As per my understanding you are trying to achieve better accuracy with a very small dataset and would like to know ways of improving the accuracy and some suggestions on architecture.
You can try to augment your dataset to increase the effective size of your dataset. Refer to this link for more details on image input. If you have audio input dataset then you can refer to this link.
You can also try transfer learning with pretrained nets which might suit your purpose if you have an image dataset - link.
If the dataset allows for using a ML algorithm for solving the problem, then it might be a good way to go about it for a small dataset. You may refer to this link for basics.
