What smart machines should learn from the neocortex / SurprizingFacts


Computers transformed work and entertainment, transportation and medicine, games and sports. And for all their power, these machines are still unable to perform the simplest tasks that a child can handle, for example, moving in an unknown room or using a pencil.

Finally, a solution to this problem becomes available. It will appear at the intersection of two areas of research: reverse brain development and a lushly flowering field of artificial intelligence. In the next 20 years, these two directions will grow together and launch a new era of intelligent machines.

Why to build smart machines we need to understand how the brain works? Although such machine learning techniques as deep-seated neural networks have recently shown impressive results, they are still very far from intellectual, from the ability to understand and act in the world around us the way we do it. The only example of the intellect, the ability to learn, plan and execute what is planned, is the brain. Therefore, we must understand the principles underlying the human intellect and use them in developing truly intelligent machines.

In our company Numenta, located in Redwood City, Pa. California, we study the neocortex (the new cortex) – the largest of the components of the brain, and the main one is responsible for the intellect. Our goal is to understand how it works and to define the principles that underlie human consciousness. In recent years, we have achieved great success in our work by identifying several properties of biological intelligence that we believe must be introduced into future thinking machines.

To understand these properties, one must begin with the fundamentals of biology. Humanly, the brain is similar to the brain of reptiles. They have a spinal cord that controls reflexes; A brain stem controlling autonomous behavior, such as breathing and heart rate; The middle brain controlling emotions and the simplest behavior. But humans, and even all mammals, have what the reptiles do not have: neocortex.

The neocortex is a strongly folded sheet 2 mm thick. If it could be stretched, it would be the size of a large napkin. In humans, it occupies 75% of the brain volume. It is this part of it that makes us intelligent.

At birth, the neocortex knows almost nothing, it learns through experience. All that we learn about the world – driving a car, making coffee in the machine, and thousands of other things that we interact with every day – is stored in the neocortex. He learns what they are, where they are and how they behave. Neocortex generates commands for motor skills, so when you cook food or write a program, this behavior is controlled by the neocortex. Language is also created and understood by the neocortex.

The neocortex, like the entire brain and nervous system, consists of cells called neurons. Therefore, in order to understand how the brain works, we need to start with neurons. Your neocortex contains about 30 billion neurons. A typical neuron has one tail-like axon and several tree-like extensions – dendrites. If we imagine that a neuron is a certain signaling system, then the axon is a transmitter, and dendrites are receivers. On the branches of the dendrites are 5000 – 10,000 synapses, each of which is connected to the same synapses of other neurons. In total, the brain has more than 100 trillion synaptic connections.

The experience you have with the world around you is that you recognize the friend's face, enjoy the music, keep the soap in your hand – appeared as a result of input data from the eyes , Ears and other sensory organs that passed before your neocortex and caused groups of neurons to work. When the neuron is triggered, its electrochemical surge travels along the axon and passes through the synapses to other neurons. If the receiving neuron receives enough input pulses, it can work in response and activate other neurons. Of the 30 billion neurons contained in the neocortex, 1-2% work at any time, which means that millions of neurons are active at any given time. The set of active neurons changes when you move and interact with the world. Your sense of the world, what you can consider a reasonable experience, is determined by the ever-changing pattern of active neurons.

In the neocortex these figures are stored due to the formation of new synapses. Their storage allows you to recognize faces and places when you see them again, and recall them from memory. For example, when you think about your friend's face, a neuronal drawing appears in the neocortex, similar to the pattern that appears when you actually see his face.

It's amazing how simple and complex the neocortex is at the same time. It is complex, because it is divided into dozens of sites, each of which is responsible for different conscious functions. In each region there are many layers of neurons, as well as dozens of types of neurons, and these neurons are combined into complex complexes.

The neocortex can be called simple, since the details of each site are almost identical. In the process of evolution, a single algorithm appeared, applicable to everything that the neocortex does. The existence of such a universal algorithm is an exciting fact, because if we can decipher it, we can get to the heart of the notion of "intelligence", and introduce this knowledge into future machines.

But is not this what AI is doing? Are not all AIs built on "neural networks", similar to those that exist in the brain? Not really. AI technologies do refer to neurobiology, but they use an overly simplistic model of a neuron that misses the main aspects of real neurons, and they do not connect like in a real and complex brain architecture. Differences are many, and they matter. That's why today's AIs are good at marking out images or recognizing speech, but they can not reason, plan, and act creatively.

Recent advances in understanding how the neocortex works lead us to guessing how Arranged thinking machines of the future. I will try to describe these aspects of biological intelligence that are necessary, but are not present in modern AI. This training is a re-installation, distributed presentations and an embodiment related to the use of motion for teaching the realities of the world.

