Showing posts with label neural network. Show all posts
Showing posts with label neural network. Show all posts

September 30, 2008

The Binding Of Memory And Hebbian Learning

Imagine this. Your physics professor is taking his regular class, and a boy sneezes. The professor is startled and the chalk falls from his hand. The whole class breaks into laughter. The memory of this particular incident gets stored in your head.

The whole process is somewhat like this. The visual scenes (your teacher’s attire, the chalk and the blackboard etc) are analyzed and processed in the visual association areas; the sounds (sneeze, laughter) get processed in the sound processed in their respective association areas and so on. Thus memory seems to be broken down to its individual elements; and these elements confine themselves in the areas in the neocortex where they were first processed.

human brain showing hippocampus
Hippocampus (means sea horse in Greek) is a small banana shaped structure inside the brain. It gets its input from these association areas via another structure called the parahippocampal cortex. Hippocampus, as if, queries those areas: ‘what happened’, ‘when did it happen’ and ‘where did it happen’? Then the hippocampus binds all those information in the form of an event. For this particular episodic memory (classroom drama); the hippocampus wires together the respective areas so that the whole event is now bound together into an ‘engram’, the proposed neuro-anatomical representative of a particular memory.

While all these are happening in the medial temporal lobe (MTL), more specifically the hippocampus, the actual memory elements are still in the neocortex. With each recapitulation, voluntary (by thinking about the incident) or involuntary (someone else’s sneezing reminds you of that event), the association gets stronger. Thinking of sneezing (memory stored in neuron A) reminds you of chalk falling (neuron B). This way as A becomes active (fires), B is associated too, and this leads to wiring them. This is known as 'Hebb’s rule', after Donald Hebb, a Canadian scientist. Simply put, it says, ‘cells that fire together, wire together’.

The strengthening of synapses as a basis of learning were later found to have been mediated in part to LTP or long term potentiation, a chemical process. While the memories are being strengthened this way, the neocortical areas become more and more inter-connected. This releases the MTL connection, as the memory finally gets settled in the neocortex. This 'plasticity' is important. The neocortex stores memories effectively, has a large storage space, but it learns slowly. The MTL learns quickly, but has little storage space. One can think of neocortex as the ‘hard disk’ and MTL as the ‘RAM’ of a computer. Freeing the MTL would enable it to acquire memory more efficiently.

No wonder my teacher said, ‘Read once, write twice, think thrice’. Its time we consolidated our memories by recap.  

Last modified: never
Reference: hyper-links, unless specifically mentioned

December 13, 2007

Negative Feedback In The Brain

negative feedback in electronic circuitsNegative feedback is used widely in instrumentation electronics, operational amplifiers and in various other streams of electronics. In negative feedback, a part of the output signal is injected back to the input, in antiphase; as shown in the picture on the left. As a result, the gain of the amplifier remains stable, the bandwidth (the frequency range) of its operation increases. In neural networks too, such type of negative feedback is quite common.

The effector (efferent, motor) arm of nervous system, that determines an action, say your hand movement, does so via alpha motor neurons supplying the skeletal or voluntary muscles.schematic diagram of neural negative feedbackAfter you have completed your hand moving job (be it the end of tying your shoelace knot or swatting that biting insect), your hand has to come back to its rest position. This task can be achieved with the use of an inhibitory interneuron (aka Renshaw cells). As shown in the adjoining figure, the yellow colored nerve fiber is the motor neuron which supplies a muscle (shown in brick red). This nerve liberates a neurotransmitter called acetylcholine, which brings about muscular contraction, after the chemical has combined with nicotinic receptors at the motor end plate. Lets make a long story short: we also see one offshoot from the same nerve fiber (also shown in yellow to imply continuity), which supplies another nerve (shown in green color). This nerve, the Renshaw cell or the inhibitory interneuron, secretes a neurotransmitter that has an inhibitory effect on the nerve it supplies. But what do we see? This nerve supplies the same excitatory nerve that initiated the contraction. As a result, contraction ceases. This type of negative feedback network is very common in physiology. The loop back network just described can be found in the spinal cord. Renshaw cells also inhibit surrounding neurons, thus sharpening contrast and focus of 'action'.

If we go further upstream, we can see a similar network in the brain. Before we actually do a movement, the cortical association areas in conjunction to the basal ganglia and the lateral cerebellum plans the movement (this has given birth to the thought controlled devices). This draft (of the plan) is then sent to the motor cortex(M1) and the premotor cortex of the brain which executes the actual action. But you know that any action you do, is overseen by the special senses such as the eye and also as a result of any action such as moving your leg, some receptors in the involved joint which calculates position sense are triggered (i.e they act like kind of pressure sensors). These overseers send these informations (visual cues, degree of stretching of the joint capsule etc) back to the premotor and motor cortex. They then modify or stop the action, in accordance with the input received.

Another instance of negative feedback is found in the cerebellum.feed forward inhibition in cerebellum The parallel fibers stimulate both the basket cells (so called because they wrap around in the form of a basket) and Purkinje cells. The basket cells also drive the Purkinje cells: but this time the response is an IPSP (Inhibitory Post Synaptic Potential) i.e it results in inhibition. This is known as feed-forward inhibition, for obvious reason.

We have seen digital ICs do gating, such as NAND/NOR/OR etc and perform various kinds of Boolean logic. The TTL IC 7400N, for example, perform NANDing. Some descending neurons from the brain, which have inhibitory output, verges on the dorsal aspect of the spinal cord through which pain sensation ascends. This gating, by presynaptic inhibition, may modify pain sensation.

Like negative feedback, positive feedback is also utilized in electronics (oscillators, for example) and in physiology (estradiol positive feedback in LH surge that brings about ovulation is one example). Brothers in arms, isn't it?

Last modified: July 7, 2008
Reference: hyper-links, unless specifically mentioned

November 09, 2007

Spiny Dendrites

illustration of neural networkThe significance of dendrites, neuronal processes, so far thought of only as 'passive components' in the brain circuitry are now being discovered anew. It was known that dendrites could be depolarized to form generator potentials (normally the inside of a cell is negative with respect to the outside. This potential difference or gap in cell voltage, is decreased when a generator potential is established), but it was later found out that they could also produce action potentials (this means that the voltage change was not restricted locally, it spread to other parts of the cell). It was thought that dendrites merely added the multiple inputs algebraically, that is, it acted like an integrator in an electronic circuit. But it was discovered that they were also capable of modifying the signals in a number of ways and they were not just extensions of the soma (cell body of the neuron) so that dendrites only contributed more surface area that were available for integration or synapse formation.

It is now known that dendrites play a very vital role in memory formation and maintenance in addition to its possible role in producing consciousness. Dendrites have tiny knob-like processes on their membranes called dendritic spines. The number and the thickness of these spines vary with age and in disease. For example, in Down's syndrome (or trisomy 21) the spines are thin and small. Their shapes and sizes change in a matter of minutes to hours. They also produce proteins using mRNAs that had migrated from the soma. These mRNAs latch onto ribosomes situated on the dendritic spines and produce proteins. These proteins modify the effects of the input of glutamatergic (glutaminergic) neurons on them. This alters NMDA and AMPA receptor activities. Long term potentiation (LTP) and thus learning is affected by the rearrangement of these spines.