Share
Embed
Transcript
42 slides
-
Motivation • Read a disturbing headline • Researched a bit
about the problem -
Motivation • Read a disturbing headline • Researched a bit
about the problem • Crop losses Source: National Academy of Science, Singh et al. -
Motivation • Read a disturbing headline • Researched a bit
about the problem • Crop losses • We 🧡 tech and doing social good -
The Project • Turns out almost all farmers have access
to smartphones - Research Study by Hughes et al. -
The Project • Turns out almost all farmers have access
to smartphones • What does this do? ◦ Uses AI to diagnose diseases from plant images early -
The Project • Turns out almost all farmers have access
to smartphones • What does this do? ◦ Uses AI to diagnose diseases from plant images early ◦ Provides actionable ways to solve diseases -
The Project • Turns out almost all farmers have access
to smartphones • What does this do? ◦ Uses AI to diagnose diseases from plant images early ◦ Provides actionable ways to solve diseases ◦ Specific ways to solve -
The Project • Turns out almost all farmers have access
to smartphones • What does this do? ◦ Uses AI to diagnose diseases from plant images early ◦ Provides actionable ways to solve diseases ◦ Specific and actionable ways to solve, AI also identifies species ◦ Works Offline, PWA -
The Project • Turns out almost all farmers have access
to smartphones • What does this do? ◦ Uses AI to diagnose diseases from plant images early ◦ Provides actionable ways to solve diseases ◦ Specific and actionable ways to solve, AI also identifies species ◦ Works Offline • Approximated decrease in plant losses: 67% -> 25% -
Existing Solutions • CropNet (Google): only Cassava plants • Plant
Village (Hughess et al.): only a dataset -
Existing Solutions • CropNet (Google): only Cassava plants • Plant
Village (Hughess et al): only a dataset • Plant Disease detector (Ramesh et al): unsatisfactory performance -
Existing Solutions • CropNet (Google): only Cassava plants • Plant
Village (Hughess et al): only a dataset • Plant Disease detector (Ramesh et al): unsatisfactory performance • Some other disease detectors: not optimized or usable on small devices -
How we built this? - The Model • Collected data
• Experimented with multiple architectures and hyperparameters -
How we built this? - The Model • Collected data
• Experimented with multiple architectures and hyperparameters -
How we built this? - The Model • Collected data
• Experimented with multiple architectures and hyperparameters • Optimize the model to run on-device (in a PWA) -
How we built this? - The Model • Collected data
• Experimented with multiple architectures and hyperparameters • Optimize the model to run on-device (in a PWA) • Optimized the model to be just 12 MBs -
How we built this? - The Model • Collected data
• Experimented with multiple architectures and hyperparameters • Optimize the model to run on-device (in a PWA) • Optimized the model to be just 12 MBs -
How we built this? - The Model • Collected data
• Experimented with multiple architectures and hyperparameters • Optimize the model to run on-device (in a PWA) • Optimized the model to be just 12 MBs Wait, this is fast!💡 -
How we built this? - Web App • Offline Support
PWA • Run the optimized model with TFJS -
How we built this? - Web App • Offline Support
PWA • Run the optimized model with TFJS ◦ Individually fetch data flow graph and weights ◦ Normalize images ◦ Resize with nearest neighbour interpolation -
How we built this? - Web App • Offline Support
PWA • Run the optimized model with TFJS ◦ Individually fetch data flow graph and weights ◦ Normalize images ◦ Resize with nearest neighbour interpolation • Finally, deploy the web app to test out with real life scenarios🚀 -
Business Plan • Farmers don't get charged • Potential plant
product companies place ads -
Business Plan • Farmers don't get charged • Potential plant
product companies place ads • After a quality check their ads appear under “How to Solve” -
Business Plan • Farmers don't get charged • Potential plant
product companies place ads • After a quality check their ads appear under “How to Solve” • Data, data and data -
Business Plan • Farmers don't get charged • Potential plant
product companies place ads • After a quality check their ads appear under “How to Solve” • Data, data and data • Collect non sensitive but useful information -
Business Plan • Farmers don't get charged • Potential plant
product companies place ads • After a quality check their ads appear under “How to Solve” • Data, data and data • Collect non sensitive but useful information • Plant images are used anonymously to improve the ML -
Business Plan • Farmers don't get charged • Potential plant
product companies place ads • After a quality check their ads appear under “How to Solve” • Data, data and data • Collect non sensitive but useful information • Plant images are used anonymously to improve the ML • Usage data is shared to plant product companies -
Impact • Can reduce crop losses by 67% -> 25%🚀
• Tested on real scenarios • Featured on Microsoft Blog and YouTube🤗 -
Thank you! And we 🧡 open-source https://git.io/JujAg View on GitHub
Research supported with Cloud TPUs from Google's TPU Research Cloud (TRC)