Re-assembly training [learning by rewiring]: the brain displays amazing properties associated with learning. First, we learn quickly. Secondly, training is gradual. We can learn something new without training the brain from scratch and not forgetting what has already been taught. Thirdly, the brain learns constantly. Moving around the world, planning and acting, we do not stop learning. Rapid, incremental and continuous learning are essential ingredients that allow smart systems to adapt to a changing world. Neurons are responsible for training, and the complexity of real neurons makes them a powerful learning machine.

In recent years, neuroscientists have learned several interesting facts about dendrites. Each of its branches works as a pattern recognition set. It turns out that 15-20 active synapses on the branch are sufficient to recognize the pattern of activity in a large set of neurons. Therefore, a single neuron can recognize hundreds of different patterns. Some of them force it to work, while others change the internal state of the cell and work as predictions of future actions.

At one time neuroscientists believed that learning occurs solely through the modification of the effectiveness of existing synapses, so that when an incoming signal occurs, the probability of inclusion The neuron by the synapse decreased or increased. But now we already know that most of the training is due to the cultivation of new synapses between the cells – there is a "re-assembly" of the brain. Up to 40% of neuron synapses are replaced daily with new ones. New synapses lead to the appearance of new communication schemes between neurons, and, consequently, to new memories. Since the branches of dendrites are practically independent, then when a neuron learns to recognize a new pattern on one of the dendrites, it does not interfere with the fact that the neuron has already learned with other dendrites.

That's why we can learn new without breaking old memories, and We do not need to train the brain from scratch every time we learn something new. Today's neural networks do not have such capabilities.

Smart machines do not necessarily have to model the complexity of biological neurons, but the opportunities available through dendrites and training through re-assembly are a must. These opportunities should be in future AI systems.

Distributed representations: the brain and the computer present information in different ways. Any combinations of zeros and ones are potentially possible in the memory of the computer, so if you change one bit, this can lead to a complete change in the meaning – just as changing the letters "o" to "and" in the word "cat" will produce an unrelated The word "whale". Such a view is unreliable.

The brain also uses the so-called. Sparse distributed representations [sparse distributed representations, SDR]. Sparse they are called because at any time the activity shows relatively few neurons. The activity of neurons is constantly changing, when you move or think, but the percentage of them is always small. If we imagine that each neuron is a bit, then the brain uses thousands of bits (much more than an 8-bit or 64-bit representation in a computer) to represent information, but only a small part of the bits at any time is 1; All the rest are 0.

Let's say you want to use the SDR to present the concept of "cat." You can use 10,000 neurons, of which 100 are active. Each of the active neurons represents an aspect of the cat, say, "pet", "fluffy", "clawed." If several neurons die or a few new ones turn on, the new SDR will still be a good idea of ​​the cat, because for the most part the active neurons will be the same. Thus, instead of an unreliable SDR representation, it turns out to be resistant to errors and noise. When we build silicon versions of the brain, they will have inherent resistance to errors.

I want to mention two features of the SDR. First, overlay makes it easy to compare two things, and lets you understand what they look like and how they differ. Suppose one SDR represents a cat and the other a bird. Both SDRs will have the same groups of neurons, representing a "pet" and "clawed", but not "fluffy". The example is simplified, but the overlapping property is important, because thanks to it, the brain immediately understands the similarity and difference of objects. This property gives him the opportunity to generalize, which lacks computers.

The second property, union, allows the brain to simultaneously represent several ideas. Imagine that I see an animal running in the bushes, but I could only see it in a glimpse, so I'm not sure what I saw. It could be a cat, a dog or a monkey. Because the SDR is allocated, the set of neurons can activate all three SDRs at the same time, and do not confuse them with each other, since SDRs do not interfere with each other. The ability of neurons to constantly form associations of SDR makes them a good tool for handling uncertainties.

These SDR properties are fundamental to understanding, thinking and planning in the brain. We can not create intelligent machines without using the SDR.

Incarnation: the neocortex receives input from the senses. Every time we move our eyes, limbs or body, the input from the senses changes. This constantly changing input is the main mechanism used by the brain for learning. Imagine that I'm giving you an object that you have not seen before. Let it be a stapler. How will you study it? You can get around it by looking at it from different angles. You can lift, hold your fingers, turn them in your hands. You can press and pull to see how it behaves. In this interactive process, you learn the shape of the stapler, your feelings from it, how it looks and how it behaves. You make movements, feel the change in the input data, do one more, again feel the change, and so on. Learning through movement is the main way of learning the brain. This will be the central component of any truly intelligent system.

I do not want to say that a smart machine needs a physical body – just that it can change sensations through movement. For example, a virtual AI can "move" through the web, navigating through links and opening files. He can study the structure of the virtual world through virtual movements, just as we walk through the building.

