User:Iutschig/proposal

Efficient Coding Hypothesis: Outline Note: bolded term is suggestion for possible headings.
1. Due to constraints in the visual system such as the number of neurons, the energy required for “neural activities” and noise, the visual processing system must have an efficient strategy for transmitting as much information as possible

• “Bottlenecking” occurs during the transmission of information from retinal receptors to the V1. Following bullets as paraphrased from Zhaoping, 302.
• Retinal receptors receive information 109 bits/sec
• However, optic nerve composed of 1 million ganglion cells, which transmit at 1 bit/sec, thus optic nerve has transmission capacity of 106 bits/sec.
• Further reduction such that overall transmission is 40 bits/sec.
• Phenomenon of inattentional blindness suggests “data deletion by information selection must occur along the visual pathway.” Is this information selection top-down or bottom-up? Bottom-up selection is requisite for responses “to unexpected events”.
• In organization of our receptors it seems that inputs are already split into different components: “color sensitivities of cones, distribution of receptors on the retina, properties of RF’s of retinal ganglion and V1 cells, and their behavioral manifestations in psychophysical performance.

2. The hypothesis suggests that natural inputs have statistical properties, which ultimately determine the structure of our visual systems.

• Two hypotheses can be tested:

• The “statistics of individual neural responses”. Neurons must “utilize full output capacity”
• The “joint statistics of the responses of a population of neurons” . The efficient coding hypothesis calls for neural responses to not be redundant.

• This hypothesis rests on critical assumptions, as identified by Simoncelli. These must be explicitly defined in order to experimentally test the Efficient Coding Hypothesis. I propose research into further discussion of these assumptions in other sources:

• Input: How do we define what is a natural image?
• Output: Do we have to specify which neurons we are studying? Moreover, what do we consider to be a neural response?

3. In the review paper by Simoncelli, the author lists Criticisms of Barlow’s Theory . The following is a list of these criticisms as summarized by Simoncelli: I suggest further research into these different critiques and gathering of sources which deal with these criticisms by examining those References listed by Simoncelli. The following are from pg. 145.

• There is evidence of “statistical dependency between neurons”
• “Over-representation in cortex”- comparing number of ganglion cells to number of neurons in visual cortex.
• It is very difficult to test the theory.
• “Importance of noise” – ignoring noise and physical constraints are too easy: this must be built into the experimental tests.

4. According to Simoncelli, there are two common methodological approaches for testing the hypothesis.

• Expose neurons to natural stimulation conditions and measure the behavior of the neural responses to develop a statistical framework
• Analyze natural images in terms of a statistical framework to create a constraint under which early sensory processing should conform under the efficient coding hypothesis. Compare this predictive model to actual sensory processing.

5. Since it is a review article, Simmoncelli lists various Experimental Tests done recently to test the validity of the Efficient Coding Hypothesis. The following is a list of the experiments mentioned. I propose doing more in-depth research on several of these experiments:

• Baddeley et al.
• Nirenberg et al.
• Reich et al.
• Wark et al. Sensory adaptation-is our coding strategy actually adaptive?

6. Progress in testing the hypothesis

• understanding early sensory processing
• development of mathematical and engineering tools
• advancement in computing and imaging technologies allowing us to manipulate image data

"==References=="