Sunday, February 25, 2018

Fast Cortical Learning Using Spike Timing

Abstract

Previously, I argued that cortical feedback connections were essential to handling sensory uncertainty and to cortical learning. In this article, I delve further into how learning works in the cortex. But, first, a word about the Yin-Yang brain.

The Yin-Yang Brain and the Attention Problem

We must keep in mind that the cortex in each hemisphere of the brain actually consists of two mirror opposite hierarchies. That is to say, every cortical column has a complement or opposite column. Although logically linked, the two reside in separate but complementary hierarchies (trees). The same is true for the minicolumns within the columns. If a connection is made to one minicolumn, a complementary connection is automatically made to its opposite. This Yin-Yang architecture also applies to the thalamus (which is where elementary pattern detectors are) and the rest of the brain.

Yin-Yang Brain
The main reason for having a Yin-Yang brain is that we live in a Yin-Yang reality. It is for this reason that biological sensors and effectors come in complementary-opposite pairs. Another reason has to do with attention, the ability to focus on one object while ignoring all others. As I will explain in a future article, the brain has an efficient way to cluster a large number of elements to form a single object. Clustering is crucial to invariant object detection and ultimately to survival. This is an unsolved problem in mainstream AI. The brain uses precise timing to solve the problem: It assumes that all the elements that comprise an object are temporally correlated. The flaw in this solution is that opposite phenomena are not temporally correlated. For example, the motion of an object moving left in the field of view is not temporally correlated with its motion to the right. In fact, the two phenomena do not use the same sensors and pattern detectors. Thus the correlation is not temporal but logical. This is not something that is learned. The brain has innate neuronal mechanisms to handle Yin-Yang logic and thereby marry two opposites into one entity.

For the sake of clarity, I will assume in this article that there is only one hierarchy.

Function and Organization of Sequence Memory


The main function of sequence memory is to combine lots of small elementary patterns into arbitrarily complex objects on the fly, even objects that it has never seen before. It is called sequence memory because the building blocks (minicolums) of the objects are also nodes in highly predictive sequences. Like pattern memory, sequence memory is a hierarchical, feedforward, multilayer, unsupervised, spiking neural network. The bottom layer receives input signals directly from pattern memory. Every layer in the hierarchy, except the top layer, sends outputs to the layer right above it.

Sequence memory consists of a large number of work-alike neuronal structures called cortical columns or macrocolumns. The main function of cortical columns is to detect unique combinations of patterns as they occur. Each column consists on average of 100 minicolumns, each of which attempts to learn a unique pattern combination.
Note: I modified the cortical column diagram that I used in the previous article to better show all the inputs, outputs and feedback connections.
Cortical Column with 5 Minicolumns
As seen above, every minicolumn has 1 feedforward output and 7 inputs (6 associate inputs and 1 primary input). The primary input of a cortical column is the only input that is common to all the minicolumns within that column. Inputs to a column originate from either pattern memory or another level in the sequence hierarchy. Outputs are connected to the level immediately above if any. Every input or output connection is paired with a feedback connection. This means that each minicolumn has 1 feedback input connection (green) from the layer immediately above its own and 7 output feedback connections (blue) that send signals down the hierarchy.

Learning in Sequence Memory

Learning in sequence memory is not about finding sequences but about finding pattern combinations around a central or primary pattern. These combinations are stored in the minicolumns. Once a minicolumn is populated and operational, it can become an actual node in a sequence or even in multiple sequences (topic for a future post). The learning process assumes that the following conditions are met:
  • There is an existing population of cortical columns waiting for input connections.
  • There are a number of pattern detectors or minicolumns that have no output connections.
Here are the learning rules:
  • A small percentage of inputs from either pattern detectors or minicolumns are chosen randomly to be the primary inputs of the cortical columns. The rest will be associate inputs.
  • A pattern detector or minicolumn can only make one output connection to a target minicolumn.
  • Within any column, learning advances one minicolumn at a time.
  • Only perfect pattern signals are used for learning.
  • A signal from a minicolumn is considered good enough if the equivalent of two or more of its connections fired.
  • An input connection to a minicolumn immediately passes the test if it fires concurrently with all existing inputs on that minicolumn.
  • Once a minicolumn has acquired all seven inputs, it is considered mature and learning continues with another minicolumn in the parent column. Unless they are located at the top level, fully populated minicolumns send their output connections up the hierarchy where the same learning method is used.
Keep in mind that this learning method will make bad connections every once in a while. As I explained in the previous article, the brain gets rid of bad cortical connections during REM sleep.

Conclusion

Learning in sequence memory is extremely fast for several reasons. First, a connection needs only pass the concurrence test once. Second, the connection is not severed if it fails the test; it can be tested on the same minicolumn multiple times. Finally, a huge number of connections can be tested simultaneously. The requirement that a minicolumn fires if only an equivalent of 2 or more of its input connections fire is very powerful. This is what allows us to recognize abstract art and see objects in the clouds. How the equivalence is computed is slightly more complicated than it sounds but I'll leave that to a future post. Stay tuned.

