Juno Deep Learning

1. Requirements

1.1. Detect Full body, Face, Head and Shoulder, Cars, Street Signs

1.2. Handle 1000 classifications per second

1.3. Fixed 120x120x3 image

2. Schedule

2.1. 2 Months (Started 5-sep-2016)

3. Learning

3.1. Boot Juno

3.2. Put Linux on Juno with Ethernet Support

3.3. Neon

3.4. OpenCL language

3.4.1. Juno with OpenCL

3.5. Im2Col Col2Im on OpenCL

4. Architechture

4.1. Resnet

4.1.1. Make it smaller until it's good enough

4.2. LeoNet

4.2.1. Check performance with augmented Data

4.2.2. Try techniques to reduce weights

4.2.3. Make it Deeper

5. Programming Tasks

5.1. Verify Matlab Implementation

5.1.1. Softmax forward

5.1.2. Implement convolution layer based on im2col/col2im

5.1.3. Relu Layer

5.1.4. Fully Connected Layer

5.1.5. Residual element-wise operations

5.2. Rewrite Layers from Matlab reference to Use Multithreading

5.3. Rewrite Layers from Matlab to OpenCL

5.4. Create a Mex Wrapper for Layers

5.5. Import learned weights from Caffe

5.6. Use Matlab Reference to validate OpenCL layers