-> mediated perception approach – parse entire frame in order to make driving decision -> behavior reflex approach – directly map input image to a driving action by a regressor
Direct perception method developed by this paper. Trained a deep CNN using a recording from 12 hours of human driving in a video game. Trained model in car distance estimation using KITTI dataset.
5 convnet layers followed by 4 fully connected layers (4096, 4096, 256, and 13 for output dimensions). Screenshots are down-sampled to 280 x 210.
They collected 484,815 images for training.
Initial learning rate of 0.01
mini batch of 64 images randomly selected from the training samples
after 140,000 iterations they stopped the training process.
7 different tracks and 22 different cars used
The convnet processes the TORCS images and estimates the 13 indicators for driving. Based on indicators and speed of car, controller will send commands to the car.
13 affordance indicators: angle toMarking_LL toMarking_ML toMarking_MR toMarking_RR dist_LL dist_MM dist_RR toMarking_L toMarking_M toMarking_R dist_L dist_R
Trained a CNN to map raw pixels from a single camera image directly to steering commands. System operating at 30 FPS.
CNNs widely used due to Large Scale Visual Recognition Challenge (ILSVRC) dataset for training and testing and due to GPUs.
CNN Layout: Input Plane (3@66x200) Normalization Normalized input plane (3@66x200) Convolutional feature map (24@31x98) Convolutional feature map (36@14x47) Convolutional feature map (48@5x22) Convolutional feature map (64@3x20) Convolutional feature map (64@1x18) Flatten 1164 neurons dense layer 100 neurons dense layer 50 neurons dense layer 10 neurons dense layer outpur (vehicle control)
convo layers perform feature extraction. They have different kernel sizes either 3x3 or 5x5 in this case.
In data augmentation, the set of frames are augmented by adding artificial shifts and rotations.
No labels were used for outline of road etc. The steering commands sent out from the network were 1/r, where r is the the turn radius.
Regression is when a model, such as a neural network, accepts input and produces a numeric output. The output of a classification model is what class the input belongs to.
In deep learning, the more data present the higher the performance.
CNNs can scan an image for patterns within the image. Recurrent NN can find patterns across several inputs, not just within a single input.
Frameworks for deep learning: TensorFlow (Google), MXNet (Amazon), Theano (Univ of Montreal), and CNTK (Microsoft).
Keras – is a high-level neural network API, written in Python and able to run on top of TensorFlow, CNTK, or Theano
Data fed into a machine learning model needs to be normalized. Zscore used for normalization.
Z = input – mean / standard deviation.
Training/Validation split (80/20) or K-Fold Cross Validation
Activation functions, also known as transfer functions, are used to calculate the output of each layer of a neural network.
ReLU (Rectified Linear Unit) used for output of hidden layers. Softmax used for the output of classification neural networks. Linear used for the output of regression neural networks.
ReLU = max(0,𝑥)
hidden layer values = A (W1 * x + b1), where A is the activation function, x is the input, W1 is the weight, and b1 is the bias.
Good idea to save big neural networks so they can be reloaded later. A reloaded nn will not require training. NN can be saved as YAML (just structure no weights) or JSON (just structure no weights) or HDF5 (structure + weights).
Overfitting occurs when a neural network is trained to the point that it begins to memorize rather than generalize.
The mean square error is the sum of the squared differences between the prediction (𝑦̂ ) and the expected (𝑦). MSE values are not of a particular unit. If an MSE value has decreased for a model, that is good. However, beyond this, there is not much more you can determine. Low MSE values are desired.
The root mean square (RMSE) is essentially the square root of the MSE. Because of this, the RMSE error is in the same units as the training data outcome. Low RMSE values are desired.