General purpose neural computer

• US 5,155,802 A
• Filed: 03/30/1990
• Issued: 10/13/1992
• Est. Priority Date: 12/03/1987
• Status: Expired due to Fees

• Current Assignee: Trustees Of The University Of Pennsylvania (University of Pennsylvania)
• Original Assignee: University of Pennsylvania
• Inventors: Van Der Spiegel, Jan, Mueller, Paul H.
• Primary Examiner(s): MacDonald, Allen R.
• Application Number:
  US07/502,053
• Time in Patent Office: 928 Days
• Field of Search: 364/513, 395/26, 395/27, 395/24, 395/11
• US Class Current: 706/39
• CPC Class Codes: G06N 3/063 using electronic means
• CPC Class Codes: G06N 3/065 Analogue means

Abstract

A general purpose programmable neural computer which parallel processes analog data. The neural computer comprises neural elements for outputting an analog signal in response to at least one input signal, synaptic circuits interfaced with the neural elements for modifying gains of the neural elements, and switching circuits interfaced with the synaptic circuits and the neural circuits for routing signals between the synapse circuits and the neural circuits and for modifying the synaptic time constants, thereby changing connection architecture of the general purpose analog computer as desired. In this manner, the neural computer of the invention can be programmed to learn different confirurations as well as different synaptic values.

Claims

1. A general purpose neural computer comprising:

electronic neuron modules for outputting an analog output signal in accordance with at least one analog input value received on respective inputs thereof;

modifiable synapse modules interconnected via interconnections with said electronic neuron modules for modifying synaptic weights of said at least one analog input value in accordance with first configuration signals;

switching modules for routing second configuration signals between each respective electronic neuron module and synapse module so as to change said interconnections between said electronic neuron modules and synapse modules; and

a processor for providing said first configuration signals to said synapse modules for modifying synaptic weights associated with said synapse modules and said second configuration signals to said switching modules for reconnecting said electronic neuron modules and synapse modules in accordance with a desired neural computing architecture.

2. A neural computer as claimed in claim 1, wherein each electronic neuron module comprises an output line for outputting said analog output signal, electronic neurons for outputting an analog signal of at least a minimum level when an analog input signal thereto is above a predetermined input threshold, and an analog multiplexer responsive to output analog signals output by each of said electronic neurons for sequentially connecting said output analog signals to said output line.

3. A neural computer as in claim 2, wherein said analog multiplexer comprises a shift register for sequentially shifting said output analog signals to said output line.

4. A neural computer as in claim 2, wherein each of said electronic neurons comprises a first current source for controlling said minimum level at said predetermined input threshold, differential amplifying means responsive to said first current source for varying said predetermined input threshold, and a second current source responsive to said differential amplifying means for establishing an externally adjustable threshold bias.

5. A neural computer as in claim 1, wherein each modifiable synapse module comprises a plurality of synapses and means associated with each of said synapses for storing said first configuration signals from said processing means as synaptic weights associated with each corresponding synapse.

6. A neural computer as in claim 5, wherein each of said switching modules comprises an array of analog switches for connecting each respective electronic neuron module to said synapse modules as specified by said second configuration signals and means for controlling time constants of transfer functions of said synapses.

7. A neural computer as in claim 6, wherein said time constants controlling means comprises a plurality of low pass filters for low pass filtering a transfer function of an associated synapse, each low pass filter having an adjustable resistance and capacitance.

8. A general purpose neural computer comprising:

electronic neurons for outputting an analog output signal in accordance with at least one analog input value received on respective inputs thereof;

modifiable synapses responsive to respective outputs of said electronic neurons for modifying synaptic weights of said at least one analog input value in accordance with input configuration signals;

adjustable low pass filters having a variable capacitor and a variable resistor for low pass filtering a transfer function of an associated synapse so as to control time constants of said associated synapse;

a processor for providing said configuration signals to said synapses for modifying synaptic weights associated with said synapses and to said low pass filters for modifying values of said variable capacitor and resistor so as to vary said time constants of said associated synapse; and

summing amplifiers for summing outputs of said low pass filters and providing a sum value to respective ones of said electronic neurons as said at least one analog input value.