Proposal #9 - Modeling the Neo-Cortex Best Proposal prize? no Overall organization/clarity of the proposal? 9 Overall project idea? 6 * Overall, is the proposal clear, concise, and well-organized? Yes this proposal is clear and organized. They review their approach, present the direction they'd like to take, and detailed the evaluation criteria for determining success. They present the material in a well defined proposal format including tile page, contents and references. * Does the proposal meet the posted proposal guidelines? Pretty much, it doesn't really comment on the idea filter to make sure it's not going to be to complicated. It also didn't talk much about practical applications to robotics, other then that it's studying the behavior of neural networks which would have some application to learning systems that might be used by AI in develop based robots. * How does the project idea fit within the framework of Developmental Robotics? My guess is that it fits into a category that is focused on core capabilities of learning systems. So it doesn't tie into the properties of embodiment or look at any of the learning through repetition. It just looks at the concept of relational data organization and retrieval. * Describe what you like BEST about the project idea. Creating Matlab models of the neural network learning process and retracing the steps would be something that could be picked up and utilize to tune current algorithms for better convergence speed and organizational accuracy. * Describe what you like LEAST about the project idea. Approach number 1 seemed to basically use existing neural networks, filling them with data and verifying the results. It didn't seem to push the limits. (Approach number 2 did seem to make up for this, but the two didn't seem to be completely tied together or build upon each other) * Do you have any concerns about the project? No, it looks like a good project for the time a lotted. It has defined evaluation criteria to determine the results and a decent amount of coding and sample data interpretation. It would be a good lead into a Graduate project that would seek to find new neural algorithms and also take these building blocks and build them into a working Robotic system. As a stretch goal, it might be interesting to figure out sets of test data that would directly overlay on an existing robotic body This would more tightly couple with an existing system and the possible behavior it could experience if the software was ported over. * Does it seem doable in the remaining time? Yes, it seems like a good learning activity to understand the process of how the neural networks of today work. * Does it seem too difficult? It doesn't seem too difficult, the Matlab model would be the most challenging part. Gathering the data sets for evaluation might also be a slight challenge. * Are there any major details left out? The discussion about how they're going to generate or borrow the data sets used in analysis. Weather they're going to generate this data, connect to sensors and gather the data, or borrow it from a previous project that did a similar study. I believe it's also missing the discussion about the type data set they're going to use to understand how the neural networks work. The reason I believe this is important is because when using a mixed set of data (images, sound, other sensory input, etc) the complexity goes exponentially up and determining success during evaluation could be come difficult. For example if modeling a Apple, you'd have to come up with relationships where the word apple is linked to pictures of apples and other relevant data. This same rough set of data would be needed for all items the robot learns, so when the robot goes operational and tries to determine what a object is, they have the necessary previously formed relationships already formed during the learning exercise. On the other hand if only one sensory input is selected to begin with and say imagery is that input. The learning process can characterize similar data from that one source and more data could be learned to make the evaluation process simpler for this size of project. * Does the idea rely upon technologies that are not currently available? It does not, but the modeling in Matlab may have some complexity that will need to be flushed out to make sure a result could be found. * Do you have any suggestions for improvement? For your second approach, try to further define what success is by describing your data set in relationship to the evaluation criteria you've defined. * Do you have any suggestions for related work that should be cited? I didn't really notice any citations for the algorithms. I've read some good summary papers that review the concept of Kohonen SOMs. They're available through Google by just searching for Kohonen SOM. * Any other comments or suggestions? I'd suggest making sure in your results that the approach 1 and 2 are more like phase 1 and 2. That way the results of the experiments could make more sense and using Matlab have some awesome graphs.