Humans can categorize that these are all images of dogs, despite their differing poses and shapes, but a computer would have a difficult time doing so. New research shows that “random projection” can be used to develop new algorithms, allowing computers to identify and categorize stimuli with the same accuracy as human subjects. (Source: Wikimedia Commons)
Humans have the incredible capacity to learn how to recognize and categorize stimuli, even if only given a small portion of the object or rotating it to a different angle. However, this task is difficult for machines, which must follow strict commands that are often unique to specific situations. Any changes to the input, such as showing an image of a real dog to a computer trained to recognize cartoon dogs, can cause the machine to produce incorrect results. A new study, published in Neural Computation, demonstrates a new algorithm for machine learning based on human models of processing stimuli. The study focuses on a method known as “random projection” (1).
The team, led by researchers at the Georgia Institute of Technology, first identified how humans develop the capacity to fill in the gaps and classify data, even when only shown a small portion of the stimulus or in differing contexts (2). Subjects were shown small, abstract images for 10 seconds each and were then given a set of 16 sketches, one of which matched a random region of the original image. The subjects were then asked to identify the matching sketch; researchers found that they needed less than one percent of the original sketch to choose the correct answer (2).
They hypothesized that humans use a method known as “random projection,” or RP, to identify the correct sketch. Most data in the real world is “high-dimensional,” containing several factors, such as shape, color, texture, depth, and context. Random projection reduces the number of variables while retaining a large amount of the original information (3). Next, the team created an algorithm that modeled RP and administered the same sketch identification task to computers; machines following the algorithm were able to complete the task with the same accuracy as human subjects (1).
The new algorithm presents a significant step toward creating computers that can learn to distinguish and categorize high-dimensional stimuli from the real world. However, the research does not prove that humans use random projection to identify objects, and further studies must be completed, both to further understand human learning and to refine computer learning algorithms (2).
References:
- Georgia Institute of Technology. (2015, December 15). How the brain can handle so much data. ScienceDaily. Retrieved from www.sciencedaily.com/releases/2015/12/151215160649.htm
- Arriaga, R. I., Rutter, D., Cakmak, M., & Vempala, S. S. (2015). Visual Categorization with Random Projection. Neural Computation, 27(10), 2132. doi 10.1162/NECO_a_00769
- Blum, A. (n.d.). Random Projection, Margins, Kernels, and Feature-Selection. Subspace, Latent Structure and Feature Selection Lecture Notes in Computer Science, 52-68. Retrieved from https://www.cs.cmu.edu/~avrim/Papers/randomproj.pdf