See Also:

Solving the Mysteries of Reciprocal Corticothalamic Feedback and Cortical Learning

Thursday, February 22, 2018

Solving the Mysteries of Reciprocal Corticothalamic Feedback and Cortical Learning

Abstract

Neurobiologists have observed (see references at bottom) that neurons in the thalamus, the part of the brain that receives input connections from the body's sensors, not only send output connections to cortical columns in the cerebral cortex, but receive reciprocal feedback inputs from the same columns. No one knows why this happens. What follows is a novel hypothesis that explains the function of the corticothalamic feedback connections as an essential part of the mechanism of sensory perception and learning.
Important note: I am neither a neurobiologist nor a neuroscientist. I get almost all my understanding of the brain by deciphering ancient Judeo-Christian occult texts. If this bothers you, then this article is obviously not meant for you. Sorry.
A Model of the Perceptual System of the Brain and the Cortex

In order to understand why the brain's perceptual system uses feedback signals, it is essential to have an idea of what it is trying to do, how it is organized and the function of its subsystems.

The diagram above is the hypothesized perceptual model. It posits that the thalamus (pattern memory) is where the brain stores a huge number of elementary pattern detectors. These send their output signals directly to the cortex (sequence memory) where they connect to a myriad neuronal structures called minicolumns. These are contained inside bigger structures called cortical columns. There are approximately 100 million cortical columns in the human brain and each has 100 minicolumns on average. Each minicolum consists of 6 associate inputs and 1 primary input. The role of a cortical column is to learn as many pattern combinations as possible. Every connected minicolumn in a column is a different manifestation of the primary input of the column. The green arrow in the diagram represents the feedback signals that return to the origins of the feedforward signals, which are the pattern detectors in the thalamus. The number of feedback connections is equal to that of the feedforward connections.

Cortical columns (see previous article) are arranged in a feedforward hierarchy of up to 20 levels or regions. Pattern signals arrive at the bottom or entry level and percolate up the hierarchy as far as they can go. The activation of a topmost minicolumn signifies that a complex object or pattern has been detected. Normally, many top minicolumns in the hierarchy will fire simultaneously depending on the complexity of the object. Think of an object as a mountain with many peaks and plateaus. How they are clustered together to form a single object is the subject of a future article. Keep in mind that the exact composition of an object is not learned. It is composed instantly even if the brain have never seen it before.
3D Reconstruction of 5 Cortical Columns in Rat Vibrissal Cortex
(Credit: Marcel Oberlaender et al)
What is important to realize is that pattern detection does not occur until and unless a minicolumn has fired. This is how the brain handles sensory uncertainty. The problem is that pattern signals arriving from the thalamus are rarely perfect due to occlusions, noise and other accidents. The minicolumns are, likewise, rarely perfect. The brain solves the uncertainty problem by using a threshold level in its minicolumns that must be reached or surpassed in order to warrant a detection event. When this happens, a topmost minicolumn emits a feedback signal that quickly cascades down the hierarchy one level at a time, branching out as it does, all the way down to the source pattern detectors in the thalamus. The signal branches out because every one of the 7 inputs to a minicolumn is paired with a reciprocal feedback output directed down the hierarchy. In other words, when a minicolumn fires or receives a feedback signal from above, it immediately outputs 7 feedback signals down the hierarchy. This grows exponentially at each level.

Solving the Mystery of Reciprocal Corticothalamic Feedback

Two questions comes to mind. First, why does the cortex use feedback signals? Second, why must the feedback signals travel all the way down into the thalamus? Why-type questions are always the best. The answers we are looking for in this case depend on gaining a good understanding of the cortical learning process:

Reciprocal Corticothalamic Feedback
  • The most important reason for having feedback signals, as explained earlier, is that this is the fastest and most energy-efficient way to solve the uncertainty problem. The solution is to enlist the contribution of many parallel inputs during the detection process. A high enough number of signals arriving at a topmost minicolumn is enough to overcome uncertainty. Contrary to common wisdom, the brain is not a probability thinker but a cause-effect thinker. The brain assumes a perfect and deterministic world. When we recognize grandma, it's not 50% or 90% grandma. It's either grandma or no grandma.
  • The cortex is the seat of episodic memory. When a minicolumn receives a feedback signal, it immediately records a memory trace and the time of the activation. This is crucial because this recording affords us not only a way to recall past events but also makes it possible to predict the future. Of course, the memory trace dissipates quickly unless it is rehearsed repeatedly.
  • Learning in the cortex consists of forming pattern combinations one minicolumn at a time. It is important that learning be as fast as possible. Random inputs are connected to a minicolumn and tested to see if they arrive concurrently. If an input passes the test only once, it immediately becomes a permanent connection. While this learning method is very fast, it can result in erroneous connections because of chance occurrences. There must be a way to correct the errors.
  • The error correction method is straightforward. Every time a minicolumn receives a feedback signal, it strengthens every input connection that just received a strong enough signal. Bad input connections that do not fire on time rarely get strengthened and so remain weak. However, these bad connections are not severed immediately. This happens at night during REM sleep.
  • Finally, the reason that the thalamus receives feedback signals is that connections to the first level of the cortical hierarchy are learned in the thalamus. The reason for this is that learning (searching for viable connections) in the thalamus is faster and easier due to the sheer number of pattern neurons. The thalamic connections must also be strengthened by feedback signals and disconnected during REM sleep if they don't behave as expected and are therefore weak.
Conclusion

