When it comes to cognitive and recognition tasks, the brain of your pet dog can still outperform the world’s most powerful digital computers. This is because data processing and data storage are inseparably linked in vastly complex neuronal networks while digital computing is slowed down by log-jams at the “von Neumann bottleneck”, as instructions and data are shuttled back and forth between memory and processor cores.
Neuromorphic engineering hopes to close the gap between biological and digital computing by mimicking neural pathways. What sort of devices could be used in such systems and how can these devices learn to learn? One of the preconditions of learning in biological systems is long term potentiation (LTP). This is the enhancement of the signal between two neurons that results from stimulating them simultaneously. It is one of the major cellular mechanisms underlying learning and memory. So synapses are plastic: the strength of the signal between two neurons can be modified, encoding memories.
Memristor devices may be able to emulate this plasticity. A memristor device is a non-linear resistive device with inherent memory: when current flows in one direction through the device, resistance increases; when current flows in the opposite direction, resistance decreases. When the current is stopped, the device retains its last resistance value. A memristor could therefore behave like a synapse in the brain: it can adopt high or low resistance states and retains information even when the current is switched off.
In their new paper Martin Ziegler, Hermann Kohlstedt and a team working in Germany have used a memristive device to demonstrate all forms of implicit or unconscious memory, including a recreation of Pavlovs classical conditioning experiment (this time no dogs were required!).
The implicit memory underpins the learning mechanisms of habituation, sensitization, and classical conditioning. Habituation and sensitization are non-associative types of learning. In habituation, repetition of a stimulus leads to an increasingly diminished response, whereas through sensitization a repeated stimulus leads to a heightened response. The team used Cu-doped Ge0.3se0.7 solid electrolyte-based non-volatile memory cells in an analogue and first demonstrated both sensitization and habituation processes.
Classical conditioning is a type of associative learning. The classic example is that of Pavlov’s dog. In the experiments carried out by Ivan Pavlov in 1929, a dog was presented with food (the unconditioned stimulus) causing him to automatically salivate (unconditioned response). If a bell (neutral stimulus) was rung at the same time as food was presented, the dog learned to associate the neutral stimulus with the unconditioned stimulus and began to salivate upon hearing the bell (conditioned response).
By applying appropriate and timed stimuli and observing the electrical behavior of the circuit, the team showed that their circuitry too could learn to associate two previously independent stimuli.
But the authors have another important finding. In order for a memristive device to behave as a realistic substitute for the basic building blocks in nerve cells, it must exhibit a threshold voltage. In these experiments, the device exhibited a negative and positive threshold voltage. Below the positive threshold voltage (or above the negative threshold voltage) the memristive device behaved as a linear resistor and only above (below) the effective threshold voltage did the device act as a memristor.
This report shows that memisistive devices can be used to emulate basic neurobiological learning phenomena. This Pavlov’s dog can learn new tricks.
Image: http://www.flickr.com/photos/tudor/3098397043/ CC BY-SA 2.0
