Computer Vision Research Hardware
Donation protected
A few years ago, an app was released that could identify birds using a set of pictures for comparison. At the time, some people asked me why I didn't make something similar for amphibians and reptiles. My answer was that the technology was not developed enough for it to work for the large and variable groups of herps. In 2017, neural network research and programming has progressed to the point where an application can be 'trained' to identify many types of living and inanimate objects. As a part of my development work for the HerpMapper project, I have been collaborating with the Visipedia research group, and they are helping me learn how to train neural networks for image recognition of amphibians and reptiles.
My primary goal is to develop and integrate an identification application for the HerpMapper project. This application would seek to identify a submitted voucher photo of an amphibian or reptile. Providing an identification, with a high degree of accuracy, would open up the use of HerpMapper to a broader audience of people unskilled in herp identification. My initial research and testing leads me to believe that a neural network application that is trained on a smaller set of localized data (a photo set) is more accurate, and will provide better assistance to people wanting to identify local herps. Take skinks in the genus Plestiodon as an example. With a neural network trained to search and compare all lizard species globally, it may give a 45% probability that the picture is of a Five-lined Skink, 40% probability of a Southeastern Five-lined Skink, and a 15% probability of a Northern Prairie Skink. Now if the network is only trained to identify herp species from Iowa, the returned probabilities may be 85% for a Five-lined Skink, and 15% for a Northern Prairie Skink, and the out-of-range Southeastern Five-lined Skink is removed from consideration. To make a useful identification application that works in the field on the user's mobile device, localized neural networks are the way to go.
The primary issue with localized neural networks is that the training process is extremely slow without the use of high-powered graphics cards and a lot of memory. Most people doing research on photo identification use the Amazon EC2 service to run training programs on specialized computer hardware. The service rate for those systems starts at $2/hr and goes up from there, depending on how many graphics cores are used. For my purposes, I believe it is more cost-efficient to purchase the necessary hardware, to be able to train as many neural networks as I need, both for HerpMapper and for other related projects. Which leads me to the point of this GoFundMe request.
NVidia's GeForce Titan X graphics cards are currently the recommended hardware for training neural networks, and at this time, the Titan X cards cost $1650 each. A single card can be used to train the neural networks, but the process must be stopped periodically to allow a verification process to run. Using two cards allows the training and verification processes to run concurrently, and makes the process more efficient overall. My GoFundMe goal of $4000 will cover the cost of two Titan X cards, and a motherboard with two PCI slots to accommodate the cards.
The two TitanX cards will allow me to quickly develop the identification applications needed for HerpMapper, and for other important uses as well. At the Midwest PARC meeting last month, I was asked if neural networks could be trained to identify blisters and lesions on snakes, and flag images as possible candidates for Snake Fungal Disease. Using a high speed dual-card system, I can develop an application that may be useful for important SFD research. I've also had conversations about developing an identification application for freshwater mussels, and another for insect identification. Audio identification may be a future avenue of exploration as well. I'm excited about these possibilities - thanks for considering my request, and spread the word!
My primary goal is to develop and integrate an identification application for the HerpMapper project. This application would seek to identify a submitted voucher photo of an amphibian or reptile. Providing an identification, with a high degree of accuracy, would open up the use of HerpMapper to a broader audience of people unskilled in herp identification. My initial research and testing leads me to believe that a neural network application that is trained on a smaller set of localized data (a photo set) is more accurate, and will provide better assistance to people wanting to identify local herps. Take skinks in the genus Plestiodon as an example. With a neural network trained to search and compare all lizard species globally, it may give a 45% probability that the picture is of a Five-lined Skink, 40% probability of a Southeastern Five-lined Skink, and a 15% probability of a Northern Prairie Skink. Now if the network is only trained to identify herp species from Iowa, the returned probabilities may be 85% for a Five-lined Skink, and 15% for a Northern Prairie Skink, and the out-of-range Southeastern Five-lined Skink is removed from consideration. To make a useful identification application that works in the field on the user's mobile device, localized neural networks are the way to go.
The primary issue with localized neural networks is that the training process is extremely slow without the use of high-powered graphics cards and a lot of memory. Most people doing research on photo identification use the Amazon EC2 service to run training programs on specialized computer hardware. The service rate for those systems starts at $2/hr and goes up from there, depending on how many graphics cores are used. For my purposes, I believe it is more cost-efficient to purchase the necessary hardware, to be able to train as many neural networks as I need, both for HerpMapper and for other related projects. Which leads me to the point of this GoFundMe request.
NVidia's GeForce Titan X graphics cards are currently the recommended hardware for training neural networks, and at this time, the Titan X cards cost $1650 each. A single card can be used to train the neural networks, but the process must be stopped periodically to allow a verification process to run. Using two cards allows the training and verification processes to run concurrently, and makes the process more efficient overall. My GoFundMe goal of $4000 will cover the cost of two Titan X cards, and a motherboard with two PCI slots to accommodate the cards.
The two TitanX cards will allow me to quickly develop the identification applications needed for HerpMapper, and for other important uses as well. At the Midwest PARC meeting last month, I was asked if neural networks could be trained to identify blisters and lesions on snakes, and flag images as possible candidates for Snake Fungal Disease. Using a high speed dual-card system, I can develop an application that may be useful for important SFD research. I've also had conversations about developing an identification application for freshwater mussels, and another for insect identification. Audio identification may be a future avenue of exploration as well. I'm excited about these possibilities - thanks for considering my request, and spread the word!
Organizer
Don Becker
Organizer
Cedar Rapids, IA