To sum up, feedback signals are an integral part of the brain's cortical learning mechanism and its ability to process imperfect sensory signals. In the cortex, they contribute to episodic memory. In the thalamus, their only function is to strengthen good connections and disconnect bad ones. In a future article, I will go over how the cortex clusters large numbers of minicolumns to form invariant objects or concepts. Clusters are also part of the brain's attention mechanism.
And I answered the second time and said to him, “What are the two olive branches (clusters) which are beside the two golden pipes, which empty the golden oil from themselves?” (Zechariah 4:12)
See Also:

Fast Cortical Learning Using Spike Timing
Feedback Connections to the Lateral Geniculate Nucleus and Cortical Response Properties
Emerging views of corticothalamic function
Stuff I've Been Working on: The Cortical Column
Fast Unsupervised Pattern Learning Using Spike Timing

Sunday, February 18, 2018

Stuff I've Been Working on: The Cortical Column


The Cortical Column

Understanding the organization and function of the cortical column is essential to figuring out how the brain works. What follows are partial results of my brain/intelligence research over the years. Unlike deep neural nets, the brain can instantly see a complex object that it has never seen before. How does it do it? It learns lots of small elementary patterns (lines, edges, bits of sounds, etc.) by creating simple sensors that reside in the thalamus.

All elementary patterns come in opposite/complementary pairs. They are the building blocks of all objects. The brain can instantly reuse them to detect any complex object on the fly. This is crucial to survival. Object detection is the job of millions of cortical columns. These are organized into two yin-yang or mirror hierarchies of up to 20 levels. The object detection process is fast and simple and requires little computation. Signals from pattern detectors simply percolate up the hierarchy according to their temporal signatures. An entire detection process, from elementary pattern detections to recognition feedback signals, takes about 10 milliseconds.

Each column can learn dozens of small pattern combinations stored in minicolumns. The combinations in every column revolve around a single pattern detector called the primary input. Only one combination can be detected at a time. Each minicolumn has one output that is sent to a higher layer. Each also receives a feedback connection from the layer above it. Feedback signals are recognition events that serve to correct incomplete pattern detections. How the combinations are learned is the topic for a future article.

The cortical hierarchy is a magnificent machine. It can do all sorts of beautiful and wonderful things that I cannot go into in this article. I will conclude by adding that an activated topmost minicolumn in the knowledge tree (a branch) represents a complex sensed object or pattern at a point in time.

I don't know when but there will be more to come. Stay tuned.

I'm Working on Stuff

A Little Taste of What I'm Working On

That's all.

Friday, February 9, 2018

Busy Days

I'm Still Alive

I'm just busy developing a smartphone application for the hearing impaired market. It's slow going but I'm hoping I can use it to raise enough money for much bigger robotics projects I have planned for the future. Stay tuned.

Friday, February 2, 2018

People Ask Me, What Do You Have Against Deep Learning?

Neurons
Yes, I Cannot Stand Deep Learning

I got a closetful of criticisms against deep learning. I have written about them in the past. I will not list them here because what would be the point? I am not really against the technology of deep learning per se. It is useful for what it does. I am just against the idea advanced by mainstream AI that deep learning is a step toward artificial general intelligence (AGI) or human-like intelligence. In this context, let me just say that, if you are researching AGI, deep learning must be thrown away like yesterday's garbage for this one specific reason if for no other: A deep neural net learns complex patterns but the brain does not. The brain can instantly see a new complex pattern without learning it. Let me say this again for emphasis because it is crucial to my position:
A deep neural net learns complex patterns but the brain does not. The brain can instantly see a new complex pattern without learning it.
Huh? That's right. In fact, almost everything the brain sees is new, that is, seen from different angles or under different lighting conditions. There aren't enough neurons and synapses in the brain to store all the possible patterns that it would need to learn in order to interact with the world. We can instantly see complex objects or patterns that we have never seen before. A deep learning system would be blind to them. We only remember high level bits and pieces of the patterns that we see. Most of the low level details are either forgotten or are written over by new experiences.

As the late philosopher and AI critic, Hubert Dreyfus, was fond of saying, the brain does not model the world. The world is its own model. The brain simply learns how to see it. There is huge difference between the two, one that I hope will, one day, be common knowledge in the scientific community. Dreyfus was saying this decades ago. He was at least a hundred years ahead of mainstream AI.

See Also:

The World Is its Own Model or Why Hubert Dreyfus Is Still Right About AI