Unexpected results from CNTK - machine-learning

Question

I'm running a fairly simple test of CNTK but not getting results that make much sense. My training/test data consist of one feature and one label. The feature is a decimal and the label will be an integer between 0-5. In a majority of cases the value of the label will be 0 or 1 and get increasingly rare as the value gets higher. 5 appears in about 16/30,000 cases.
What's odd is that when I output the results they indicate that each possible label has about an equal chance of occurring. I would expect 0 or one to be the most likely and 5 the be extremely unlikely. I was hoping SO could shed some light on what I might be doing wrong here. I included some sample data, sample output and a config file below.
Config:
# Parameters can be overwritten on the command line
# for example: cntk configFile=myConfigFile RootDir=../..
# For running from Visual Studio add
# currentDirectory=$(SolutionDir)/<path to corresponding data folder>
RootDir = ".."
ConfigDir = "$RootDir$/Config"
DataDir = "$RootDir$/Data"
OutputDir = "$RootDir$/Output"
ModelDir = "$OutputDir$/Models"
# deviceId=-1 for CPU, >=0 for GPU devices, "auto" chooses the best GPU, or CPU if no usable GPU is available
deviceId = 0
command = Simple_Demo_Train:Simple_Demo_Train
precision = "float"
traceLevel = 1
modelPath = "$ModelDir$/simple.dnn"
outputNodeNames = ScaledLogLikelihood
#######################################
# TRAINING CONFIG #
#######################################
Simple_Demo_Train = [
action = "train"
# Notation xxx:yyy*n:zzz is equivalent to xxx, then yyy repeated n times, then zzz
# Example: 10:20*3:5 is equivalent to 10:20:20:20:5
SimpleNetworkBuilder = [
# 2 input, 2 50-element hidden, 2 output
layerSizes = 1:50*3:6
trainingCriterion = "CrossEntropyWithSoftmax"
evalCriterion = "ErrorPrediction"
layerTypes = "Sigmoid"
initValueScale = 1.0
applyMeanVarNorm = true
uniformInit = true
needPrior = true
]
SGD = [
# epochSize = 0 means epochSize is the size of the training set
epochSize = 0
minibatchSize = 25
learningRatesPerMB = 0.5:0.2*20:0.1
momentumPerMB = 0.9
dropoutRate = 0.0
maxEpochs = 10000
]
# Parameter values for the reader
reader = [
readerType = "UCIFastReader"
file = "$DataDir$/train.txt"
miniBatchMode = "partial"
randomize = "none"
verbosity = 1
features = [
dim = 1 # two-dimensional input data
start = 0 # Start with first element on line
]
labels = [
start = 1 # Skip two elements
dim = 1 # One label dimension
labelDim = 5 # Two labels possible
labelMappingFile = "$DataDir$/mapping.txt"
]
]
]
########################################
# TEST RESULTS #
# (computes prediction error and #
# perplexity on a test set and #
# writes the output to the console.) #
########################################
Simple_Demo_Test = [
action = "test"
# Parameter values for the reader
reader = [
readerType = "UCIFastReader"
file = "$DataDir$/test.txt"
miniBatchMode = "partial"
randomize = "none"
verbosity = 1
features = [
dim = 1 # two-dimensional input data
start = 0 # Start with first element on line
]
labels = [
start = 1 # Skip two elements
dim = 1 # One label dimension
labelDim = 5 # Two labels possible
labelMappingFile = "$DataDir$/mapping.txt"
]
]
]
########################################
# OUTPUT RESULTS #
# (Computes the labels for a test set #
# and writes the results to a file.) #
########################################
Simple_Demo_Output=[
action = "write"
# Parameter values for the reader
reader = [
readerType = "UCIFastReader"
file = "$DataDir$/test.txt"
miniBatchMode = "partial"
randomize = "none"
verbosity = 1
features = [
dim = 1 # two-dimensional input data
start = 0 # Start with first element on line
]
labels = [
start = 1 # Skip two elements
dim = 1 # One label dimension
labelDim = 5 # Two labels possible
labelMappingFile = "$DataDir$/mapping.txt"
]
]
outputPath = "$OutputDir$/SimpleOutput" # Dump output as text
]
Sample Training Data:
0.86 2
0.84 0
6.818182 0
1.34 1
1 1
0.92 0
0.7692308 0
0.755102 1
0.86 2
5.466667 0
0.96 0
0.9459459 1
1 4
1 0
0.8421053 2
5.5 0
0.84 2
1.2 2
1.32 1
0.98 0
1 1
1.2 2
5.4 1
1.06 2
0.98 1
1.041667 3
0.82 2
7.333333 0
Sample Output:
3.18673 3.18266 3.19894 3.18264 3.2388 3.235
3.18683 3.18272 3.19895 3.18264 3.23872 3.23491
3.18668 3.18263 3.19894 3.18263 3.23884 3.23505
3.18653 3.18255 3.19893 3.18263 3.23895 3.23518
6.53459 4.97457 3.46288 3.3192 0.668835 0.204602
3.18667 3.18263 3.19894 3.18263 3.23884 3.23505
3.18657 3.18258 3.19893 3.18263 3.23892 3.23515
3.18655 3.18257 3.19893 3.18263 3.23894 3.23516
3.18665 3.18262 3.19894 3.18263 3.23886 3.23507
3.18656 3.18257 3.19893 3.18263 3.23893 3.23515
3.18654 3.18256 3.19893 3.18263 3.23895 3.23517
3.18688 3.18274 3.19895 3.18264 3.23869 3.23487
3.18675 3.18267 3.19894 3.18264 3.23879 3.23498
3.18679 3.18269 3.19895 3.18264 3.23875 3.23494
3.1866 3.18259 3.19893 3.18263 3.2389 3.23512
3.18655 3.18256 3.19893 3.18263 3.23894 3.23517
3.18652 3.18255 3.19893 3.18263 3.23896 3.23519
3.18656 3.18257 3.19893 3.18263 3.23893 3.23515
3.18656 3.18257 3.19893 3.18263 3.23894 3.23516
3.18688 3.18274 3.19895 3.18264 3.23869 3.23487
3.18698 3.1828 3.19896 3.18265 3.23861 3.23477
Mapping File:
0
1
2
3
4
5

Answer 1

With the information provided, it is pretty difficult to give a definitive answer to your question. What is most likely going on: Your network did not yet learn much, and is still in a rather "vague" state, where weights did not yet deviate much from their initial values. This is most likely because you have only used very little training, compared to the number of weights you are learning (mind you that you are using 3 layers of 50 neurons, on a 1D input!). When asked to predict on your test data, most of the test data is dissimilar to any training data, and your network will predict its "best guess", which is that all classes are equally likely.
To check that some learning is actually going on, try for example reducing your problem to a two-class setting, 0 versus the rest, and reduce the network complexity.