Quantifying the predictive uncertainty of neural networks via residual estimation with I/O kernel
A residual estimation with an I/O kernel (“RIO”) framework provides estimates of predictive uncertainty of neural networks, and reduces their point-prediction errors. The process captures neural network (“NN”) behavior by estimating their residuals with an I/O kernel using a modified Gaussian process (“GP”). RIO is applicable to real-world problems, and, by using a sparse GP approximation, scales well to large datasets. RIO can be applied directly to any pretrained NNs without modifications to model architecture or training pipeline.
1. A process for estimating residuals of a neural network (NN) model to determine uncertainty in the NN model's predictions of a value of at least one of a physical, chemical or electrical variable, the process comprising:
training, by a processor, a NN model to make one or more predictions of the value of the at least one of a physical, chemical or electrical variable using a training data ( , y) input set, wherein the training data input set includes (x i ,y i )) n i=1 wherein y i are the expected outcomes by the NN model given input x i;
storing, by the processor, an output data set from the NN model, including the one or more predictions resulting from operation on the training data input set, wherein the output data set includes (x i , ŷ i )} n i=1 , wherein ŷ i are the predicted outcomes by the NN model given input x i ; and
training, by the processor, a Gaussian process (GP) to estimate residuals of the NN model when applied to raw input data x* using the training data input set (x i , y i )) n i=1 and the output data set (x i , ŷ i )) n i=1,
wherein the training of the Gaussian process (GP) includes:
calculating, by the processor, residuals r={r i =y i −ŷ i } n i=1 wherein r denotes the vector of all residuals and ŷ denotes the vector of all NN model predictions;
calculating, by the processor, an n×n covariance matrix at all pairs of training points based on a composite kernel K c (( , ŷ), ( , ŷ)), where each entry is given by k c ((x i , ŷ i ), (x j , ŷ j ))=k in (x i , x j )+k out (ŷ i , ŷ j ), for i,j=1,2, . . . , n; and
optimizing, by a gradient-based optimizer, GP hyperparameters σ 2 in , l in , σ 2 out , l out , and σ 2 n by maximizing log marginal likelihood
log p ( r| ,ŷ )=−½ r T ( K c (( , ŷ ), ( , ŷ ))+σ 2 n I ) −1 r −½log| K c (( , ŷ ), ( ,ŷ ))+σ 2 n I|−n /2log 2π.
2. The process according to claim 1 , wherein the NN model is a fully connected feed-forward network.
3. The process according to claim 1 , further comprising:
applying, by the processor, the trained Gaussian process (GP) to predictions ŷ * of a neural network (NN) model applied to raw input data x * , the applying including:
calculating, by the processor, residual mean;
calculating, by the processor, residual variance; and
returning, by the processor, distribution of calibrated prediction ŷ′ * .
4. The process according to claim 3 , further comprising
calculating, by the processor, residual mean in accordance with {circumflex over ( r )} * =k T * (K c (( , ŷ), ( , ŷ))+σ 2 n I) −1 r;
calculating, by the processor, residual variance in accordance with var({circumflex over (r)} * )=k c ((x * , ŷ * ), (x * , ŷ * ))−k T * (K c (( , ŷ), ( , ŷ))+σ 2 n I) −1 k * ; and
returning, by the processor, a distribution of calibrated prediction ŷ′ 2 in accordance with ŷ′ * ˜ (ŷ * +{circumflex over ( r )} * , var({circumflex over (r)} * )).
5. The process according to claim 3 , wherein the NN model is a fully connected feed-forward network.