This brings us to the important discovery made at Numenta last year. In the neocortex, the data from the sensations are processed by the site hierarchy. When data passes from one level of the hierarchy to another, more complex features are extracted from them, until at some point it becomes possible to recognize the object. Deep learning networks also use a hierarchy, but they often require 100 levels of processing for image recognition, and the neocortex costs only four to achieve the same result. Also, networks with in-depth training require millions of training examples, and the neocortex can learn new objects with just a few movements and sensations. The brain does something fundamentally different from what a typical artificial neural network does-but what?

German Helmholtz, a German physicist of the nineteenth century, was one of the first to propose an answer to this question. He saw that although our eyes move three or four times per second, our visual perception remains stable. He calculated that the brain takes into account the movements of the eyes, otherwise it would seem to us that the whole world is jumping back and forth. Similarly, if you touch anything, you would be confused if the brain were to handle only tactile sensations, and did not know that your fingers are moving. This principle of combining movements with changes in sensations is called sensorimotor integration. As well as where sensorimotor integration occurs in the brain – it was a mystery.

We discovered that sensorimotor integration occurs in all areas of the neocortex. This is not a separate step, but an integral part of the processing of sensations. Sensory motor integration is a key part of the "intellectual algorithm" of the neocortex. We have a theory and a model of how neurons can do this and it overlaps the complex anatomy of the neocortex region.

What are the consequences of this discovery for machine intelligence? Consider two types of files that you can find on your computer. One is an image taken by the camera, and the other is a computer-designed design, for example, an Autodesk file. An image is a two-dimensional array of visual details. CAD-file is also a set of details, but each of them is related to the arrangement in three-dimensional space. CAD-file simulates three-dimensional objects, and not how the object looks from a certain perspective. With a CAD file, you can predict how the object will look from any point of view, and determine how it will interact with other three-dimensional objects. With the image of this you will not do. We discovered that each area of ​​the neocortex learns 3D object models in much the same way as a CAD program. Each time your body moves, the neocortex perceives the current motor command, converts it to a position in the object's coordinate system, and combines this position with the data from the senses to build 3D models of the world.

In retrospect, this observation has the meaning. Smart systems need to learn multidimensional models of the world. Sensomotor integration does not occur in several parts of the brain – this is the main principle of its work, part of the algorithm of the intellect. Smart machines are required to work this way.

Three of these major aspects of the neocortex – re-assembly training, distributed representations and sensorimotor integration – will be the cornerstones of computer intelligence. Thinking machines of the future can ignore many aspects of biology, but not these three. No doubt, we are waiting for other discoveries in the field of neuroscience that shed light on other features of consciousness that will need to be included in similar machines in the future, but we can begin with what we know today.

From the earliest Days AI critics rejected the idea of ​​trying to emulate the human brain, usually arguing that "aircraft do not wave wings." In fact, Wilbur and Orville Wright studied the birds in detail. To create the lift, they studied the shape of the wings of birds and checked them in the wind tunnel. For the driving force, they turned to an area different from aviation – the propeller and the motor. Для управления полётом они наблюдали, как птицы крутят крыльями для создания крена и используют хвосты для поддержания высоты. Именно это они и проделали. Самолёты до сегодняшнего дня используют этот метод, хотя мы крутим только одним краем на крыльях. Короче говоря, братья Райт изучали птиц и затем решили, какие элементы полёта необходимы для полёта людей, и какие можно проигнорировать. Именно это мы и будем делать, создавая думающие машины.

Думая о будущем, я волнуюсь из-за того, что наши цели недостаточно амбициозны. Для сегодняшних компьютеров очень здорово заниматься классификацией изображений и распознавать речь, но мы не подходим близко к созданию по-настоящему умных машин. Я считаю, что для нас жизненно важно заняться этим. Будущие успехи и даже выживание человечества может зависеть от этого. К примеру, если мы собираемся заселять другие планеты, нам понадобятся машины, действующие в нашу пользу, для полётов в космосе, строительства сооружений, добычи ресурсов и независимого решения сложных проблем в окружающей среде, в которой люди не смогут выжить. На Земле мы сталкиваемся с проблемами болезней, климата и нехватки энергии. Умные машины могут помочь нам. К примеру, вполне возможно сделать умные машины, чувствительные и способные работать на молекулярных масштабах. Они могли бы рассуждать о сворачивании белков и экспрессии генов так же, как мы с вами рассуждаем о компьютерах и степлерах. Они могли бы думать и действовать в миллион раз быстрее человека. Такие машины могли бы лечить заболевания и поддерживать наш мир в обитаемом состоянии.

В 1940-х пионеры компьютерной эры чувствовали, что компьютеры далеко пойдут и будут весьма полезны, и что они, вероятно, преобразуют человеческое общество. Но они не могли точно предсказать, как компьютеры изменят наши жизни. Так же и мы можем быть уверены, что по-настоящему умные машины преобразуют наш мир к лучшему, даже если сегодня мы не можем предсказать, как именно. Через 20 лет мы оглянемся назад и поймём, что в наше время прорывы в теории мозга и машинного обучения начали эпоху настоящего машинного интеллекта.