Training Set-Free, Vote-Based Neural Nets to Capture Group Intelligence
- 详细技术说明
- A method was devised for using human guesstimates to initialize nodes in a machine learning network to strengthen the predictive output.
- *Abstract
-
Current neural networks can find unknown relationships from highly complex and/or non-linear data input and their predictive outcomes. When data is very limited or even absent, a machine generated algorithm may produce random behavior and poor network predictions. External or human input allows the network to adjust to individual nodes and connections, and combine this with a machine generated algorithm to strengthen the predictive output.
Dr. Bruce Kristal and colleagues at Weill Cornell Medical College devised a method of utilizing human guesstimates to initialize nodes in a machine learning network. The host machine will propose at least one question on a predetermined topic to multiple users and use those answers to strengthen the value of the outcome. By giving the set of questions and a choice of results to experts in a field (and weight answers according to professional rank), the responses can be fed into the network and the host network can identify connections between the nodes to generate a response. By analyzing using external input and machine generated algorithms, a hybrid training set is achieved and can be used to predict outcomes where no data could previously be achieved.
By utilizing this method, a network can generate decisions that directly builds diverse predictions using additional synthetic (human) examples to reduce the current necessity of training the network with known input and outcome data required to achieve high accuracy. Input which introduces human knowledge, wisdom or insight into network connections, including interrelationships between variables of interest in the network give the framework to develop software in multiple avenues of use, including speech and handwriting recognition, stock analysis, drug development, event wagering, medical diagnosis, detection of credit card fraud and classification of DNA sequences.
- *Licensing
- Vibhu Sachdevsachdev@cornell.edu (212) 746-6187
- 其他
-
- Patent: US 8,095,480
- Patent application: 20120084238
- 国家/地区
- 美国
