, combining innate (pre-trained) behaviors with adaptive learning.
READ MORE:
Encoding innate ability through a genomic bottleneck
Sergey Shuvaev, Divyansha Lachi, Alexei Koulakov, and Anthony Zador
PNAS, September 12, 2024. 121 (38) e2409160121
Abstract:
Animals are born with extensive innate behavioral capabilities, which arise from neural circuits encoded in the genome. However, the information capacity of the genome is orders of magnitude smaller than that needed to specify the connectivity of an arbitrary brain circuit, indicating that the rules encoding circuit formation must fit through a “genomic bottleneck” as they pass from one generation to the next. Here, we formulate the problem of innate behavioral capacity in the context of artificial neural networks in terms of lossy compression of the weight matrix. We find that several standard network architectures can be compressed by several orders of magnitude, yielding pretraining performance that can approach that of the fully trained network. Interestingly, for complex but not for simple test problems, the genomic bottleneck algorithm also captures essential features of the circuit, leading to enhanced transfer learning to novel tasks and datasets. Our results suggest that compressing a neural circuit through the genomic bottleneck serves as a regularizer, enabling evolution to select simple circuits that can be readily adapted to important real-world tasks. The genomic bottleneck also suggests how innate priors can complement conventional approaches to learning in designing algorithms for AI.
Comentarios