As a technical people, we usually see AI solutions as a bunch of really smart algorithms operating on statistical models, doing nonlinear computations. In general something extremely abstract, what its roots in programming languages.
But, as “neural network” term may suggest, many of those solutions are inspired by biology, primarily biological brain.
Some time ago, DeepMind researchers published paper: Neuroscience-Inspired Artificial Intelligence, where they highlighted some AI techniques which directly or indirectly come from neuroscience. I will try to sum it up, but if you would like to read full version, it can be found under this link:
Roots of AI
One of many definitions describes AI as hypothetical intelligence, created not by nature but artificially, in the engineering process. One of the goals of it is to create human-level, General Artificial Intelligence. Many people argue if such an intelligence is even possible, but there is one thing which proves it: it’s a human brain.
It seems natural that neuroscience is used as a guide or an inspiration for new types of architectures and algorithms. Biological computation very often works better than mathematical and logic-based methods, especially when it comes to cognitive functions.
Moreover, if current, still far-from-ideal AI techniques can be found as a core of brain functioning, it’s pretty likely that in some time in the future engineering effort pays off.
At the end, neuroscience can be also a good validation for existing AI solutions.
In current AI research, there are two key fields which took root in neuroscience — Reinforcement Learning (learning by taking actions in the environment to maximise reward) and Deep Learning (learning from examples such as a training set which correlates data with labels).
First widely spread computational model for neural networks (directly inspired by neuroscience) created by Warren McCulloch and Walter Pitts proved that Artificial Neural Networks are able to compute any arithmetical and logical functions.
A bit later other researchers proposed mechanisms inspired by neural plasticity — adaptation of brain neurons during learning process. It made neural networks able to learn incrementally via supervisory feedback (Perceptron algorithm) or by finding patterns in unstructured data (Hebian learning) — first attempts to unsupervised learning.
After that, thanks to the development of backpropagation algorithm (method to calculate the error contribution of each neuron after a batch of data) multi-layer neural networks started to occur. What is interesting here, the implications of development in this area for intelligence understanding were firstly appreciated by a group of neuroscientists and cognitive scientists who researched connectionism or parallel distributed processing (PDP) — set of approaches assuming that brain functions based on highly parallelized information processing, proving that human cognition and behavior emerge from dynamic, distributed interactions within networks of simple neuron-like processing units.
Current state-of-the-art Convolutional Neural Networks are highly inspired by PDP approach. In both biological and artificial systems, successive nonlinear computations transform raw visual input into an increasingly complex set of features.
Another the most widely used AI tool, Reinforcement Learning was also inspired by nature, mostly research in animal learning. Temporal difference learning, a prediction-based Machine Learning method considers that subsequent predictions are often correlated in some sense. TD methods learn from differences between temporally successive predictions, rather than having to wait until the actual reward is delivered. It is useful in environments with extremely big search spaces (like game of Go).
Recent AI techniques
Neural Networks and Reinforcement Learning are directly inspired by nature. But brain itself isn’t just the one, global learning system. It’s modular network, with distinct but interacting subsystems underpinning key functions such as memory, language, and cognitive control.
Many of them (and their internal “processes”) are now widely used as a guide or inspiration for modern AI techniques.
Most of old CNN networks worked directly on entire images or video frames — it means that each pixel on them have equal priority at the beginning of the processing. Currently, thanks to visual attention, processing resources are centered on representational coordinates, on a series of regions analysed in turns. It helped to ignore irrelevant objects and increased performance what led to an impressive performance at difficult multi-object recognition tasks, outperforming conventional CNNs, both in terms of accuracy and computational efficiency.
Many neuroscientists argue that human intelligent behavior relies on multiple memory systems — reinforcement-based mechanisms (learned incrementally and through repeated experience) and instance based mechanisms (“one shot” to encode experiences rapidly).
One of the memory-based solution in AI, integrated into deep Q-network (DQN) used to play Atari 2600 video games is “experience replay”. The network stores a subset of the training data in an instance-based way, and then ‘‘replays’’ it offline, learning a new from successes or failures that occurred in the past. The replay buffer in DQN may be considered as a very primitive hippocampus, permitting complementary learning like in a natural brain. It’s pretty similar to the process of information consolidation into neocortex, during a sleep or resting time.
Experiences stored in a memory buffer can not only be used to gradually adjust the parameters of a deep network but can also support rapid behavioral change based on an individual experience. Typical deep Reinforcement Learning architectures fail in one-shot learning, but the ones which are able to select actions based on the similarity between the current situation input and previous events stored in memory, are able to achieve success in it.
Another type of memory taking a part in human intelligence is working memory. It’s a cognitive system with a limited capacity that is responsible for temporarily holding information available for processing. Even if now working memory seems to contrast with Long short-term memory (LSTM), still an approach of allowing information to be kept and maintained until appropriate output is required is clearly inspired by it.
While ordinary LSTM networks have functions of sequence control and memory storage closely intertwined (what is an opposition to human working memory), there are already new types of AI architectures directly inspired by our brains, which separate those two mechanisms — the differential neural computer (DNC). Here again, mimicking human brain made DNC much better than LSTM in wide range of complex memory and reasoning tasks (e.g. in finding the shortest path through a graph-like structure, such as a subway map).
Artificial neural networks suffer from the problem of catastrophic forgetting (one task is forgotten as a new one is learned). To address this issue biological brain seems to have specialised mechanisms to protect knowledge about previous tasks during learning on new ones. Findings in neuroscience like decreased synaptic lability (lower rates of plasticity) in a proportion of strengthened synapses are now an inspiration in development of AI algorithms and led to deep networks with implementation of elastic weight consolidation (EWC).
Research in that field shows better performance without changing network capacity because multiple tasks can be learned in the same time — weights are shared between tasks with related structure. And this led to continual learning in deep RL at large scale.
The future of AI and neuroscience
Is it everything? Of course not! Those were just single examples of AI solutions inspired or guided by our recent knowledge in neuroscience and cognitive science. Some of AI systems match or even outperformed humans in tasks like object recognition or adversarial environments (Game of Go, Atari games).
“Google DeepMind created an artificial intelligence program using deep reinforcement learning that plays Atari games and improves itself to a superhuman level.”
Machines are able to simulate human speech that is almost impossible to distinguish from “the real” counterparts. AI can translate between multiple languages (or even create a new one to increase translation efficiency), create art of music like artists do.
But of course we are still far, far away from a human-level intelligence. We still don’t know yet everything about our brains, roots of our behaviors and our nature.
However as a field of AI is continuously evolving, the same happens to neuroscience. There are better tools for brain imaging and our constantly growing knowledge in bioengineering helps us to understand better processes and computations which happen inside our heads.
Moreover, development is happening in both directions. While neuroscience is obviously an inspiration for AI researchers and engineers, an ability to simulate some of the processes in a controlled environment will bring us some answers in our understanding of our brains (like neural networks inspired PDP researchers).
Dreams, creativity, self-awareness or consciousness. There are many of mysteries waiting to be discovered. Millions of years of evolution standing behind probably the most complex and greatest nature work — the human brain. With a constant development in AI, neuroscience and many other fields of science, each day makes us one step closer to understand our nature.