MNIST Example Error

I installed DyNet manually for C++ (and set it up for python) with the following ```cmake .. -DEIGEN3_INCLUDE_DIR=../../eigen -DPYTHON=`which python3` -DBACKEND=eigen -DMKL=True -DMKL_ROOT=/opt/intel/mkl -DENABLE_CPP_EXAMPLES=ON```

I tried running the compiled examples (xor and the rnnlm worked) but when I ran the mnist example I got this error
```
mnist: /usr/local/src/dynet-base/eigen/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h:392: Eigen::TensorEvaluator, Device>::TensorEvaluator(const XprType&, const Device&) [with BinaryOp = Eigen::internal::scalar_product_op; LeftArgType = const Eigen::TensorConversionOp > > >; RightArgType = const Eigen::TensorMap >; Device = Eigen::DefaultDevice; Eigen::TensorEvaluator, Device>::XprType = Eigen::TensorCwiseBinaryOp, const Eigen::TensorConversionOp > > >, const Eigen::TensorMap > >]: Assertion `dimensions_match(m_leftImpl.dimensions(), m_rightImpl.dimensions())' failed.
Aborted (core dumped)
```
I added some prints to the examples and the error seems to come from the `cg.backward(loss_expr);` line.

I am using Ubuntu 16.04 and compiled with g++ 5.4

Reproduce the MNIST example using the reported CMake options on Ubuntu 16.04 with g++ 5.4. Start at the cg.backward(loss_expr) call and inspect the Eigen TensorEvaluator assertion for mismatched dimensions. Done means the MNIST example completes without aborting.

Written by the indexing model from the issue text.